# Wajarri v3 Research Plan and Work Log

**Program:** English to Wajarri (`eng_Latn` to `wbv_Latn`)  
**Canonical source:** `/mnt/donto-data/donto-resources/research/language-programs/wajarri-v1/WAJARRI-V3-RESEARCH-PLAN-AND-WORK-LOG.md`  
**Public HTML:** `https://mobtranslate.com/docs/wajarri-v3-research-plan.html`  
**Public Markdown:** `https://mobtranslate.com/docs/wajarri-v3-research-plan.md`  
**Started:** 2026-07-31  
**Current phase:** v3 is complete as a narrowly bounded controlled release. Three independently seeded A40 runs, CPU float32 parity, immutable Hugging Face publication, shared-Space serving, public model-registry publication, production homepage routing, PostgreSQL cache identity, and desktop/mobile browser acceptance have passed. The release supports six dictionary-bound subject-slot constructions only; it is not a free-form Wajarri translator. No new Wajarri target sentence was invented for this release.  
**RunPod policy:** Wajarri worker `77tes61jw5a024` was deleted only after complete result recovery and independent local checksum verification. The post-deletion provider census proves it absent; unrelated Alpha worker `wtupxv15debnvh` was not touched. No Wajarri RunPod worker remains active.

This is the single living research plan and work log for the next Wajarri model. It replaces scattered progress notes. New evidence, corpus decisions, experiments, failures, and promotion rulings are appended here.

## Executive decision

Do not train for more epochs on the present corpus. Publish and serve v3 only through its fail-closed six-construction contract; keep dictionary lookup outside the model and route every unsupported sentence to the ordinary prompt-based path.

The final experiment succeeded by narrowing the task instead of pretending that sparse evidence could support unrestricted generation. The model receives a masked English construction, the decoder is seeded with the target-language and `<copy>` control tokens, and the model generates one of six complete source-bound predicates. A deterministic renderer replaces the single first-position slot with an exact Wajarri dictionary headword. This produced 35/35 exact templates and rendered clauses for all three independent seeds and exact GPU-to-CPU parity. It does not license novel inflection, arbitrary syntax, or a speaker-validated naturalness claim.

Wajarri v2 is a strong closed-set lexical reconstruction artifact, but it is not a competent sentence translator. The selected checkpoint reconstructs 1,076 of 1,179 one-target prompts exactly (91.26%), yet scores 0 of 5 exact and 9.62 chrF++ on the small complex historical diagnostic. The selected checkpoint consumed only the first 640 presentations of a planned 3,712-presentation recovery schedule: 509 lexical presentations, 111 synthetic sentence presentations, 14 historical sentence presentations, and six fixed-utterance presentations. The remainder of that schedule was never presented to the selected checkpoint and must not be described as its training trajectory.

The paired source-record audit also rules out continuing the current raw definition-context objective. On the same 1,684 records, direct prompts reproduce the source target 1,094 times while definition-context prompts do so 878 times. Definition context breaks 227 direct successes and rescues only 11 direct failures. This does not prove that definition text is intrinsically harmful; it proves that the present ungoverned prompt is harmful for this checkpoint. v3 will test explicit POS/sense labels and sentence-conditioned lexical use instead.

The checkpoint-level retention diagnosis is now complete. It found one safe control checkpoint, no safe treatment checkpoint, and a monotonic tradeoff in which composition fitting overwrote governed lexical forms. The immediate task is therefore source-backed contrast acquisition and a task-separated sentence-adapter design, followed by a new controlled experiment. More raw epochs, repeats of the present 36-row matrix, or undifferentiated sentence volume are not justified by the evidence.

### Current controlling result

The controlling result is corrected confirmation run `wbv-v3-forced-slot-confirmation-runpod-20260802-a2`. It passes only the controlled subject-slot route. The first apparent three-seed result was invalidated because Transformers silently used Trainer seed 42 in all three runs; the replacement run explicitly bound both `seed` and `data_seed`, failed unless the runtime values matched, and required three distinct selected adapter hashes.

| Seed | Selected update | Raw template exact | Rendered clause exact | Batch 1/16 mismatches | Retention | Synthetic exact | Synthetic chrF++ | One-target lexical exact | Adapter SHA-256 |
| ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- |
| 17 | 40 | 35 / 35 | 35 / 35 | 0 | 54 / 55 | 22 / 36 | 75.4246 | 813 / 1,179 | `a4a91d4e6a6ba49509a8ab432c0674686c96b4611f58db8b63089f26707ed3d6` |
| 42 | 40 | 35 / 35 | 35 / 35 | 0 | 54 / 55 | 22 / 36 | 75.3028 | 816 / 1,179 | `d59a4a276d4968e6d4e453dd534ebca6edfaacf936e1bd97cf7c3eb4be1c8801` |
| 73 | 40 | 35 / 35 | 35 / 35 | 0 | 54 / 55 | 22 / 36 | 75.1763 | 824 / 1,179 | `8574727bbf9cfb3e5032e9d922d7af61d6eb62f4b6b0976d6fc06b18b7dd82f7` |

Every seed had zero blank, source-copy, repeated-four-gram, unresolved-slot, and unresolved-bracket faults on the authorized route. The winning seed-17 adapter then produced byte-identical normalized predictions under A40 bfloat16 and CPU float32 at batch sizes 1 and 16. The prediction SHA-256 was `b17ab8d44ccaaa8dba962ecf46486d731d7aaf8d4c6a7adeed430044612cf25f` in all four conditions.

The lower one-target lexical scores are not hidden. They demonstrate why v3 must not replace deterministic dictionary lookup or the public lexical route. The model is authorized only for this composition contract:

```text
public English:       The bilby is running.
dictionary binding:  bilby = marruwa
model input:          <translate> The <copy> is running.
decoder prefix IDs:   [2, 256204, 256208]
raw model output:     <copy> jamarnimanha.
rendered result:      marruwa jamarnimanha.
```

The other complete predicates are `yanajimanha` (coming towards the speaker), `yanmanha` (going away from the speaker), `nyinamanha` (sitting), `nyinarangamanha` (sitting down), and `garrimanha` (standing). The router requires one exact dictionary headword and one exact frozen construction. It refuses malformed slot outputs. Unsupported constructions never invoke this adapter and continue through the ordinary dictionary-prompt route.

| Public component | Immutable identity |
| --- | --- |
| Model | [`ajaxdavis/mobtranslate-wajarri-v3-controlled`](https://huggingface.co/ajaxdavis/mobtranslate-wajarri-v3-controlled) commit `1a26382648e294f2c102102ef914c7b0c86bec20` |
| Model tag | `v3.0.0-controlled-subject-slot-s17-step40-20260802` |
| Dataset | [`ajaxdavis/mobtranslate-wajarri-synthetic-corpus-v1`](https://huggingface.co/datasets/ajaxdavis/mobtranslate-wajarri-synthetic-corpus-v1) commit `302f3a93eb255a2fc70692b7c429ba5ee8300687` |
| Shared Space | [`ajaxdavis/alpha-v0-historic`](https://huggingface.co/spaces/ajaxdavis/alpha-v0-historic) commit `03b6d4ed9c9ffb8ac2e55676447084af2c7aa5c8` |
| Homepage release | `20260802T103148Z-72668fe89dc5` |

The release claim is therefore positive but narrow: exact controlled reconstruction and rendering over six source-bound predicates and exact dictionary-bound subjects. Free-form translation, productive morphology, historical-text competence, discourse competence, and fluent-speaker validation remain unproven and unauthorized.

### First causal screen result

The one-seed `B0`/`C0`/`T1` A40 screen completed on 2026-07-31. `B0` is the untouched public v2 adapter, `C0` is a step- and token-matched retention continuation, and `T1` is the 36-row composition intervention. The treatment caused measurable sentence learning, but it failed the lexical noninferiority gate and is not a release candidate.

| Development measure | `B0` | `C0` | `T1` |
| --- | ---: | ---: | ---: |
| Exact sentence | 0 / 13 | 0 / 13 | 3 / 13 |
| Both required slots present | 0 / 13 | 0 / 13 | 3 / 13 |
| Required subject present | 7 / 13 | 9 / 13 | 7 / 13 |
| Required predicate present | 1 / 13 | 1 / 13 | 7 / 13 |
| Mean chrF++ | 42.03 | 43.18 | 60.56 |

| Full regression measure | `B0` | `C0` | `T1` |
| --- | ---: | ---: | ---: |
| One-target lexical exact | 1,076 / 1,179 | 1,072 / 1,179 | 1,061 / 1,179 |
| Controlled-synthetic exact | 13 / 36 | 12 / 36 | 12 / 36 |
| Controlled-synthetic chrF++ | 70.90 | 71.54 | 70.02 |
| Fixed-utterance retention | 55 / 55 | 55 / 55 | 55 / 55 |

Relative to `B0`, `T1` gained eight exact one-target lexical rows and lost 23, for a net loss of 15. Its 36 unique composition rows were each presented 16 times, concentrating training on only four predicates. This establishes that direct composition supervision can move the model; it does not establish broad Wajarri grammar or sentence reliability. `T1` is not promoted, not uploaded, and not served.

The next experiment therefore lowers the composition dose and interleaves broad lexical anchors. It also uses a fresh 32-cell development matrix made from eight source-bound subjects not used in the first matrix. No new run may tune against the consumed 13-cell development set.

## Evidence audited on 2026-07-31

### Current model

| Measure | Wajarri v2 result |
| --- | ---: |
| One-target direct lexical exact | 1,076 / 1,179 (91.26%) |
| Definition-context exact | 878 / 1,684 (52.14%) |
| Controlled-synthetic holdout exact | 13 / 36 |
| Controlled-synthetic holdout chrF++ | 70.90 |
| Complex historical holdout exact | 0 / 5 |
| Complex historical holdout chrF++ | 9.62 |
| Mechanical faults under frozen decoder | 0 |

The selected model remains public at [ajaxdavis/mobtranslate-wajarri-v2](https://huggingface.co/ajaxdavis/mobtranslate-wajarri-v2). Its release card correctly states that free-form translation has not passed a natural-language gate.

### Final recovery schedule and actual selected-checkpoint exposure

The complete planned context-recovery schedule was:

| Schedule role | Presentations | Share |
| --- | ---: | ---: |
| Definition-context lexical recovery | 1,684 | 45.37% |
| Direct lexical recovery | 1,211 | 32.62% |
| Sentence-context retention | 760 | 20.47% |
| Fixed-utterance retention | 32 | 0.86% |
| Batch-alignment padding | 25 | 0.67% |

The schedule contained 3,712 presentations but only 3,093 unique parent rows. Presentation count is not independent evidence count. More importantly, the public checkpoint is step 20 and consumed only the following prefix:

| Consumed role | Presentations | Share |
| --- | ---: | ---: |
| Definition-context lexical recovery | 287 | 44.84% |
| Direct lexical recovery | 222 | 34.69% |
| Synthetic sentence retention | 111 | 17.34% |
| Historical sentence retention | 14 | 2.19% |
| Fixed-utterance retention | 6 | 0.94% |
| **Total actually consumed** | **640** | **100%** |

Its complete documented project lineage is:

| Lineage stage | Consumed presentations | Unique exposed pairs or rows |
| --- | ---: | ---: |
| v1 lexical/context/synthetic checkpoint 800 | 44,800 | 599 |
| v2 lexical-dose checkpoint 2,220 | 71,040 | 1,409 |
| v2 sentence-recovery checkpoint 25 | 800 | 421 |
| v2 context-recovery checkpoint 20 | 640 | 640 |

These are optimizer presentations, not independent linguistic observations. Upstream NLLB pretraining exposure is unknown.

### Synthetic sentence coverage

There are 211 controlled pairs across only six construction families:

| Construction family | Rows |
| --- | ---: |
| Positive stative, small | 69 |
| Positive stative, good | 42 |
| Positive stative, hungry | 30 |
| Positive stative, cold | 25 |
| Present intransitive, going away | 24 |
| Present intransitive, sitting down | 21 |

The synthetic holdout is not an independent grammar test. It reuses all six construction families, all six predicate forms, and 12 of the 13 training subjects. Its score measures recombination within a small closed inventory.

All 23 selected-model errors on this holdout preserve the expected predicate while substituting the subject noun. Examples include:

| English | Reference | Model |
| --- | --- | --- |
| The old woman is cold. | `Jugarnu janda.` | `Jardu janda.` |
| The turtle is going away. | `Gandarra yanmanha.` | `Galyba yanmanha.` |
| The sandhill is good. | `Dungguru barndi.` | `Burlga barndi.` |
| The younger brother is sitting down. | `Buwa nyinarangamanha.` | `Matarna nyinarangamanha.` |

This identifies lexical selection inside sentence generation as a primary error, not failure to reproduce the six predicates.

### Sentence lineage, task pairing, and split integrity

The 211-row public synthetic corpus contains two record shapes: 57 rows carry an explicit grammar binding inline, while 154 legacy rows carry their sentence and lexical-compatibility decisions in external review ledgers. All 154 legacy rows join to both required ledgers. They are reviewed research synthesis, not speaker-attested references. Every v3 row will carry its binding and review provenance inline so record identity alone is sufficient to audit eligibility.

The private v2 training payload contains 166 synthetic training parents. Of these, 164 have the same record identity in the public corpus. Two older training IDs duplicate public holdout source-target pairs under different newer IDs. One of those two training parents occurs in the selected checkpoint's first 640 presentations. Neither duplicate occurs in the reported 36-row evaluation subset, so this defect does not explain its 13 exact matches; however, those two public holdout identities are contaminated and are retired. A separate evaluated legacy ID was deduplicated to a newer public record with the same pair. v3 must assign splits to normalized source-target groups before record-level deduplication.

Thirty-five of the 36 evaluated sentence rows have an exact join to a local dictionary target for the expected subject. Their paired outcomes are:

| Isolated direct task | Sentence task | Rows |
| --- | --- | ---: |
| Exact | Exact | 12 |
| Exact | Failed | 18 |
| Failed | Exact | 1 |
| Failed | Failed | 4 |

The 18 sentence failures after exact isolated reconstruction are the clearest evidence for sentence-conditioned lexical realization or composition failure. The five isolated-task failures are not five independent lexical phenomena: they collapse to three source-record clusters. Three butterfly rows consistently select a competing surface across both tasks; the sandhill row selects different competing forms by task; and the kookaburra sentence is exact even though isolated reconstruction selects another recorded form. Those clusters require sense, synonym, variety, and orthographic adjudication before accepted references or training overlays change.

Canonical lineage artifacts are under `analysis/benchmark-runs/wbv-v3-v2-sentence-lineage-audit-v1.1/`. The contract SHA-256 is `9559e794a2f22114a77766e4bc927213d112958f4008459e01ac65f2ab049b56`; the report SHA-256 is `7d31914dd85a2e5d554c1804a421e4faa2bbc08ca6c5f873234d468b6b967af3`. All artifact checksums pass, and a clean rerun is byte-identical. The earlier v1 audit is retained and explicitly superseded because it conflated an evaluated legacy row's external review provenance with the inline provenance of its deduplicated public replacement.

### Lexical failure character

Among the 103 one-target direct failures:

| Failure class | Count |
| --- | ---: |
| Wrong known Wajarri target | 62 |
| Different surface form | 27 |
| Near surface form | 11 |
| Over-generation | 2 |
| Under-generation | 1 |

Exact recall by reference-token length is 89.6% to 93.8% across every bucket from one token through five-plus tokens. Tokenizer fragmentation may matter, but it does not explain the main lexical failure pattern. The current model usually generates well-formed known vocabulary and selects the wrong item.

### Grammar and source frontier

The current grammar edition contains:

| Evidence class | Count | Current use |
| --- | ---: | --- |
| Douglas numbered examples | 269 | OCR candidates, not semantically admitted |
| Visually curated Douglas examples | 40 | Transcription accepted; alignment unresolved |
| Table blocks | 11 | Paradigm review required |
| Morphotactic statements | 16 | Historical analysis requiring reconciliation |
| Source-anchored grammar assertions | 69 | Mostly not independently accepted for generation |
| Historical sentence witnesses | 29 | 24 train, 5 development diagnostic |
| Productive synthetic templates | 6 | Narrow present/stative scope |
| Independent natural speaker-validated rows | 0 | Natural promotion gate remains closed |

Seven downloaded ICTV videos total about 20 minutes and contain Wajarri teaching, profiles, oral history, narrative, cooking, and place material. They have no accepted transcript or alignment yet. The 50words review contains 55 source pairings, including eight fixed utterances and 165 audio links. Several bilingual Wajarri books are identified but not yet acquired as aligned text.

## Research basis

The revised design follows several results from low-resource machine-translation research:

1. The current [FormosanBank NLLB adaptation guide](https://huggingface.co/blog/FormosanBank/nllb-200-mt) recommends directional checkpoints, source/document-level split discipline, a single-token language code, correct NLLB decoder configuration, train-only tokenizer adaptation, and corpus improvement before model tuning.
2. [GATITOS](https://aclanthology.org/2023.emnlp-main.26/) finds that curated bilingual lexicons improve low-resource translation and that lexicon quality matters more than indiscriminate noisy scale.
3. [Rare-word data augmentation](https://aclanthology.org/P17-2090/) improves translation by placing low-frequency words in new sentence contexts rather than merely replaying isolated pairs.
4. A recent [Indigenous-language synthetic-data study](https://aclanthology.org/2026.loresmt-1.10/) finds gains from curated plus filtered synthetic data and deterministic language-specific normalization.
5. The [Mambai retrieval study](https://aclanthology.org/2024.eurali-1.1/) finds gains from combining dictionary entries with retrieved parallel examples, while showing that manual-derived tests can badly overstate performance on native-speaker material.
6. A [2026 PEFT comparison](https://aclanthology.org/2026.americasnlp-6.4/) finds full fine-tuning strongest, OFT statistically close, and LoRA a strong efficiency-quality tradeoff across 13 Indigenous language pairs. Data quality and task coverage remain prior questions.
7. Published 20,000-pair results used [20,000 high-quality parallel pairs and a 16,000-sentence evaluation set](https://aclanthology.org/2025.mmloso-1.14/). That scale is not comparable to repeating six deterministic templates.
8. Wajarri is suffixing and agglutinative, with split-ergative case, bound pronouns, TAM contrasts, and subordinate-clause dependencies. These must be represented explicitly rather than inferred from a word list. See [Marmion's Wajarri overview](https://academic.oup.com/book/46706/chapter-abstract/410242905).

## Phase 1: benchmark and failure ledger

**Status:** Complete for the published Wajarri v2 baseline. The ledger is frozen and reproducible; later models must be scored against the same rows plus newly sealed suites.

### Lexical suites

Build and freeze:

| Suite | Population | Primary purpose |
| --- | ---: | --- |
| `L-CENSUS` | 2,511 dictionary source records: 1,684 current and 827 historical | Complete source-record accounting |
| `L-DIRECT` | 1,408 normalized direct prompts | Ambiguity-aware reconstruction |
| `L-ONE` | 1,179 one-target prompts | Strict exact reconstruction |
| `L-CONTEXT` | 1,684 current source-record definition prompts | Closed definition-conditioned reconstruction |
| `L-SENSE` | To be frozen after sense-level review | Genuine POS- and sense-qualified selection |

The older count of 3,016 was the sum of 1,408 direct prompt groups and 1,608 grouped definition-context rows. It was not a count of dictionary source records. The selected v2 report contains 3,092 lexical outcomes: 1,408 direct prompt outcomes plus one definition-context outcome for each of 1,684 current source records. These are related views over the dictionary, not 3,092 independent lexical facts.

Every model failure receives a stable row ID and the following machine-readable fields:

- accepted reference set and source-record identities;
- prediction and normalized prediction;
- exact, grapheme CER, chrF++, blank, copying, and degeneration outcomes;
- known-target substitution and nearest-reference analysis;
- documented project-lineage exposure;
- target token fertility and target frequency;
- source-provided sense, variety, and lexical metadata;
- dynamically discovered semantic cluster;
- fixed-versus-regressed comparison against every later run.

Semantic groups will be discovered with embeddings and then reviewed. They will not be maintained as a brittle list of English keyword rules.

### Frozen v2 baseline findings

The checksum-bound run evaluated 3,188 rows and produced 1,143 failures. Of those, 1,115 are lexical failures, 23 are controlled-synthetic sentence failures, and five are historical sentence failures. There were no blanks, source copies, repeated four-token loops, or retention failures under the selected decoder.

| Frozen suite | Exact | Result |
| --- | ---: | ---: |
| Direct lexical, all prompt groups | 1,099 / 1,408 | 78.05% |
| Direct lexical, one accepted target | 1,076 / 1,179 | 91.26% |
| Direct lexical, multiple accepted targets | 18 / 224 | 8.04% |
| Definition-context, current source records | 878 / 1,684 | 52.14% |
| Controlled-synthetic holdout | 13 / 36 | 36.11% |
| Historical diagnostic | 0 / 5 | 0% |
| Fixed-utterance retention | 55 / 55 | 100% |

The direct benchmark also contains five explicitly reconciled cross-source conflicts; all five score exact against the declared union of references. They are retained as their own ambiguity class rather than silently folded into one-target or multi-target rows.

The one-target result is a closed-set reconstruction measurement, not lexical generalization: all 1,179 prompts had an exact input-target exposure somewhere in the documented model lineage, and 103 still fail. In the lexical-dose stage alone, 1,194 of all 1,408 direct prompts had an exact accepted pair exposed; these score 1,091/1,194 (91.37%), while the 214 prompts without that exact exposure score 8/214 (3.74%). This is a post-hoc association, not a causal estimate.

The selected recovery checkpoint saw 287 of the 1,684 definition-context rows once. Those rows score 157/287 (54.70%); the other 1,397 score 721/1,397 (51.61%). One presentation is not a meaningful context-dose experiment. A controlled dose study remains justified, but simply completing the old schedule would not answer whether definition context, explicit sense labels, or sentence use caused an improvement.

The stronger paired result is operationally decisive. Across the same 1,684 source records, direct prompting is source-target exact for 1,094 records and raw definition-context prompting for 878. The paired outcome classes are 867 both exact, 227 direct only, 11 context only, and 579 neither. Raw definition context therefore has a -12.83 percentage-point paired exact-rate difference. The records are clustered by prompt and dictionary structure, so this is a descriptive frozen-benchmark result rather than an independent-binomial population estimate.

Across all lexical failures, the model generated another recorded Wajarri dictionary target 685 times. The existing coarse surface labels contain 691 wrong-known-target outcomes, 306 different forms, 47 near forms, 37 over-generated forms, and 34 under-generated forms. This makes lexical competition and sense selection the main measured failure mode. More isolated replay alone is not the preferred intervention.

All 23 controlled-synthetic failures preserve at least one final reference word, and inspection shows the expected predicate is retained while the subject slot changes. Because all 36 rows come from only six already-trained construction families, this test does not discriminate broad grammar competence. The next sentence suite must hold out complete construction families and morphology combinations.

Canonical Phase 1 artifacts are under `analysis/benchmark-runs/wbv-v3-v2-baseline-failure-ledger-v1.1/`. The contract SHA-256 is `0f020d91acc724fa949fec3ab1ae940ac05096ce840dfd1df651917b28ddd107`; the summary SHA-256 is `4be01f8540fac12d9ad53c245a1f3ffa968ddbf1ae50e361143d8d893ca39fe7`. All recorded artifact checksums pass, and a clean second generation produced identical hashes.

### Sentence and morphology suites

Create four distinct suites:

| Suite | Initial target | Split rule |
| --- | ---: | --- |
| Closed-template development | Existing 36 | Development-consumed; never final evidence |
| Historical development | Existing 29 | Source identity retained; five seen diagnostic rows remain development-consumed |
| Morphology and contrast | 400-600 | Hold out lexeme families and feature combinations |
| Compositional sentence test | 300 | Hold out complete template families and lexical combinations |
| Natural final test | At least 150 | At least three independent speaker/text/source clusters |

Sentence rows will expose gold slots for subject, predicate, arguments, case, TAM, polarity, clause count, dependency relation, and construction family. Aggregate chrF++ cannot hide failures in negation, participant roles, or clause completeness.

## Phase 2: evidence conversion

The frozen pre-generation commission is now materialized at `analysis/corpus-requirements/wbv-v3-3000-sentence-commission-v0.1.0/`. It specifies work to be reviewed and generated; it is not itself a synthetic corpus.

| Commission component | Frozen count |
| --- | ---: |
| Construction families | 24 |
| Future unique pairs per family | 125 |
| Future train / development / sealed-test pairs | 2,400 / 300 / 300 |
| Historical Douglas source witnesses | 40 |
| Current SCSA curriculum source witnesses | 20 |
| Source-specific lexical failures retained for review | 817 |
| Highest-priority lexical items selected for sentence overlay | 500 |
| Douglas phenomenon tags assigned to families | 49 / 49 |
| Families supported by bounded existing templates | 2 |
| Families still requiring productive-rule review | 22 |

The 24 families cover property, identity, possession, location, intransitive and transitive TAM, negation, imperatives, role reversal, questions, spatial cases, cause/purpose, temporal dependency, and object-modifying dependency. Every family receives exactly 100 train rows and 25 held-out rows. Development and sealed-test allocations alternate 12/13 by family so both total 300. The design also requires at least 600 morphology or participant-role contrasts, 500 reviewed lexical-competition rows, and 300 additional glossary-conditioned presentations. Isolated dictionary pairs do not consume the 3,000-sentence budget.

All 60 published sentence witnesses remain review-only. The original SCSA PDF is now archived with SHA-256 `ebcd824b6b6679e7c9e875bbff5b3bc27b0f795b56889eb08f76c12868aefe0d`; the source-fidelity audit verifies all 86 candidate records and 106 cited spans against it with minimum target- and English-example token recall of 1.0. That resolves source capture, not linguistic acceptance. The 40 Douglas transcriptions still require semantic/alignment and current-orthography adjudication; the 20 commissioned SCSA pairs still require segmentation, morphology, sense, and variety review before they can license productive generation. All 24 families remain `generation_authorized=false`, and RunPod remains blocked.

1. Complete semantic alignment and current-orthography review for the 40 visually curated Douglas examples.
2. Review the remaining 269 numbered examples in construction-priority order.
3. Reconcile 11 table blocks and 16 morphotactic statements with current Wajarri evidence.
4. Preserve recorded and source-hypothesized paradigm forms as different states; never train on hypothesized cells as attested forms.
5. Transcribe and align the seven downloaded ICTV videos.
6. Integrate the eight accepted 50words fixed utterances and other exact public bilingual witnesses where split and source identity permit.
7. Acquire and align suitable passages from the identified Wajarri dictionary, teaching books, and bilingual narratives.

The Douglas and Marmion disagreement over switch-reference is not resolved. Switch-reference generation remains disabled until the analysis and surface realizations are reconciled.

## Phase 3: controlled corpus expansion

Synthetic scale is conditional:

| Stage | Unique pair target | Stop/go requirement |
| --- | ---: | --- |
| Pilot | 3,000 | At least 18 validated construction families and clean audits |
| Expansion | 8,000 | Improvement on unseen lexeme/construction development suites |
| Full | 15,000-20,000 | Independent gains without family-specific regression |

The pilot should cover, when validated by the evidence ledger:

- stative, identity, locative, and possessive verbless clauses;
- intransitive and transitive clauses with role reversal;
- present, past, and future TAM;
- verbal and nominal negation;
- imperatives, polar questions, and content questions;
- ergative/absolutive, locative, allative, ablative, dative, causal, and purposive functions;
- free and bound pronouns, including attachment position;
- object modification, coordination, temporal relations, and dependent clauses;
- morphology-focused minimal contrasts and legitimate reduplication;
- short discourse continuations after independent evidence exists.

Each pair must carry immutable provenance, grammar-rule IDs, lexeme IDs, semantic slot restrictions, morphology analysis, generator version, split-group ID, split, and audit status. Split groups are assigned from normalized source-target identity plus governed construction and lexeme families before any record-level deduplication. Unique evidence pairs and repeated optimizer presentations are always counted separately.

No macro construction family may exceed 8% of the pilot. No predicate form may exceed 3%. Lexical bindings must obey animacy, valency, semantic-class, and variety restrictions. English prompts must include natural statements, questions, and commands rather than telegraphic template fragments.

### Glossary-conditioned task

Train and evaluate sentence use of retrieved dictionary entries separately:

```text
<translate> The old woman is cold.
<glossary> old woman = jugarnu; cold = janda
```

This asks the model to apply supplied lexical knowledge in grammar, rather than recover arbitrary hidden dictionary mappings. It directly targets the subject-substitution pattern observed in the current synthetic holdout.

Long-tail dictionary words remain lookup-first in production. Context generation will focus on semantically useful core vocabulary and the diagnosed failure population, not force rare names, flora, and fauna into implausible sentences merely to increase a counter.

## Phase 4: controlled model experiments

Do one-seed screening under identical token and update budgets:

| Arm | Treatment |
| --- | --- |
| `B0` | Frozen Wajarri v2 baseline |
| `C0` | Step- and token-matched sentence retention continuation |
| `S` | Balanced grammar-grounded sentence pilot |
| `SG` | `S` plus glossary-conditioned sentence use |
| `FSG` | Same winning data from a fresh frozen Wajarri-compatible base |
| `TOK` | Winning data with a train-only validated tokenizer extension |

The initial mixture target is 65% ordinary grammar-grounded sentences, 15% morphology/contrast rows, 10% glossary-conditioned sentences, 5% natural or historical retention, and no more than 5% isolated lexical replay.

The primary causal question is whether sentence and glossary supervision improve unseen construction and lexical-combination behavior at fixed compute. The secondary question is whether inherited lexical adaptation helps or harms sentence learning.

Only after selecting the data recipe will the program compare LoRA r16, OFT, and one bounded full fine-tune. The winning recipe is then repeated with seeds 17, 42, and 73. Recipe selection uses aggregate development performance, not the luckiest seed.

All weight comparisons use one frozen greedy decoder. Beam search, repetition guards, and reranking are evaluated afterward as serving interventions over fixed weights.

## Phase 5: promotion gates

Dictionary lookup, lexical reconstruction, and sentence generation have independent gates.

### Dictionary route

- deterministic exact retrieval;
- correct sense and variety display;
- source provenance preserved;
- no model claim attached to lookup output.

### Lexical model

- report every `L-CENSUS`, `L-DIRECT`, `L-ONE`, `L-CONTEXT`, and governed `L-SENSE` row;
- zero blank, source-copy, or uncontrolled-extra-text outputs;
- no material regression from v2's 91.26% one-target exact result;
- report confidence-adjusted frozen-benchmark performance without claiming general reliability.

### Sentence model

- statistically supported improvement over v2 on the new compositional benchmark;
- improvement in subject, predicate, case, TAM, polarity, and clause-completeness slots;
- no construction family may fail while being hidden by aggregate scores;
- zero participant reversal, polarity reversal, or dropped required clause in the critical subset;
- no material regression on retained sentence families;
- blinded natural-language review before any reliability claim.

Automatic metrics are diagnostics. They do not authorize natural translation by themselves; low-resource learned metrics can correlate weakly with human judgments. See the [ACL evaluation study](https://aclanthology.org/2024.acl-short.58/).

## Product experiment

Benchmark four routes on the same sealed rows:

1. NLLB adapter alone;
2. NLLB with retrieved glossary entries;
3. retrieval-grounded large-language-model translation;
4. NLLB draft plus retrieved dictionary, grammar evidence, and constrained post-editing.

The homepage should continue to expose route provenance. Known one-word queries use deterministic dictionary lookup. Sentence output remains a research draft until the sentence gate passes.

## Resource estimate

| Work | Expected scale |
| --- | --- |
| Completed `C2`/`T2` one-seed screen | About 20 minutes of A40 worker time at USD 0.44/hour; nominal compute about USD 0.15 before provider billing granularity |
| Next checkpoint-sentinel diagnostic | Estimated 20-35 A40 minutes, about USD 0.15-0.26 before startup and billing granularity; not authorized |
| Source review for the first contrast commission | 48-96 new unique pairs only after the evidence gates below pass |
| Broader factorial expansion | 80-160 new unique pairs only after the first contrast screen passes |
| Multi-seed confirmation | Deferred until one recipe passes all checkpoint and full-regression gates |

These are planning bounds, not promises. Training begins only after data hashes, splits, token accounting, and evaluation code are frozen.

## Work log

### 2026-07-31: research audit and decision

- Audited Wajarri v2 release metrics, prediction files, schedule composition, tokenizer analysis, grammar edition, source inventory, and current synthetic corpus.
- Recomputed the one-target failure classes and recall by token-length bucket.
- Established that wrong-known-word substitution and absent sentence grammar are more important than tokenizer fragmentation at the current stage.
- Established that the existing synthetic holdout is a closed-template development suite, not independent grammar evidence.
- Reviewed current NLLB adaptation, lexical augmentation, synthetic-data, PEFT, retrieval, tokenizer, and low-resource evaluation research.
- Froze the execution order: benchmark, evidence conversion, 3,000-pair pilot, controlled screening, conditional expansion, multi-seed confirmation.
- Verified that no Wajarri or MobTranslate RunPod pod is running. The unrelated Alpha pod was not touched.

### 2026-07-31: Phase 1 started

- Began implementation of the complete lexical and sentence failure ledger.
- Next gate: materialize a checksum-bound benchmark manifest and deterministic baseline report before commissioning new synthetic rows.

### 2026-07-31: Phase 1 completed

- Added the reusable deterministic analyzer `training/translation/build_wajarri_v3_baseline_ledger.py` and four focused unit tests.
- Bound the exact model revision, adapter hash, decoder, dictionary census, four training-lineage exposure ledgers, five evaluation suites, prediction files, and the declared five-row benchmark reconciliation in one immutable analysis contract.
- Verified all expected row counts and input hashes before analysis. The analyzer failed closed on the first run when the five cross-source accepted-reference unions were not yet declared in the contract; the reconciliation was then bound explicitly rather than ignored.
- Materialized 3,188 complete evaluation rows, 1,143 failures, 1,115 lexical failures, 1,684 source-record outcomes, 817 source-specific failures, 685 known-target confusion groups, 2,423 prediction-collapse groups, and 94 mechanically discovered orthographic surface families.
- Verified all output hashes and repeated the complete analysis in fresh scratch storage with byte-identical artifact hashes.
- Corrected the population accounting: 2,511 source records exist in the census; 3,016 was an older sum of two grouped benchmark views.
- Corrected the exposure narrative: the selected public checkpoint consumed 640 presentations, not the full planned 3,712-row schedule.
- Rejected an immediate extra-epoch run. The measured next targets are sense-qualified lexical competition, grammar-grounded lexical use, held-out construction families, morphology contrasts, and subject-slot selection.

### 2026-07-31: Phase 2 started

- Began converting the 40 visually curated Douglas examples and the failure clusters into a source-bound construction-priority queue.
- Froze a 3,000-pair commission across 24 balanced construction families, with 2,400 train, 300 development, and 300 sealed-test rows.
- Assigned all 49 catalogued Douglas phenomenon tags to the pilot rather than silently dropping difficult constructions.
- Materialized a 60-row source-pair review queue, an 817-row lexical-failure review queue, and a 500-item sentence-overlay priority subset.
- Retired the current raw definition-context objective after the paired audit found 227 regressions against 11 recoveries.
- Added `training/translation/build_wajarri_v3_sentence_commission.py` and four focused tests. Eight baseline-plus-commission tests pass.
- Verified all commission checksums and regenerated the complete artifact byte-for-byte identically in fresh storage.
- Commission contract SHA-256: `95fc8f8120585b3f1d1c7843f641daff3f69d26ebb5dec514c9570885c423bbd`. Report SHA-256: `d9d80ad486b98f497439b18c5b201e490fcba47db9d381fae8162d67535a442e`.
- Next gate: adjudicate source witnesses and productive morphology family by family, then build and reject-or-promote a 300-pair dry-run census before expanding to 3,000 pairs.

### 2026-07-31: sentence lineage and composition audit completed

- Added `training/translation/build_wajarri_v3_sentence_lineage_audit.py` and four focused regression tests for review-ledger joins, private/public parent identity, pair-level split collisions, and deduplicated evaluation identities.
- Reconciled the public 211-row corpus, private 3,569-row training payload, 3,712-presentation checkpoint schedule, 36-row synthetic evaluation, 186 sentence reviews, 444 compatibility decisions, and 1,684 source-record lexical outcomes under one checksum-bound contract.
- Corrected actual selected-checkpoint sentence exposure from 125 synthetic rows to 111 synthetic plus 14 historical presentations.
- Verified that all 154 sparse public synthetic rows have both required external review records; the remaining 57 carry explicit grammar bindings inline.
- Found two normalized source-target pairs assigned to private train and public holdout under different IDs. One was consumed by the selected checkpoint. Neither appears in the reported 36-row score, but both public holdout identities are retired.
- Found one evaluated legacy ID replaced in the public corpus by an inline-bound record with the same pair. Preserved evaluation-record and public-record provenance as separate fields.
- Paired sentence behavior with exact local lexical targets on 35 rows. Direct reconstruction succeeds on 30; sentence generation succeeds on 13. Eighteen sentence failures remain after exact direct reconstruction.
- Reduced the five direct-task failures to three source-record clusters instead of treating correlated rows as independent evidence.
- Preserved the first v1 audit and issued v1.1 as its explicit supersessor. Contract SHA-256: `9559e794a2f22114a77766e4bc927213d112958f4008459e01ac65f2ab049b56`. Report SHA-256: `7d31914dd85a2e5d554c1804a421e4faa2bbc08ca6c5f873234d468b6b967af3`.
- Verified every artifact checksum, a byte-identical clean rerun, and all 12 related unit tests in the pinned `regex==2026.7.19` environment.
- Bound v3 split policy to normalized pair groups before deduplication and prioritized sentence-conditioned lexical overlays plus glossary uptake over more isolated replay.
- Next gate: adjudicate the three direct-failure source clusters and the 60 source-pair witnesses, then materialize the 300-pair dry-run census. RunPod remains blocked until that census passes.

### 2026-07-31: live cluster probe and source-bound triage completed

- Probed the immutable public Hugging Face Space model `mobtranslate-wajarri-v2`, revision `v2.0-context-guarded-nr4-s17-context-recovery-r16-s20-20260731`, adapter SHA-256 `05e6c40172d93b99538b3a21a04bbbd14a6c767180f85b83d5f00064a6858fb4`.
- All nine requests were model-bound and successful; one matched its source-scoped accepted surface exactly. Eight rows were comparable with the frozen A40 run: seven reproduced the same output and one sandhill lexeme differed (`nyilira` on A40 BF16, `nyili` on the Space CPU BF16 runtime).
- This does not establish a particular kernel as the cause. It establishes that runtime backend and generation environment belong in the behavioral identity of an artifact and must be recorded in future comparisons.
- Live-probe contract SHA-256: `893a8af169ab9eb9759d913eb6c696a1da91ea43aeb93404a3967bc5ef0737f8`. Report SHA-256: `10ca434327b13eef99b006ee39c352750fdc5bb1041861c6d48fd9447cc81756`.
- Built three source-bound dossiers covering five sentence rows and nine live probes. Butterfly remains blocked on source surface and nominal-slot adjudication; sandhill does not support the earlier informal claim that the model was correct; kookaburra requires separation of an exact sentence realization from an ambiguous isolated lemma/paraphrase.
- No model output was promoted into the dictionary or reference set. All three clusters remain development-only and outside sealed evaluation.
- Cluster-triage contract SHA-256: `51130f4007112145e028be5530aaaa1f660de46835af8ecbac6ec5634f02d745`. Report SHA-256: `daa330f49be3741cead1c7c99da327d543620664612cd12ba913d581c3d00192`.

### 2026-07-31: curriculum-source fidelity checked

- Archived the original SCSA Wajarri curriculum PDF with SHA-256 `ebcd824b6b6679e7c9e875bbff5b3bc27b0f795b56889eb08f76c12868aefe0d`.
- Verified 86 extracted candidate records and 106 cited spans against the PDF. Minimum target-side and English-example token recall is 1.0.
- The 20 explicit bilingual witnesses remain source-faithful review candidates, not automatically accepted productive templates. Source capture, sentence alignment, morphology, productive generalization, and naturalness are separate decisions.

### 2026-07-31: model-bound 20-sentence pre-census completed

The live census used 20 source-bound sentence previews and nine unique lexical realizations. It produced no blanks and no source copies.

| Outcome | Count |
| --- | ---: |
| Direct lexical realizations exact | 7 / 9 |
| Sentences exact | 2 / 20 |
| Both lexemes exact but sentence failed | 10 |
| Lexical failure and sentence failure | 8 |
| Expected subject present | 8 / 20 |
| Expected predicate present | 4 / 20 |
| Both expected slots present | 2 / 20 |
| Common-noun family exact | 2 / 15 |
| First-person family exact | 0 / 5 |

Representative failures show distinct mechanisms rather than one generic quality problem:

| English | Source-bound candidate | Model | Observed failure |
| --- | --- | --- | --- |
| The child is running. | `Mayu jamarnimanha.` | `Mayu yanmanha.` | Subject retained; predicate collapses to going |
| The dog is sitting. | `Duthu nyinamanha.` | `Duthu nyinarangamanha.` | Specific sitting-down form replaces sitting |
| I am coming. | `Ngatha yanajimanha.` | `ngurragi yanmanha` | Fixed first-person/home-going phrase overreaches |
| The woman is sitting. | `Nyarlu nyinamanha.` | `Jurdu nyinarangamanha.` | Sister/female-relative subject and sitting-down frame substitute |
| The child is standing. | `Mayu garrimanha.` | `Jura nyinarangamanha.` | Competing child form plus wrong posture frame |

The census contract SHA-256 is `10426dd564eafd8c83ddaa1963c1714aedf9ac18354521524787db15877d2987`; the report SHA-256 is `972394210283ec0d127781b05653b72921cf15d7de1ae3f7d1db0f5463dc62be`. Every output checksum passes.

### 2026-07-31: selected-checkpoint lineage attribution completed

- Joined every pre-census prediction to the first 640 selected-checkpoint presentations and the 1,684-row dictionary source ledger.
- None of the 20 candidate sentence prompts occurred exactly in the selected checkpoint prefix.
- The failed outputs contain 11 non-reference Wajarri tokens. Ten are recorded somewhere in the dictionary source ledger and seven occurred in the selected checkpoint prefix. The model is generally selecting known or recently presented Wajarri surfaces, not emitting arbitrary strings.
- `nyinarangamanha` occurred in 14 selected-checkpoint presentations and `yanmanha` in 15. Their repeated use in unrelated sitting, standing, running, and coming prompts is consistent with an overlearned narrow-frame hypothesis.
- `ngurragi` occurred in the fixed utterance `ngurragi yanmanha` for “I’m going home.” Its use for both “I am going” and “I am coming” is evidence of phrase-level overreach, not productive first-person composition.
- The selected prefix contains three exact English prompt conflicts for `child`: `Mayu` and `Jura` both occur in cold, going-away, and small sentences. The v3 intervention must not replay identical English prompts with competing targets unless the prompt is sense- or source-qualified.
- `Jurdu` occurred as the subject for sister sentences, while the model repeatedly used it for woman. The source-scoped woman form `nyarlu` also fails direct reconstruction, so this is both a lexical and sentence-selection problem.
- These are exposure associations and hypothesis generators, not a causal estimate. The next fixed-compute intervention is designed to test the hypotheses.
- Lineage contract SHA-256: `590b0ce9705d22bf0abf6b3def454003f091a56fb0cacfab155603815947db96`. Report SHA-256: `d8db4889896c2802d78c0679aa8b9616e62047a435900ad445dbed0646d738a5`.

### 2026-07-31: narrow composition intervention frozen

This section supersedes the earlier “300-pair dry run, then 3,000 pairs” next gate. The 3,000-pair commission remains useful as a long-range coverage plan, but current evidence supports a smaller causal screen first.

The new commission exhaustively classifies 14 subjects by five bounded present-intransitive predicates. It composes only complete source-bound forms; it infers no stem, suffix, conjugation, or unseen surface.

| Cell disposition | Count | Role |
| --- | ---: | --- |
| Novel training candidate | 36 | Nine training subjects by four new predicates |
| Existing clean retention | 10 | Already-issued going-away pairs |
| Development candidate | 13 | Held-out child, dog, and first-person combinations |
| Blocked reference conflict | 1 | Child going-away has competing `Mayu` and `Jura` targets |
| Blocked subject adjudication | 5 | Woman/`nyarlu` versus observed `Jurdu` competition |
| Excluded semantic compatibility | 5 | Fish is not forced into a terrestrial posture/motion matrix |
| **Total classified** | **70** | Every possible cell has an explicit disposition |

Four contrastive requirements are mandatory before screening: coming versus going away; sitting versus sitting down; first-person going away versus going home; and `Mayu`/`Jura` plus `Nyarlu`/`Jurdu` subject competition.

The original census used four generic-going prompts even though `yanmanha` is bound to going away from the speaker. It also used four generic-coming prompts for the multi-target English cue `coming`, while the selected source-scoped form `yanajimanha` is specifically motion towards the speaker. Those eight prompts are retired from exact comparison and replaced in the intervention by `going away` and `coming towards the speaker`. Both original exact sentence results came from the retired generic-going set. The remaining source-qualified live subset is therefore 0/12 exact, while direct reconstruction remains 7/9 exact.

The commission contains zero training-eligible rows. This is intentional: it freezes what must be reviewed without silently converting dictionary concatenation into accepted language. The first v1 commission is preserved with its published hashes and explicitly superseded because its `coming` prompt was not sense-qualified. The controlling v1.1 contract SHA-256 is `b698bca492916fec56dc0f2bdf211f80363652d639fcc9a6b4cf6a366e9b4367`. Report SHA-256: `10dfc96e1a98b19aef81d63b4800049d69551311e55956cef783c6ea56bfef22`. Manifest SHA-256: `979ec403054352e79561eec98b91b46425feadf6b33b0523e3da21fdb303d048`. All checksums pass, and a clean rerun is byte-identical.

### Controlling execution order

1. Review the 36 novel cells against the bound dictionary forms, existing intransitive template, source-specific predicate semantics, and explicit subject compatibility decisions.
2. Resolve or omit the one child reference conflict and five woman cells. A blocked row cannot delay the clean core experiment and cannot be silently guessed.
3. Materialize one immutable intervention payload that separates unique evidence rows from optimizer replay. Preserve the ten clean going-away rows as retention; do not count them as new evidence.
4. Freeze three step- and token-accounted arms from identical v2 weights: untouched `B0`, retention-only continuation `C0`, and composition treatment `T1`. Pair row order, batch boundaries, learning-rate schedule, maximum updates, and decoder policy.
5. Run one screening seed on RunPod only after payload hashes, token counts, model/tokenizer identity, and development scoring are frozen.
6. Require improvement on held-out subject-predicate cells, no material loss on the complete lexical census, and no regression on retained sentence families. A lexical score cannot authorize sentence translation.
7. If the narrow screen fails, inspect cell-level substitutions and revise the task or conflict policy. Do not buy a larger run with 3,000 repetitions of a failed construction.
8. If it succeeds, confirm the recipe with seeds 17, 42, and 73, then activate additional grammar families from the 3,000-pair commission in evidence-backed batches.

The earlier RunPod block is now satisfied by the source review, paired schedule, corrected model-bound baseline, and sealed kit described below.

### 2026-07-31: source review issued the bounded internal treatment

- Reviewed every proposed composition cell against source-specific predicate semantics, subject compatibility, dictionary/audio convergence where available, and the existing controlled intransitive pattern.
- Issued 36 rows for one internal noncommercial screen, 13 corrected development rows, and ten retained going-away cells. Eleven cells remain blocked or excluded; no blocked row was silently guessed.
- Evidence grades are deliberately unequal: sitting has convergent curriculum, morphology, dictionary, and audio support; standing has convergent curriculum-sentence, dictionary, and audio support; deictic coming remains cross-era and source-scoped; running remains exploratory dictionary-plus-audio evidence.
- None of the 36 rows is speaker-attested, public reference gold, or sufficient for model promotion. Public-release-authorizing rows remain zero.
- Source-review contract SHA-256: `72255698524bf1e28bb26b41b4654a3fe4175284aa9bb94019d1fce141f39384`. Report SHA-256: `fe00a8ba28257d625a9d94ee5bda510798fa98679fddc3a9359edafbc450743a`.

### 2026-07-31: paired fixed-compute schedules frozen

Three rejected designs are preserved because they changed the causal question:

1. Same-subject control matching was 9.97% shorter in non-padding tokens and consumed five old holdout rows.
2. Independent nearest-row matching remained outside the 1% token gate.
3. Restricting controls to only 40 inline-reviewed rows could not reach the treatment target-token budget.

The controlling v1.3 schedule admits externally reviewed legacy rows only when both the sentence-review and lexeme-compatibility ledgers pass. It uses no old holdout row.

| Property | `C0` retention control | `T1` composition treatment |
| --- | ---: | ---: |
| Distinct rows | 36 | 36 |
| Presentations | 576 | 576 |
| Non-padding tokens | 10,592 | 10,592 |
| Optimizer updates | 18 | 18 |
| Physical batch / accumulation | 4 / 8 | 4 / 8 |
| Checkpoints | 3, 6, 9, 12, 15, 18 | 3, 6, 9, 12, 15, 18 |

Both arms start from the exact public adapter weight SHA-256 `05e6c40172d93b99538b3a21a04bbbd14a6c767180f85b83d5f00064a6858fb4`. They share seed 17, learning rate `2e-5`, warmup, batch boundaries, LoRA topology, and greedy decoder. The only intended difference is the treatment evidence.

Schedule contract SHA-256: `28103ce918fe123bd22bd17dd5cbe3382de79105dce555d0dd0ea59181185acb`. Report SHA-256: `c0faa943b45d5313c90fa9b03c71ea86b836af94896ebfae732b621a2e17b9dc`.

### 2026-07-31: corrected development baseline frozen

The immutable public Space was probed on all 13 source-qualified development cells:

| Measure | Baseline |
| --- | ---: |
| Exact sentence | 0 / 13 |
| Both required slots present | 0 / 13 |
| Required subject present | 7 / 13 |
| Required predicate present | 1 / 13 |
| Blank or source-copy fault | 0 |

Examples make the failure mechanism concrete:

| English | Reviewed development form | Public model |
| --- | --- | --- |
| The child is coming towards the speaker. | `Mayu yanajimanha.` | `mayu yanmanha` |
| The child is running. | `Mayu jamarnimanha.` | `Mayu yanmanha` |
| The child is sitting. | `Mayu nyinamanha.` | `Junda nyinarangamanha` |
| The dog is coming towards the speaker. | `Duthu yanajimanha.` | `Duthu yanmanha` |
| I am going away. | `Ngatha yanmanha.` | `Ngurragi yanmanha` |
| I am sitting. | `Ngatha nyinamanha.` | `Ngatha nyinarangamanha` |

The model often retains a plausible subject while collapsing the predicate to a familiar going-away or sitting-down frame. This is why the first v3 experiment targets controlled composition instead of adding undifferentiated epochs or thousands of unreviewed sentences.

Baseline contract SHA-256: `70f2e0a9e6533b6616c19cc763635ff9edb1a02669090801b464cd44f96e6c64`. Report SHA-256: `ef2e415f59f70e0e6c825c7e2b24060d5e07cdfda18209a614999053c1c1d19e`.

### 2026-07-31: sealed RunPod screen completed and rejected for promotion

- Built `wbv-v3-composition-screen-runpod-20260731-a1` around the exact public adapter, exact released tokenizer, immutable Wajarri base commit `ddae6103913b7cab299d687bee2dc7904fd59459`, all 3,188 regression rows, and the two fixed schedules.
- The driver evaluates untouched `B0`, trains `C0` and `T1`, scores every checkpoint on the 13 consumed development cells, then evaluates each selected arm over the complete frozen regression population.
- Treatment must beat both `B0` and `C0` on exact sentences or complete slot realization. It must also preserve one-target lexical exact within one percentage point, synthetic chrF++ within two points, fixed-utterance exact, and zero mechanical faults.
- The package passed 57 focused tests, Python compilation, every file checksum, and a static preflight reporting 13 development rows, 3,188 regression rows, 576 presentations per arm, and 18 updates per arm.
- Reproducible kit archive SHA-256: `e27969f63758d8631a1ccf99e2925651ee34965a4d5963fa5e52edfa55e7f463`.
- Launched RunPod pod `7pk5wnwgtazb0z` on one A40 at USD 0.44/hour. The first package failed before model loading because its unpinned runtime resolved to an incompatible future Torch/CUDA combination. Revision `r1` fixed the runtime but failed after untouched-baseline evaluation and before the first optimizer update because schedule rows omitted the model-visible `pair_kind` field. Both failures are preserved as preflight artifacts rather than hidden.
- Revision `r2` completed on Torch 2.8.0 and CUDA 12.8. The selected `C0` step-3 adapter has SHA-256 `d66329f6fc82f55085b15ffb1ae93baab9c6a2d94cca51d2224acb9cbcbdb457`; the selected `T1` step-18 adapter has SHA-256 `733b03d9e1ffb8a89082d880852d91b4f99c58b7152a03896b7450a9d1b8e710`.
- `T1` improved development exact from 0/13 to 3/13 and predicate presence from 1/13 to 7/13, but one-target lexical exact fell from 1,076/1,179 to 1,061/1,179. The experiment status is `NEGATIVE_OR_INCONCLUSIVE_INTERNAL_COMPOSITION_SCREEN`; promotion is false.
- The untreated baseline already contains one repeated-token diagnostic fault on a retired raw definition-context row. `C0` and `T1` retain that fault and each add a source-copy output for `telephone`. The original absolute-zero all-suite fault gate was therefore unreachable even for `B0`. The completed negative ruling is unchanged; future contracts separate deployment-relevant zero-fault suites from baseline-relative retired-context diagnostics.
- Recovered 86 retained manifest files and documented 31 omitted public-base files in `RECOVERY-MANIFEST.json`. Result archive SHA-256: `8aae443082322e8757a19d04ae632ea0d5f24556a1e6964eaa9eec9838aa8839`; result JSON SHA-256: `69c9589184e0190e8961352257a3a978afa029827aa5dc1d5b8f4c5fdfd82e07`; result-manifest SHA-256: `3670ef220ee4972a3c8a6f8f7d4774041c164ce1e5944bc20b919507abd63c5d`.
- Added a deterministic result analyzer that emits paired development changes, all lexical changes, sentence changes, mechanical-fault rows, diagnosis, and the next experiment requirements. Its report SHA-256 is `19788bb2079ddc39018a2d9faed6cd60dea1e23c5b70b4a176e80595d3867050`.
- Deleted pod `7pk5wnwgtazb0z` after artifact recovery. A final `runpodctl pod list` returned an empty list. There is no idle Wajarri GPU spend.

### 2026-07-31: fresh held-subject development matrix frozen

- Built 32 development-only cells from eight source-bound subjects crossed with four predicates: coming towards the speaker, running, sitting, and standing.
- The subjects are aunt `Maraji`, cousin `Manyu`, dingo `Ngubanu`, adolescent girl `Warlugura`, grandmother `Aba`, horse `Nguurru`, red kangaroo `Marlu`, and old man `Juja`.
- Every isolated subject target is reconstructed exactly by untouched `B0`; none of the 32 sentence pairs overlaps the old intervention train or development matrices; no training row was issued from this set.
- These are controlled source-composition diagnostics, not attested sentences, sealed test data, or evidence of naturalness. They can select an internal recipe once, after which they are consumed.
- Contract SHA-256: `075ae3055973129c1025d72ae5359cbdcb78f3837d48d234fd369c5615f6b438`. Development matrix SHA-256: `c2d3e596124f032f0b82bb19bf787e878c7c39eec4fb028aac83979768a2a4fb`. Subject-review SHA-256: `451a346a4e3c5977ba5b74a2602da9e73d9babbde1a4c93ef9635c088ee13ad3`.

### 2026-07-31: next mixed-replay schedules prepared

The next causal comparison is `B0` versus `C2` versus `T2`:

| Property | `C2` matched control | `T2` mixed treatment |
| --- | ---: | ---: |
| Presentations | 576 | 576 |
| Optimizer updates | 18 | 18 |
| Composition presentations | 288 | 288 |
| Lexical presentations | 288 | 288 |
| Unique composition rows | 36 | 36 |
| Unique lexical rows | 144 | 144 |
| Source non-padding tokens | 4,548 | 4,548 |
| Target non-padding tokens | 4,066 | 4,066 |
| Total non-padding tokens | 8,614 | 8,614 |

- `T2` presents every composition row eight times rather than 16 and interleaves two presentations of each of 144 distinct one-target lexical anchors.
- The anchor set includes all 23 exact dictionary rows lost by `T1`, excludes all eight fresh development subjects, and is dynamically stratified by observed tokenizer length. Closed-set reporting must separate all 1,179 prompts, the 144 trained anchors, and the 1,035-row unanchored complement.
- `C2` is matched to `T2` on task counts, update count, source tokens, target tokens, and total non-padding tokens. The first aggregate-only attempt matched total tokens but shifted 62 tokens from target to source; it is preserved under the explicit `superseded-source-target-imbalance` name and cannot be executed.
- Canonical contract SHA-256: `b4fc887a8422e0fb41e550a517fb25a9dfcabdfe95b0e6a5c662a8c3bcf3696b`. `C2` schedule SHA-256: `32f683dd61a385ec0a0f803d959956d5f0e1ec8ecc0260b44ee64a5dfccee3ac`. `T2` schedule SHA-256: `2bc2b43f6fa9345039e0af7cbd19ae5ef5753822961a5c34c70c9efb4347aa97`. Token-accounting SHA-256: `05bd834ef8979af99ff7ba439de708c5baf43510de23c00f490f038c76e0edcd`.
- The builders, result analyzer, driver policy changes, and related Wajarri pipeline now pass 65 focused unit tests and Ruff. Every generated artifact checksum passes.
- RunPod execution is deliberately still blocked. The remaining preflight work is to evaluate untouched `B0` on the fresh 32-cell matrix, extend the driver to report anchored and unanchored lexical populations separately, freeze the corrected mechanical-gate policy, build a new immutable kit, and rerun its static preflight. Only then is a paid GPU screen justified.

### 2026-08-01: lexical regression localized and static GPU kit passed

The new partition analyzer was run against the recovered `B0`, `C0`, and `T1` predictions before designing any additional data. The 144-anchor population was defined to contain all 23 `B0` to `T1` losses, so its role is diagnostic and preventative rather than a random population estimate.

| One-target partition | `B0` exact | `C0` exact | `T1` exact |
| --- | ---: | ---: | ---: |
| All one-target prompts | 1,076 / 1,179 | 1,072 / 1,179 | 1,061 / 1,179 |
| Selected lexical anchors | 144 / 144 | 140 / 144 | 121 / 144 |
| Mandatory prior-regression anchors | 23 / 23 | 19 / 23 | 0 / 23 |
| Unanchored complement | 932 / 1,035 | 932 / 1,035 | 940 / 1,035 |
| Matched lexical controls | 144 / 144 | 144 / 144 | 144 / 144 |

This changes the failure diagnosis. `C0` did not reduce exact reconstruction anywhere in the 1,035-row unanchored complement; its four losses all belonged to the sensitivity set. `T1` improved the unanchored complement by eight exact rows while losing all 23 mandatory anchors. The aggregate decline was therefore concentrated in a small continuation-sensitive population rather than distributed across the lexicon. That is evidence for interleaving direct replay of the damaged mappings. It is not evidence that the 23 words are a natural statistical class, because they were selected from the observed losses.

The `T2` treatment now receives two presentations of each of 144 anchors while its composition dose is halved from 16 to eight presentations per unique composition row. `C2` receives the same number of lexical and sentence presentations and is matched exactly on source, target, and total non-padding tokens. This directly tests whether targeted lexical replay preserves the sensitive mappings while retaining the sentence-composition signal.

Implementation and preflight changes:

- Generalized the development evaluator from a hardcoded 13-row assumption to a required contract-bound count and retained fail-closed duplicate, schema, and two-slot validation.
- Added explicit result partitions for all one-target rows, trained anchors, matched controls, the unanchored complement, and mandatory prior-regression anchors.
- Replaced hardcoded `C0`/`T1` driver roles with contract-declared control and treatment arms.
- Added separate exact-retention gates for the unanchored complement, all 144 anchors, and all 23 mandatory anchors.
- Corrected mechanical policy: public-serving suites require zero faults; the retired definition-context diagnostic may contain inherited faults but may not worsen relative to same-run `B0`.
- Preserved the public v2 adapter as the exact starting point and kept automatic model promotion false.

The immutable preflight kit is `experiments/kits/wbv-v3-mixed-replay-screen-preflight-20260801-a1`. It contains 37 checksum-bound files and is 34 MB compressed. Static preflight reports 32 development rows, 3,188 regression rows, 576 presentations per arm, and 18 optimizer updates per arm. All 78 focused Wajarri tests pass, Ruff passes, every kit checksum passes, and the archive passes gzip and tar integrity checks.

| Artifact | SHA-256 |
| --- | --- |
| Build contract | `9ee5f3045c5610bd2e51886a093dca0da3e19e658435d1f1a1e22cfa6015b2b7` |
| Immutable kit contract | `68b311c1b9ea0fd4d65819f6167e62b2f49de276c5a7dda54fe9cf5e93c7f3b0` |
| Kit checksum ledger | `c3724fc6f133b5f0b5af01aa3e6bab9e44b4264b8a1d89be7db80388b86c627a` |
| Reproducible kit archive | `99b0855a6b94fc27f07ed22b7c6618888766a8e70fedb53129ad6f274d5c1a51` |
| `B0` lexical partition report | `1cbdb71029a4c96a13a0b967eea1c21d61d16a11b2918d2a1272df9f41ac1dab` |
| `C0` lexical partition report | `2a1bfb65fc38dfa133fd8fb8ac5eac43a5fc2bb6b59db1936c25f0cd0af129cf` |
| `T1` lexical partition report | `6cb034fe74ff8d33171134bd9a3ef4814d4c0f440f39684b15405a9d70ddf79a` |

The kit is intentionally not executable on paid compute. Its immutable contract contains `paid_compute_authorized_by_user=false`, and a tested bootstrap interlock exits before checking CUDA or creating an output directory. No RunPod pod is active. The next operational step, if paid execution is approved, is to reissue an authorized derivative of this exact preflight kit, verify that only the authorization/run identity changed, run one A40 seed, recover every result, and delete the worker immediately.

### 2026-08-01: `C2`/`T2` mixed-replay A40 screen completed and rejected

The previously prepared comparison was authorized, executed as `wbv-v3-mixed-replay-screen-runpod-20260801-a1-r2`, recovered, and shut down. Both arms began from the immutable public Wajarri v2 adapter and received 576 presentations over 18 optimizer updates. The selected control was `C2` step 18; the selected treatment was `T2` step 15.

| Fresh 32-cell development measure | `B0` | `C2` | `T2` |
| --- | ---: | ---: | ---: |
| Exact sentence | 0 | 0 | 4 |
| Both required slots present | 0 | 0 | 4 |
| Required subject present | 12 | 12 | 12 |
| Required predicate present | 0 | 0 | 11 |
| Mean chrF++ | 35.95 | 36.93 | 54.42 |
| Repeated-token faults | 2 | 0 | 0 |

The treatment effect was real and progressive, rather than a single lucky checkpoint:

| `T2` update | Exact | Both slots | Predicate present | Mean chrF++ | Repeated-token faults |
| ---: | ---: | ---: | ---: | ---: | ---: |
| 3 | 1 | 1 | 2 | 38.78 | 2 |
| 6 | 2 | 2 | 4 | 42.42 | 2 |
| 9 | 3 | 3 | 8 | 49.02 | 1 |
| 12 | 3 | 3 | 9 | 51.62 | 0 |
| 15 | 4 | 4 | 11 | 54.42 | 0 |
| 18 | 4 | 4 | 11 | 53.96 | 0 |

The corresponding `C2` checkpoints remained at zero exact and zero expected predicates throughout. Direct composition supervision therefore changes sentence behavior; ordinary matched continuation does not explain the gain.

The full regression result nevertheless rejects `T2`:

| Full measure | `B0` | `C2` | `T2` |
| --- | ---: | ---: | ---: |
| All one-target lexical exact | 1,076 / 1,179 | 1,074 / 1,179 | 1,067 / 1,179 |
| Trained-anchor exact | 144 / 144 | 141 / 144 | 129 / 144 |
| Mandatory sensitive-anchor exact | 23 / 23 | 20 / 23 | 8 / 23 |
| Unanchored-complement exact | 932 / 1,035 | 933 / 1,035 | 938 / 1,035 |
| Controlled-synthetic exact | 13 / 36 | 11 / 36 | 13 / 36 |
| Controlled-synthetic chrF++ | 70.90 | 71.06 | 71.12 |
| Complex historical exact | 0 / 5 | 0 / 5 | 0 / 5 |
| Complex historical chrF++ | 9.62 | 9.84 | 10.27 |
| Fixed-utterance retention | 55 / 55 | 55 / 55 | 55 / 55 |

`T2` gained seven exact lexical rows, lost 16, and retained 1,060, for a net change of minus nine. It passed the narrow sentence-effect test but failed all-one-target, trained-anchor, mandatory-anchor, and retired-diagnostic noninferiority. Public-serving suites had zero mechanical faults. The formal status is `NEGATIVE_OR_INCONCLUSIVE_INTERNAL_COMPOSITION_SCREEN`; promotion is false. Neither selected adapter was uploaded or served, and the public Space remains on Wajarri v2.

| Completed-run artifact | SHA-256 |
| --- | --- |
| `C2` selected adapter | `dd65431ee49d53591d47a00f0ccc5f88a0ac48ab2ebd65e16c82057cdd37a78f` |
| `T2` selected adapter | `b31a785ac28456137a3059f092a4930a9ab7c256b348bd3c81dcfe2fc2dc4bc3` |
| Result | `6d5d33b3b2354c42d8c4994d04e784d7d1fd35810a3532d7907ce256347c7ef8` |
| Result manifest | `02cb61eff158b0aba8849cd23a87fe8f20f133d150cf8ad570ddbce14342f07f` |
| Run-complete record | `1b47965c99f6287eee9567143961a6d7d00102041fb9054b4472f4bcfa82a022` |
| Recovery manifest | `3e161119d5becf5c65d8a6605f5002d21e1b40a554fd5dd7dbd9a8ac37954f31` |
| Post-hoc analysis report | `496b3e9d08a867f579111145bb570300b66993d386a8139e5ec9aca04d2ae2b2` |

The A40 worker `1vgczur074a6k9` was deleted after recovery. A fresh `runpodctl pod list --output json` on 2026-08-01 returned `[]`.

### 2026-08-01: qualitative failure localization

The lexical losses are not uniform. All 31 one-token anchor targets survived. Losses occurred in 3 of 35 two-token targets, 9 of 44 three-to-four-token targets, and 3 of 34 targets with five or more model tokens. Token length contributes to sensitivity, but it does not explain the wrong-known-word substitutions or derivational flattening.

All 15 `T2` losses from the 144 baseline-exact anchors belong to the 23-row mandatory sensitivity set:

| English prompt | Governed target | `T2` output | Source record |
| --- | --- | --- | --- |
| shoes | `jinabuga` | `jinabu` | `wbv-src-local-001279` |
| small brown venomous snake | `jidu` | `jirdu` | `wbv-src-local-001356` |
| climbing up or into | `garlbajunmanha` | `garlbajimanha` | `wbv-src-local-000318` |
| feeling sorry | `gurninyimanha` | `gurnimanha` | `wbv-src-local-000536` |
| spoiling | `walhiyimanha` | `walhimanmanha` | `wbv-src-local-001398` |
| bush | `buuju` | `bariya` | `wbv-src-local-000266` |
| playfully | `bunthanthu` | `bunthanyimanha` | `wbv-src-local-001099` |
| boil | `gurraburda` | `gurrgurda` | `wbv-src-local-000214` |
| becoming full | `ngumulimanha` | `mimigimanha` | `wbv-src-local-000605` |
| swelling | `yurayimanha` | `yurayimanmanha` | `wbv-src-local-001462` |
| spreading out | `garrayijunmanha` | `garrimanmanha` | `wbv-src-local-001401` |
| straightening | `thubarnmanha` | `thubarnimanha` | `wbv-src-local-001432` |
| making a pile of sand | `budirimanmanha` | `budimanmanha` | `wbv-src-local-001066` |
| sitting something | `nyinajunmanha` | `nyinamanha` | `wbv-src-local-001314` |
| doing in vain | `widigunmanha` | `widimanha` | `wbv-src-local-000460` |

Every one of these 15 source records has a decoded speaker-attributed word recording, but every form is still marked `morphologicalAnalysisStatus=unanalysed` and none has a governed sentence example in dictionary edition `wajarri-dictionary-50words-attested-v1.2.0`. It would therefore be unsound to invent sentence contexts or morphological paradigms for them merely to protect a benchmark score. The immediate protection mechanism is checkpoint-level lexical sentinels and direct replay. Sentence generation for these forms remains blocked until a source supplies their syntactic and morphological use.

The fresh composition errors are also structured:

- Standing `garrimanha` is the positive control: the selected treatment produced the expected predicate in all 8 cells and exact sentences in 3.
- Running `jamarnimanha` was present in 3 of 8 cells and exact in 1.
- Sitting `nyinamanha` was present in 0 of 8 cells; the model often substituted the source-distinct sitting-down form `nyinarangamanha`.
- Coming-towards `yanajimanha` was present in 0 of 8 cells; the model often substituted going-away `yanmanha`.
- Aunt `Maraji`, dingo `Ngubanu`, and grandmother `Aba` yielded 4 of 4 expected subject forms; adolescent girl `Warlugura`, cousin `Manyu`, horse `Nguurru`, old man `Juja`, and red kangaroo `Marlu` yielded 0 of 4 exact treatment sentences.

This points to two separable problems: preservation of morphologically complex lexical forms during continuation, and controlled selection of source-qualified predicates and subjects inside a sentence. More repetitions of the current templates would conflate them.

### 2026-08-01: paired public-Space glossary probe

A deterministic builder derived 36 training and 32 development glossary-input variants from existing reviewed rows. The target side is byte-identical to the corresponding plain row: this adds 68 input variants and **zero new Wajarri target sentences**.

The public CPU Space was then probed sequentially with each plain development prompt and its glossary-conditioned partner. All 64 responses came from the expected immutable Wajarri v2 model identity; there were no retries, queue waits, blank outputs, or source copies. CPU float32 results are kept separate from the A40 BF16 experiment.

| Paired 32-cell measure | Plain v2 | Glossary-conditioned v2 | Net gain |
| --- | ---: | ---: | ---: |
| Exact sentence | 0 | 6 | +6 |
| Both required slots present | 0 | 6 | +6 |
| Required subject present | 12 | 23 | +11 |
| Required predicate present | 0 | 9 | +9 |

All six exact gains were sitting clauses: adolescent girl, aunt, cousin, dingo, grandmother, and red kangaroo. The same prompt structure still produced `Ngaru nyinamanha` for expected `Nguurru nyinamanha` and `jarda nyinamanha` for expected `Juja nyinamanha`. Glossary conditioning is therefore a useful retrieval intervention, not a solved translator. It improved the source slot and one predicate family without teaching the deictic-motion, running, or standing contrasts reliably.

The probe made 64 requests in 306.30 wall seconds and reported 227.885 model seconds. Its contract SHA-256 is `07abcd5d10cd895e5d4280f8e359b21695e4f43db8eae5c504404328e006e464`; report SHA-256 is `5af517a2dcb9fe0150b3e31f231030430a5d390b49b920dc168392b7e1cee703`; manifest SHA-256 is `b71a4c516aeff0d0ea3fa000403a94335bede35c967f17a459a788d7309fbc5e`. The glossary-fixture contract SHA-256 is `5c600ff7c9259626c9736c3b0eb49dfaa3d71d1de8b36ff1e9376252d6a3af06`; its report SHA-256 is `4f81e8618164197c3bc0b38bbcfceee12ac301c10fb1d3e86d8458f08471102f`. Both checksum ledgers pass.

### New sentence requirements before the next training stages

The phrase “generate more synthetic sentences” now resolves into four gated stages. Counts below are unique English-Wajarri pairs, not optimizer presentations, prompt mirrors, paraphrase counts, or independent speaker attestations.

| Stage | New Wajarri target pairs | What is required | Gate before proceeding |
| --- | ---: | --- | --- |
| A: checkpoint-sentinel rerun | **0** | Re-run the existing matched `C2`/`T2` schedules while measuring all 144 baseline-exact anchors and all 23 mandatory anchors at every checkpoint; also measure plain and glossary development conditions | An admissible checkpoint must retain all protected anchors and beat `B0` and `C2` on sentence composition |
| B: evidence acquisition | **0 synthetic pairs** | Find source-backed usage and morphological analyses for the 15 sensitive forms above; adjudicate `yanajimanha` versus `yanmanha`, `nyinamanha` versus `nyinarangamanha`, and exact running/subject forms | No form may enter sentence generation from an isolated gloss plus an unanalysed headword alone |
| C: first contrast commission | **48-96** | 16-32 deictic-motion contrasts, 16-32 sitting-versus-sitting-down contrasts, 8-16 exact running contrasts, and 8-16 subject/orthography contrasts | Every target must be licensed by an activated source rule and lexeme; split whole subject and morphology families before generation; reject if lexical sentinels or retained sentences regress |
| D: broader factorial expansion | **80-160** | Cross 20-40 vetted animate subjects with the four source-qualified predicates; glossary mirrors may be added as input variants but do not count as new target pairs | Only after Stage C improves held-family development and passes full lexical, sentence, degeneration, and provenance gates |

Standing should remain a lightly represented positive control rather than receive indiscriminate new volume. The previously designed 3,000-pair corpus is retained only as an expansion ceiling after Stages A-D establish that the intervention generalizes. There is currently no evidence-based reason to generate 20,000 Wajarri sentences.

For every issued pair, the repeatable language-program contract requires:

1. exact source IDs for every lexical form and grammar rule;
2. a declared construction family, morphology features, subject family, and variety status;
3. complete-form targets rather than generated suffix strings;
4. group-level train/development/test assignment before paraphrasing or deduplication;
5. separate counts for attested, research-synthesis, glossary-input, and speaker-reviewed evidence;
6. deterministic lint for source binding, slot completeness, contamination, duplicates, orthography, and forbidden extrapolation;
7. a frozen model, tokenizer, decoder, row, token, update, and learning-rate contract before GPU use;
8. per-checkpoint lexical sentinels so a later checkpoint cannot be selected after silently erasing known forms.

### 2026-08-01: checkpoint-sentinel preflight frozen

The next diagnostic driver now evaluates untouched `B0` and every `C2`/`T2` checkpoint at updates 3, 6, 9, 12, 15, and 18 on:

- the 32 plain held-subject development cells;
- their 32 input-only glossary variants;
- all 144 baseline-exact lexical sentinels;
- the nested 23-row mandatory sensitivity set.

A checkpoint is admissible only if it has zero sentinel mechanical faults, loses no baseline-exact anchor, and retains all 23 mandatory anchors. Development score may choose only among admissible checkpoints. If none exists, the run emits `NO_ADMISSIBLE_CHECKPOINT_DIAGNOSTIC_ONLY` and preserves diagnostic adapters without pretending to have selected a candidate.

The immutable preflight-only kit is `experiments/kits/wbv-v3-checkpoint-sentinel-preflight-20260801-a1`. It contains 44 checksum-bound files plus the checksum ledger, 41,964,483 uncompressed bytes, and a 34,804,052-byte reproducible archive. Static preflight confirms 32 plain development rows, 32 glossary development rows, 3,188 full-regression rows, 576 presentations per arm, and 18 updates per arm.

| Next-kit artifact | SHA-256 |
| --- | --- |
| Build contract | `14ba16b81671fc05360c0060efc3da61b0077ec2107b946a8959566db0e8891b` |
| Immutable kit contract | `5b6090f9508e1efe543f78cba3ba161a49a4fd09d0aa38f19f6a94d11a8df2eb` |
| Kit checksum ledger | `bd39a9ef20cebc96cddbfc28970c597f1dba0dff8ae41e3c3d89d6e0c066712a` |
| Reproducible kit archive | `d74df70cb34b927a27bfc2c1bad84259550c66fde6c6b191491934f75092047d` |

All 97 focused Wajarri v3 unit tests pass, focused Ruff checks pass, the bootstrap passes `bash -n`, all 44 bound checksums pass, and both gzip and tar archive checks pass. A local bootstrap test verified every checksum and then refused with `Paid GPU execution is not authorized by this immutable contract`; it created no output directory. The contract remains `paid_compute_authorized_by_user=false`, and a final live RunPod check still reports no pods.

This section supersedes the earlier preflight’s next-step wording. The next paid action is not another broad epoch run and is not currently authorized. If separately authorized, it is one bounded A40 checkpoint-sentinel diagnostic from this exact kit, followed by immediate result recovery and worker deletion. Its outcome decides whether Stage C sentence commissioning is warranted.

### 2026-08-01: authorized checkpoint-sentinel A40 run completed and rejected

The user authorized the bounded next RunPod experiment. An authorized derivative was built without changing the frozen scientific design. Relative to the preflight contract, only creation time, run identity, authorization state, and the paid-compute Boolean changed; normalized remainder hashes matched before execution.

Three secure A40 allocations never reached nonzero uptime or exposed SSH and were deleted without running the workload: `igklrbz482bypv`, `6vaeouwg0enqxf`, and `8fp4wszfenxwms`. The successful worker was `re80p6zctrxdtv`, an A40 in `EU-SE-1` started from the official `runpod-torch-v280` template and exact image `runpod/pytorch:1.0.2-cu1281-torch280-ubuntu2404`. Admission verified an NVIDIA A40 with 49,140 MiB VRAM, PyTorch `2.8.0+cu128`, CUDA 12.8 availability, more than 160 TB free in the provider workspace, and every one of the 44 kit checksums before model loading. The worker rate was USD 0.44/hour. It was retained for roughly 30 minutes including result packaging and transfer, implying approximately USD 0.22 before provider rounding.

The run evaluated untouched `B0`, every `C2` and `T2` checkpoint at updates 3, 6, 9, 12, 15, and 18, the 32-cell plain development matrix, the paired 32-cell glossary-input matrix, all 144 baseline-exact lexical sentinels, all 23 mandatory sensitive sentinels, and the 3,188-row full regression bundle.

#### Complete checkpoint census

| Arm | Update | Admissible | Plain exact | Plain predicate | Plain chrF++ | Glossary exact | Glossary predicate | Sentinel exact | Mandatory exact |
| --- | ---: | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| `C2` | 3 | **yes** | 0 / 32 | 0 / 32 | 35.83 | 6 / 32 | 10 / 32 | 144 / 144 | 23 / 23 |
| `C2` | 6 | no | 0 / 32 | 0 / 32 | 35.86 | 6 / 32 | 11 / 32 | 143 / 144 | 22 / 23 |
| `C2` | 9 | no | 0 / 32 | 0 / 32 | 36.75 | 4 / 32 | 10 / 32 | 142 / 144 | 21 / 23 |
| `C2` | 12 | no | 0 / 32 | 0 / 32 | 36.91 | 4 / 32 | 10 / 32 | 140 / 144 | 19 / 23 |
| `C2` | 15 | no | 0 / 32 | 0 / 32 | 36.91 | 5 / 32 | 11 / 32 | 141 / 144 | 20 / 23 |
| `C2` | 18 | no | 0 / 32 | 0 / 32 | 36.93 | 4 / 32 | 10 / 32 | 141 / 144 | 20 / 23 |
| `T2` | 3 | **no** | 1 / 32 | 2 / 32 | 38.78 | 7 / 32 | 12 / 32 | 141 / 144 | 20 / 23 |
| `T2` | 6 | no | 2 / 32 | 4 / 32 | 42.42 | 7 / 32 | 13 / 32 | 138 / 144 | 17 / 23 |
| `T2` | 9 | no | 3 / 32 | 8 / 32 | 49.02 | 7 / 32 | 13 / 32 | 135 / 144 | 14 / 23 |
| `T2` | 12 | no | 3 / 32 | 9 / 32 | 51.62 | 7 / 32 | 13 / 32 | 130 / 144 | 9 / 23 |
| `T2` | 15 | no | 4 / 32 | 11 / 32 | 54.42 | 8 / 32 | 15 / 32 | 130 / 144 | 9 / 23 |
| `T2` | 18 | no | 4 / 32 | 11 / 32 | 53.96 | 8 / 32 | 15 / 32 | 129 / 144 | 8 / 23 |

The `C2` selector correctly chose update 3, the only admissible control checkpoint. No `T2` checkpoint was admissible, so its update-3 adapter is explicitly retained under `diagnostic-adapters/T2` with selection status `NO_ADMISSIBLE_CHECKPOINT_DIAGNOSTIC_ONLY`. The later four-exact treatment checkpoint was not allowed to conceal its 14 sentinel and 14 mandatory-anchor losses.

#### Selected-checkpoint full regression

| Full measure | `B0` | `C2` step 3 | `T2` step 3 diagnostic |
| --- | ---: | ---: | ---: |
| One-target lexical exact | 1,076 / 1,179 | 1,076 / 1,179 | 1,072 / 1,179 |
| Trained-anchor exact | 144 / 144 | 144 / 144 | 140 / 144 |
| Mandatory-anchor exact | 23 / 23 | 23 / 23 | 19 / 23 |
| Unanchored-complement exact | 932 / 1,035 | 932 / 1,035 | 932 / 1,035 |
| Controlled-synthetic exact | 13 / 36 | 13 / 36 | 13 / 36 |
| Controlled-synthetic chrF++ | 70.90 | 72.67 | 70.95 |
| Complex historical exact | 0 / 5 | 0 / 5 | 0 / 5 |
| Complex historical chrF++ | 9.62 | 9.62 | 9.84 |
| Fixed-utterance retention | 55 / 55 | 55 / 55 | 55 / 55 |

Both continued adapters passed public-serving mechanical checks, one-target and unanchored tolerance, controlled-synthetic tolerance, and fixed-utterance retention. `T2` failed trained-anchor and mandatory-anchor retention. Both `C2` and `T2` also failed retired definition-context diagnostic noninferiority: `B0` already repeats `winyjinyji` for the green-cricket prompt, while both continuations additionally copy the English source `telephone` instead of producing governed `wayangunda`. Thus `C2` is a useful safety-preserving control, not a release.

The treatment’s earliest protected losses were not random short words:

| Prompt | Governed form | `T2` step-3 output |
| --- | --- | --- |
| bush | `buuju` | `bariya` |
| boil | `gurraburda` | `gurrgurda` |
| Sturt's Desert Pea | `marluguru` | `marlba-marlba` in the full 1,408-row lexical run |
| doing, in vain | `widigunmanha` | `widimanha` |

The Sturt's Desert Pea row exposes an additional release blocker. The 144-row sentinel invocation returned correct `marluguru`, but the full lexical evaluator returned `marlba-marlba` from the same compact adapter weight SHA-256. The evaluation order and padding context differ even though the decoder is deterministic. This is an observed batch-shape invariance failure until reproduced and explained; all future release candidates must return the same normalized output for every gate row at batch sizes 1, 8, 16, and 64.

Qualitatively, the treatment learns a few cells while leaving the key contrast failures intact:

| English | Governed reference | Public `B0` | `T2` step 3 | `T2` step 15 diagnostic |
| --- | --- | --- | --- | --- |
| The aunt is standing. | `Maraji garrimanha.` | `Maraji garrayi-garrayi` | `Maraji garrimanha.` | `Maraji garrimanha.` |
| The adolescent girl is coming towards the speaker. | `Warlugura yanajimanha.` | `Muthura yanmanha.` | `Muthura yanmanha.` | `Muthura yanmanha.` |
| The horse is sitting. | `Nguurru nyinamanha.` | `Dundulyji nyinarangamanha.` | `Dundulyji nyinarangamanha.` | `Dundulyji nyinarangamanha.` |
| The red kangaroo is standing. | `Marlu garrimanha.` | `Burun jardu jarlgu nyinya.` | `Burun jardu jarlgu nyinya.` | `Jalgu garrimanha.` |

This is narrow matrix fitting, not broad sentence translation. At step 15 the four plain exact cells are aunt-standing, dingo-standing, grandmother-running, and grandmother-standing. Coming-towards remains 0 of 8 at every treatment checkpoint. Sitting remains confused with sitting-down in the plain condition. Glossary conditioning helps more than unconditioned generation, but even the most fitted checkpoint is only 8 of 32 exact on the consumed matrix.

#### Recovery and shutdown evidence

The remote result manifest bound 204 files. Every checksum passed on the worker, the deterministic result archive passed gzip integrity, its local SHA-256 matched the remote value, extraction succeeded on `/mnt/donto-data`, and all 204 manifest checksums passed again locally. Only compact selected/diagnostic adapters and complete evaluations were retained; optimizer state and unselected checkpoint weights were deleted under the run's deletion ledger.

| Artifact | SHA-256 |
| --- | --- |
| Authorized build contract | `3d48e09bdeccf228037b42848c8dd189778ca7da8f87ed335ad02fdf68ba97dc` |
| Authorized immutable kit contract | `868e992be46ecadb4756a638b4be23d6f01c0c70f8073b19e27841498f27e48a` |
| Kit checksum ledger | `258a50238e90e1992738ea4d541cd0472b0f38bf179a48523484e8541bd8ceef` |
| Kit archive | `ee985a8a52dded9094410989294599aadacf0121122d95a63758c8918b0d2295` |
| `C2` compact adapter | `f78c4bbb129ff74406838187985e02037e2f70ca905ee264517d4003e03660b0` |
| `T2` diagnostic adapter | `320ce95d175829204586de806fe5dc28112c37fb9d51d5e2c3d9c5e36ee9e155` |
| Result | `83206aaa848b4112639c9918190e42307c92df17a65204484041da3512116a75` |
| Result manifest | `a4b909106c4e3798f94087b2d76caa932104c1b549e70303a572baafc904d53d` |
| Run-complete record | `54e3ebb5cc6485b64e7be2b15956b0b1b1f158ff3f3070051672f51bcd55b3fe` |
| Recovered 1,615,084,259-byte archive | `2e29604a2c6dea5cfbee84730f7f8bdc58eea42ce2a4dd1a55b0f892f7429d44` |

The canonical recovered directory is `experiments/results/wbv-v3-checkpoint-sentinel-runpod-20260801-a1/`. Worker `re80p6zctrxdtv` was deleted after local verification, and a fresh `runpodctl pod list --output json` returned `[]`.

#### Hugging Face ruling and live state

No new Wajarri adapter was uploaded. The public model remains `ajaxdavis/mobtranslate-wajarri-v2` at repository commit `357a7e5bde96990a3d5131bca12da6dacce98b3a`, model version `v2.0-context-guarded-nr4-s17-context-recovery-r16-s20-20260731`. The shared Space `ajaxdavis/alpha-v0-historic` remains on CPU Basic at commit `950b261a7fe6304b94aa63d03b7f0126cc8bbf15` and advertises the same immutable Wajarri identity. A live model-bound request for `The aunt is standing.` returned `Maraji garrayi-garrayi`, not governed `Maraji garrimanha.`. This confirms that hosting works and sentence translation remains unsolved.

#### What remains before another model can be released

The next work is not another epoch run and not a 20,000-row blind expansion:

1. Complete source-backed use and morphological analyses for the sensitive forms, beginning with the four update-3 full-regression losses and the earlier 15-row loss set.
2. Commission **48-96 new governed target pairs**: 16-32 coming-towards versus going-away contrasts, 16-32 sitting versus sitting-down contrasts, 8-16 running contrasts, and 8-16 subject/orthography contrasts. Glossary-input mirrors remain input variants and do not count as new Wajarri targets.
3. Keep the public v2 lexical adapter immutable. Build a separate `<translate>` sentence adapter under the same frozen base and tokenizer, while dictionary lookup and `<lexeme>` reconstruction continue through deterministic lookup or v2. One Hugging Face Space can route both adapters under one API without asking one adapter to optimize incompatible tasks.
4. Train the sentence adapter only after family-level splits are frozen. Compare plain and retrieved-glossary inputs at fixed token exposure, checkpoint every one to two updates, and include batch-size invariance in every checkpoint gate.
5. Use the entire 1,179-row one-target lexical suite as a diagnostic and preservation/distillation surface, not only 144 matched sentinels. A sentence adapter may be task-routed, but sentence outputs must still preserve required governed lexical slots.
6. Require a held-family sentence gain, zero deployment faults, stable outputs across batch sizes, synthetic and fixed-utterance noninferiority, and source-backed natural review before opening a final test or uploading a new public adapter.
7. Only after the first 48-96-pair intervention generalizes may the program commission the 80-160 broader factorial set. The 3,000-pair corpus remains a later ceiling; there is no evidence for generating 20,000 Wajarri pairs now.

The shortest technically defensible path to a useful Hugging Face system is therefore a task-routed lexical-plus-sentence model service backed by new source-governed contrasts. The current experiment proved that additional updates on the existing matrix cannot produce that system safely.

### 2026-08-02: copy-token screen recovered; subject-slot architecture issued

Subsequent controlled screens did not justify expanding unsupported Wajarri morphology. The strongest sentence continuation, `D6`, improved the bounded composition matrix but still rewrote copied dictionary forms. `T7` then made `<copy>` a dedicated single tokenizer token and task control. That intervention reduced neither the underlying one- and two-character rewrites nor the lexical regression enough to pass.

| Measure | Untouched `B0` | `D6` | `T7` selected update 480 |
| --- | ---: | ---: | ---: |
| Plain composition exact | 0 / 11 | 9 / 11 | 9 / 11 |
| Inline held-subject exact | not selected | 21 / 24 | 22 / 24 |
| Neutral single-copy exact | not selected | 48 / 64 | 48 / 64 |
| Neutral dual-copy exact | not selected | 24 / 64 | 25 / 64 |
| Fixed-utterance retention | 55 / 55 | 54 / 55 | 54 / 55 |
| Controlled-synthetic exact | 13 / 36 | 23 / 36 | 23 / 36 |
| Controlled-synthetic chrF++ | 70.9009 | 75.7673 | 75.0276 |
| Direct lexical exact | 1,099 / 1,408 | not selected | 838 / 1,408 |
| Definition-context lexical exact | 877 / 1,684 | not selected | 590 / 1,684 |

The T7 adapter is an internal starting point, not a release. Its selected weight SHA-256 is `02e04f84ae261bfb7fcfca4a05e221b0e2982036496ef9a8f311a526c7446d9c`. It had 13 direct-lexical and 47 context-lexical mechanical faults. The model remained 0 of 5 exact on the historical diagnostic. No T7 weight was uploaded, registered, mounted, or routed to the homepage.

#### Evidence frontier re-audited

The primary-source audit found no machine-readable Wajarri sentence collection that could responsibly turn the unsupported `-manha` inventory into hundreds of training clauses:

- PARADISEC's catalogue API exposed the RUIL 50 Words item but no larger Wajarri sentence corpus suitable for this run.
- AIATSIS finding aids identify substantial archival Wajarri recordings, but item-specific access, copyright, depositor, and community permissions may apply. They are acquisition leads, not silent training data.
- Douglas's 1981 *Watjarri* grammar is an essential diagnostic source but is all-rights-reserved in the available catalogue record.
- ASJP supplies a small CC BY 4.0 lexical list, not sentence grammar; the available Wiktionary inventory is also only lexical.

The complete-predicate frontier therefore remains six source-bound surfaces: `yanajimanha`, `yanmanha`, `jamarnimanha`, `nyinamanha`, `nyinarangamanha`, and `garrimanha`. The program explicitly rejected suffix-based generation over the remaining unanalysed headwords. More row count would not repair missing linguistic evidence.

#### Deterministic subject-slot intervention

The next intervention removes the part T7 could not perform reliably. The model no longer regenerates a dictionary subject. It emits one literal slot plus one complete governed predicate:

```text
model output:       <copy> jamarnimanha.
dictionary binding: <copy> = marruwa
rendered output:    marruwa jamarnimanha.
```

The renderer fails unless there is exactly one `<copy>` marker in the first output position. It performs one deterministic replacement and does not alter the predicate. Two paired arms isolate how the slot is declared:

| Arm | Model-visible source | Raw target |
| --- | --- | --- |
| `M8` masked source | `<translate> The <copy> is running.` | `<copy> jamarnimanha.` |
| `D8` declared source | `<translate> The bilby is running. <glossary> bilby = <copy>` | `<copy> jamarnimanha.` |

The commission contains 103 training parent pairs, 11 composition-development parents, and 24 held-subject-development parents. It creates 206 training representations and 70 slot-development representations but **zero new Wajarri sentences**. `M8` collapses to six unique model-visible pairs, one per complete predicate; `D8` preserves 103 distinct English-subject inputs. No sealed-test row was read.

Each arm has exactly 7,680 scheduled presentations and 480 optimizer updates. Every update contains four fixed-utterance retention rows, four plain sentence rows, four inline sentence rows, and four subject-slot rows. Non-slot presentation positions are paired between arms. Token exposure is reported rather than assumed equal:

| Arm | Source tokens | Target tokens | Total non-padding tokens |
| --- | ---: | ---: | ---: |
| `M8` | 97,597 | 65,884 | 163,481 |
| `D8` | 113,290 | 65,880 | 179,170 |

The 140-row consumed development screen has 11/11/24/24 ordinary plain/inline composition and held-subject rows plus 11 masked composition, 11 declared composition, 24 masked held-subject, and 24 declared held-subject slot rows. Frozen checkpoints are updates 20, 40, 60, 80, 120, 160, 240, 320, 400, and 480. A candidate must render all 11 matched composition and all 24 matched held rows exactly, retain at least 9/11 ordinary plain composition and 22/24 ordinary inline held rows, preserve the T7 full-regression tolerances, have zero deployment faults, and return identical raw and rendered outputs at batch sizes 1 and 16.

The evaluation path was hardened before launch. The prior full-regression decoder used `skip_special_tokens=True`, which could erase an illegal generated `<copy>` token before scoring. The new evaluator preserves all generated task controls, records `unresolved_task_token` as a mechanical fault, separately scores raw template exactness and deterministic rendered exactness, and never conflates either with ordinary sentence exactness. Fifteen focused local tests pass.

#### Immutable A40 launch state

The corrected authorized kit is `experiments/kits/wbv-v3-subject-slot-screen-runpod-20260802-a2/`. Every bound file and the complete kit checksum ledger passed locally and again on the worker. Static preflight confirmed two arms, 7,680 presentations per arm, 480 updates, ten adapter snapshots, 140 development rows, and no sealed test.

| Launch artifact | SHA-256 |
| --- | --- |
| Authorized build contract | `8eac140ed0a402b5ba60db52627cb6b01d148a09959237edfec97c547db73336` |
| Immutable kit contract | `b11057e7101c24eb69b09c2770d35954cbd3413d180ee9fd7dff789cd24edf98` |
| Kit checksum ledger | `85edae8a571d6d440d9f693873eeee3fd3d7b723f481a87135dfdf70ec4796c7` |
| T7 initial adapter | `02e04f84ae261bfb7fcfca4a05e221b0e2982036496ef9a8f311a526c7446d9c` |

The first worker, `03bfrytfzbtrx3`, reached D8 update 175 before an operator audit found two selector defects inherited from the older experiment runner: fixed-utterance equality would reject a genuine 54-to-55 improvement, and cross-representation diagnostic faults were incorrectly included in the primary arm fault scope. The worker was stopped and deleted before checkpoint selection. Its partial weights are not retained and it cannot produce a model claim.

Two new tests now prove that retention improvements satisfy noninferiority and that primary mechanical selection is scoped to the four ordinary endpoints plus the arm's two matched slot endpoints. Cross-representation exactness remains a secondary diagnostic tie-break, while batch-1/batch-16 invariance remains global and hard. All 17 focused tests pass.

Corrected worker `5ge03ph6aewrki` started at USD 0.44/hour on a secure Swedish A40. It has a hard provider-side termination deadline of 2026-08-02 09:51:35 UTC. Every a2 kit checksum passed locally and remotely before execution. No result or model-quality claim has yet been made. The worker will be deleted only after the result archive is recovered, independently hashed, extracted on `/mnt/donto-data`, and every result-manifest file verifies locally.

### 2026-08-02: subject-slot screen rejected, decoder-prefix mechanism discovered

The corrected paired screen completed at 2026-08-02 06:29:48 UTC. Under the preregistered greedy decoder, neither `D8` nor `M8` emitted `<copy>` once on any matched slot row at any of ten checkpoints. The result is therefore `NEGATIVE_OR_INCONCLUSIVE_SUBJECT_SLOT_SCREEN`; no adapter is authorized for publication or serving by that run.

| Selected result | `T7` baseline | `D8` step 80 | `M8` step 80 |
| --- | ---: | ---: | ---: |
| Matched rendered slot exact | not applicable | 0 / 35 | 0 / 35 |
| Valid generated slot marker | not applicable | 0 / 35 | 0 / 35 |
| Plain composition exact | 9 / 11 | 9 / 11 | 9 / 11 |
| Inline held-subject exact | 22 / 24 | 20 / 24 | 21 / 24 |
| Batch-1/batch-16 mismatches over 140 rows | baseline diagnostic | 8 | 9 |
| Direct lexical exact | 838 / 1,408 | 753 / 1,408 | 815 / 1,408 |
| Definition-context lexical exact | 590 / 1,684 | 523 / 1,684 | 593 / 1,684 |
| Fixed-utterance retention | 54 / 55 | 52 / 55 | 54 / 55 |
| Controlled-synthetic exact | 23 / 36 | 21 / 36 | 21 / 36 |
| Controlled-synthetic chrF++ | 75.0276 | 76.6330 | 74.9159 |
| Historical exact | 0 / 5 | 0 / 5 | 0 / 5 |

The result does not support more synthetic Wajarri volume. It isolates an interface failure: the complete source-bound predicate was often present, but the added special token was never the greedy argmax. The M8 training audit proves this was not a disconnected embedding row. `<copy>` received nonzero gradient on all 1,920 backward calls; its tied encoder, decoder, shared-input, and output-head row changed. D8 is additionally rejected for declared-subject interference, retention loss, and batch instability.

| Recovered a2 artifact | SHA-256 |
| --- | --- |
| `RESULT.json` | `9056c16b86d1262cb4c21d738a3fdfd1ec61f678d9f4e3a151a8ffa32aa2a633` |
| `RUN-COMPLETE.json` | `dbdbcd044b099cecb6ddb1eac389a0e3c7d42007722daec82b0f71776e2da750` |
| `RESULT-MANIFEST.json` | `af90e7a5077bb5fa8122317a323015890b68190595450751f1fd631767651276` |
| Selected D8 adapter | `18be2f62f36aea65d2f189c7943d20edb7f757599c4264871f74fd3912dc2d92` |
| Selected M8 adapter | `b924b5cdc2fcd2194a0d97a6e278d4ca075d069603a538c9fba0c26469294679` |
| Recovered 70,996,207-byte archive | `a351cb81629e81dcc0da6aa4bae6f59f30a3e19dc13c4bf2280ab4041721aa82` |

The remote manifest bound 142 files. All 142 passed on the worker, the archive hash matched after transfer, gzip integrity passed, extraction completed on `/mnt/donto-data`, and all 142 files passed again locally. Worker `5ge03ph6aewrki` was then deleted. The subsequent pod census contained only unrelated Alpha worker `wtupxv15debnvh`.

#### Post-hoc forced decoder-prefix probe

Because `<copy>` is interface metadata rather than Wajarri, the next diagnostic seeded the decoder with the exact prefix `[decoder_start=2, wbv_Latn=256204, <copy>=256208]` and allowed M8 to generate the complete predicate. Two attempted invocations produced no model result and are not counted: the first omitted the evaluator's token-extension helper; the second used `forced_decoder_ids`, which Transformers 4.48 rejects in favour of `decoder_input_ids`. The supported decoder-prefix invocation then completed.

On only the route M8 was trained to serve, the result is exact:

| Forced-prefix M8 endpoint | Batch 1 | Batch 16 |
| --- | ---: | ---: |
| Masked composition | 11 / 11 | 11 / 11 |
| Masked held-subject | 24 / 24 | 24 / 24 |
| Combined raw template exact | 35 / 35 | 35 / 35 |
| Combined rendered clause exact | 35 / 35 | 35 / 35 |
| Marker, position, and predicate correct | 35 / 35 | 35 / 35 |
| Mechanical faults | 0 | 0 |

The complete 140-row diagnostic had ten batch mismatches, but none belonged to the matched masked route. All 35 intended route outputs were identical across batch sizes. Declared-subject and ordinary rows remain invalid under this decoder and are not authorized routes. Representative exact compositions include:

```text
<translate> The <copy> is running.
<copy> jamarnimanha.
dictionary binding marruwa -> marruwa jamarnimanha.

<translate> The <copy> is sitting down.
<copy> nyinarangamanha.
dictionary binding miginy -> miginy nyinarangamanha.
```

The probe archive is 48,917 bytes with SHA-256 `8936ac4fdf9782cefcfef2df0cbfa6958ad92ccf91086ec32c53c8e1a386f297`. Its six files, including exact evaluator code, all pass the included checksum ledger locally. This is a post-hoc architecture discovery, not release evidence by itself.

#### Frozen three-seed confirmation now running

The confirmation reruns M8 from T7 for seeds 17, 42, and 73. Seed 17 is freshly trained; its post-hoc adapter is not reused as confirmation. Every seed receives the same 7,680-presentation schedule, 480 updates, token exposure, optimizer trajectory, and ten checkpoint positions. Checkpoint evaluation is restricted to the 11 masked composition and 24 masked held-subject rows under the supported decoder prefix. Selection chooses the earliest 35/35, zero-fault checkpoint. The selected adapter is then checked at batch sizes 1 and 16 and against the consumed full regression suites.

A seed passes only if all 35 raw and rendered outputs are exact, all controls and predicates are present, route faults and batch mismatches are zero, fixed-utterance retention is at least T7, synthetic exact loses no more than two rows, synthetic chrF++ loses no more than two points, and no unresolved task token appears. Even a three-seed pass authorizes only a fail-closed controlled route over the six governed predicates. It does not authorize free-form translation.

| Confirmation artifact | SHA-256 |
| --- | --- |
| Authorized build contract | `50e1d61f6116ba0626f4317e9364c4607cd24af5296689e0e866d9e50246818b` |
| Immutable kit contract | `a9fcfa3584d12a1da5d10056d3bd7bcda6cd4823eda0dc4fab0b9f00da591521` |
| Kit checksum ledger | `ed3dd087591dc03c51870c7e5d58e6d4b8172d2d9c162c4271f1cc0784d0b5fd` |
| Reproducible 35,453,128-byte kit archive | `4501c0a98e1a9497dacaa75d98e21e42ed4f59053466e361b64a1d9862c1b9f0` |

All 34 kit files pass locally and remotely; static preflight and 17 focused local tests pass. An initial worker, `d75nut7548tuec`, exposed only a 20 GB filesystem and was deleted before upload or training because it failed the 30 GiB reserve. Replacement worker `77tes61jw5a024` has an explicit 50 GB mounted workspace, a secure Swedish A40, a USD 0.44/hour rate, and provider auto-termination at 09:00 UTC. The unrelated Alpha worker is not part of this run.

### 2026-08-02: first confirmation invalidated by a seed audit

The first confirmation run, `wbv-v3-forced-slot-confirmation-runpod-20260802-a1`, initially reported the same positive route result for declared seeds 17, 42, and 73. Artifact recovery and evaluation were valid, but a post-run trainer audit found that the run was not a valid multi-seed confirmation.

The command-line seed was written into each manifest, but the constructed `Seq2SeqTrainingArguments` did not receive either `seed` or `data_seed`. Transformers therefore reset all three Trainer instances to its default seed 42. The three selected update-40 adapters were byte-identical:

```text
declared seed 17 -> d59a4a276d4968e6d4e453dd534ebca6edfaacf936e1bd97cf7c3eb4be1c8801
declared seed 42 -> d59a4a276d4968e6d4e453dd534ebca6edfaacf936e1bd97cf7c3eb4be1c8801
declared seed 73 -> d59a4a276d4968e6d4e453dd534ebca6edfaacf936e1bd97cf7c3eb4be1c8801
```

That run proves deterministic repeatability under one effective stochastic state, not robustness across three independent seeds. Its adjudicated status is `INVALID_AS_MULTI_SEED_CONFIRMATION`; it never authorized a release. The complete invalid-run evidence was retained rather than rewritten:

| Invalid a1 artifact | SHA-256 |
| --- | --- |
| `RESULT.json` | `50535a692d4132a81c68b645db1074b519ed0ee5dbc4907a65fe5c71613c33f0` |
| `RUN-COMPLETE.json` | `33407cd0a06579d2a69470ab44440f6c8c021f52deffa603e4b4106459c9f8f7` |
| `RESULT-MANIFEST.json` | `2640d0067a7c88b10b38cacbeb42c575265c10852558d9d2397c56718d2008be` |
| Recovered 105,094,918-byte archive | `36a43404c482e06e4703894fa5ce83c50b3e6a1162223bf8cd5d2d82649e33cb` |

All 200 a1 manifest files still verify locally. The corrective implementation binds both Trainer seed fields to the declared run seed, records the runtime values, fails on a mismatch, and requires three distinct selected adapter hashes. The replacement a2 kit contract SHA-256 is `83788fad2533713c61d6d12e9d212a7c5bfe60dbba6a29efd6fc8ff016791e14`; its archive SHA-256 is `71ae31582f1bfe1f726d415db937ebd393f335f9d029db6673130429a0dfa4a8`.

### 2026-08-02: corrected independent confirmation passed the controlled route

Replacement run `wbv-v3-forced-slot-confirmation-runpod-20260802-a2` retrained seeds 17, 42, and 73 independently from the same T7 start. Every seed used the same 7,680-presentation M8 schedule, 480 maximum updates, checkpoint positions, optimizer contract, six masked model-visible pairs, and 35-row consumed route screen. The only intended stochastic change was the bound seed and data seed.

The selector chose update 40 for every seed because it was the earliest checkpoint satisfying all route gates. Each adapter generated the exact raw template and exact rendered clause for all 35 rows at batch sizes 1 and 16. Each had zero route faults and identical outputs between batch shapes. The three weight hashes are distinct, proving the stochastic states were no longer collapsed.

| Confirmation gate | Seed 17 | Seed 42 | Seed 73 |
| --- | ---: | ---: | ---: |
| Trainer seed bound to declared seed | pass | pass | pass |
| Trainer data seed bound to declared seed | pass | pass | pass |
| Distinct selected adapter hash | pass | pass | pass |
| Raw slot template exact | 35 / 35 | 35 / 35 | 35 / 35 |
| Rendered clause exact | 35 / 35 | 35 / 35 | 35 / 35 |
| Required predicate present | 35 / 35 | 35 / 35 | 35 / 35 |
| Slot count and first position valid | 35 / 35 | 35 / 35 | 35 / 35 |
| Mechanical faults | 0 | 0 | 0 |
| Batch-1/batch-16 mismatches | 0 | 0 | 0 |
| Fixed-utterance retention | 54 / 55 | 54 / 55 | 54 / 55 |
| Controlled-synthetic exact | 22 / 36 | 22 / 36 | 22 / 36 |
| Controlled-synthetic chrF++ | 75.4246 | 75.3028 | 75.1763 |
| Historical exact | 0 / 5 | 0 / 5 | 0 / 5 |

Seed 17 is the released adapter. Its selected weight SHA-256 is `a4a91d4e6a6ba49509a8ab432c0674686c96b4611f58db8b63089f26707ed3d6`. Seed 42 is `d59a4a276d4968e6d4e453dd534ebca6edfaacf936e1bd97cf7c3eb4be1c8801`; seed 73 is `8574727bbf9cfb3e5032e9d922d7af61d6eb62f4b6b0976d6fc06b18b7dd82f7`.

The route result does not erase the negative evidence. All three adapters remain 0/5 exact on the complex historical diagnostic, and direct one-target lexical reconstruction is below public v2. The positive claim is therefore restricted to exact template generation under the forced control prefix and deterministic subject rendering. Dictionary lookup remains the product authority for known words.

| Corrected a2 artifact | SHA-256 |
| --- | --- |
| `RESULT.json` | `6abbad036a146c3a15695f1c0ca31e7427f20f2f4ad7a712350cc261fac92fb1` |
| `RUN-COMPLETE.json` | `a76fb09f7e8f46200536fbd1164ec0c7a57c6a18ba158fecdb6914c6ff02895f` |
| `RESULT-MANIFEST.json` | `2f7f5d2d97827c7641c035f64e3fce99a660e3c8f4ec223dee4af8606a742626` |
| Deterministic recovered archive | `c445e98201cd4dbc484e74a27123aef4de3daeb12393998f8ce398fc9702b700` |

The a2 manifest binds 200 files. All 200 passed on the worker, archive transfer and gzip integrity passed, the local archive hash matched, extraction succeeded on `/mnt/donto-data`, and all 200 files passed again locally.

### 2026-08-02: CPU float32 parity passed

The first automated CPU invocation stopped before model loading because the standalone evaluator package omitted `nllb_runtime_token_extension.py`. This was a packaging failure, not a model failure, and no parity claim was made. The launcher was corrected to require and checksum the helper before execution. Its SHA-256 is `1c9156464016efc43edf82df6fbc0afa6ffc590133322af6dd9e3c29fc89c695`; the runtime helper SHA-256 is `cf1637bb17a358391770cb307e326151f63a85a9b2c2f083b531cce5d6aa5e2f`.

The corrected run loaded the exact seed-17 adapter on CPU float32 and compared it with the A40 bfloat16 evidence. All 35 rows matched exactly at batch sizes 1 and 16. The normalized prediction SHA-256 was the same in all four conditions:

```text
b17ab8d44ccaaa8dba962ecf46486d731d7aaf8d4c6a7adeed430044612cf25f
```

Every CPU identity, row-set, raw-template, rendered-clause, predicate, slot-shape, fault, and batch-invariance gate passed. The parity report SHA-256 is `9758ec536ce5500f84f544d03ea67cd7b79edbd9b69d361b1bb061de2be9352f`; its result manifest is `8a23898a6d98f703408ab41930e636e69a24f93780ea02231f9e2504e6dbf5b1`; the deterministic parity archive is `317fbad1d01547ab5c178ab5f5844a8fbe11f47d2dd33a807da8f23eb1d9daf8`. All six parity-manifest files verify locally.

### 2026-08-02: worker recovery and deletion completed

Wajarri worker `77tes61jw5a024` was kept only through result recovery, local archive hashing, extraction, 200-file result verification, and CPU parity completion. It was then deleted. A fresh provider census proved that worker absent and left unrelated Alpha worker `wtupxv15debnvh` untouched.

| Shutdown evidence | SHA-256 |
| --- | --- |
| Worker record before deletion | `b1b0c3da5b9bbdd635fee8ea5f6af23f94a80ebb2f7ef5245fa3e2a1f0f0ae4f` |
| Provider census after deletion | `44aefc19b19bc35d7f70a51749af394862fe939253a7c5ad2970acd65b00b449` |
| Post-provider deletion ledger | `91483094471a24e7abfc46ca2d7fb35226566348d7530e2adf30e46796090c4d` |

The manifest-bound `DELETION-LEDGER.json` inside the confirmation output records training-time cleanup and is distinct from the later provider-worker ledger. During release preparation, an attempted overwrite of that manifest-bound file was rejected by the checksum builder. The original was restored from the independently verified archive, all 200 manifest files were reverified, and the provider deletion record remained a separate artifact. This is why the release contains both ledgers under different names.

### 2026-08-02: immutable Hugging Face dataset published

The controlled dataset release is public at [ajaxdavis/mobtranslate-wajarri-synthetic-corpus-v1](https://huggingface.co/datasets/ajaxdavis/mobtranslate-wajarri-synthetic-corpus-v1).

| Dataset identity | Value |
| --- | --- |
| Commit | `302f3a93eb255a2fc70692b7c429ba5ee8300687` |
| Immutable tag | `v3.0.0-controlled-subject-slot-s17-step40-20260802` |
| Files | 59 total; 58 checksum-manifest rows |
| Training parent pairs | 103 |
| Unique masked model-visible pairs | 6 |
| Scheduled presentations | 7,680 |
| Newly invented Wajarri target sentences | 0 |
| Sealed-test rows read during release construction | 0 |
| Anonymous publication verification | 59/59 present; 58/58 checksums exact |
| Publication verification SHA-256 | `a590e514cf36a50a6b62566e1dd4d2907d70e9d79da41e1a4fe280d90347db6a` |

The release preserves source records, schedules, model-visible masked pairs, development evidence, checksums, and the invalid-a1 adjudication. Repeated schedule presentations are not described as unique sentences, and glossary or masked input variants are not counted as new Wajarri targets.

### 2026-08-02: immutable Hugging Face model published

The controlled adapter is public at [ajaxdavis/mobtranslate-wajarri-v3-controlled](https://huggingface.co/ajaxdavis/mobtranslate-wajarri-v3-controlled).

| Model identity | Value |
| --- | --- |
| Model ID | `mobtranslate-wajarri-v3-controlled` |
| Version | `v3.0.0-controlled-subject-slot-s17-step40-20260802` |
| Commit | `1a26382648e294f2c102102ef914c7b0c86bec20` |
| Immutable tag | `v3.0.0-controlled-subject-slot-s17-step40-20260802` |
| Base | `ajaxdavis/mobtranslate-wajarri-v1` at `ddae6103913b7cab299d687bee2dc7904fd59459` |
| Adapter SHA-256 | `a4a91d4e6a6ba49509a8ab432c0674686c96b4611f58db8b63089f26707ed3d6` |
| Release files | 31 checksum-bound files plus repository metadata |
| Release record SHA-256 | `589d6a8a8b1de0b2c8a64c89abe79c453dd75c84d849179dcedd0db56572a809` |
| Publication verification SHA-256 | `4f9d35aff3fe14c16c590161b364f785c8821a2501b0c26c2f619fd040eec9a7` |

Anonymous downloads verified every published file byte-for-byte. Both the branch commit and immutable tag resolve to the same release content. The repository contains the adapter, tokenizer/runtime extension files, frozen templates, GPU evidence, CPU parity evidence, training provenance, invalid-a1 adjudication, worker-deletion evidence, checksums, requirements, and a complete hosting guide.

The release builder exposed and fixed one filesystem trap before publication. `adapter_config.json` was initially hardlinked from the recovered result; rewriting its portable base reference would therefore have mutated the source artifact. The builder now copies that file before editing and a test proves the source remains byte-identical. The source recovered adapter and all confirmation checksums remain unchanged.

The NLLB-derived model follows the upstream CC BY-NC 4.0 constraint. Application code licensing does not override model or data terms.

### 2026-08-02: shared Hugging Face Space deployed and tested

The shared adapter service [ajaxdavis/alpha-v0-historic](https://huggingface.co/spaces/ajaxdavis/alpha-v0-historic) is deployed at commit `03b6d4ed9c9ffb8ac2e55676447084af2c7aa5c8`. Wajarri runs on CPU float32 under the existing bounded on-demand adapter cache. It is not mounted on the MobTranslate host and does not require a continuously running RunPod GPU.

The Wajarri service advertises only task `subject_slot`; generic Wajarri `translate` and `lexeme` tasks are rejected. Startup binds the exact model repository, version, base revision, `<copy>` token ID 256208, target-language token ID 256204, runtime special-token initialization, six templates, and checksums. Twenty-nine Space tests pass.

| Live English construction | Bound subject | Exact Space output |
| --- | --- | --- |
| coming towards the speaker | `marruwa` | `marruwa yanajimanha.` |
| going away from the speaker | `marruwa` | `marruwa yanmanha.` |
| running | `marruwa` | `marruwa jamarnimanha.` |
| sitting down | `marruwa` | `marruwa nyinarangamanha.` |
| sitting | `marruwa` | `marruwa nyinamanha.` |
| standing | `marruwa` | `marruwa garrimanha.` |

All six live requests returned HTTP 200 and exact output. An unsupported construction, a missing dictionary binding, and generic Wajarri `/v1/translate` each returned HTTP 422 `invalid_request`. The Space had zero failed Wajarri model requests in final acceptance. The live acceptance report SHA-256 is `fabf8453b7192c62867e65cebd11af6924321e07e93cf3b4bc7dc577e4596571`.

The bounded cache may unload Wajarri when the two warm adapters occupy memory. In that state `/health` reports Wajarri as `available`, not failed; the next uncached Wajarri request reloads the exact pinned adapter. MobTranslate's PostgreSQL result cache prevents repeated text from repeatedly paying that load cost.

### 2026-08-02: production homepage route accepted

Production web release `20260802T091947Z-72668fe89dc5` deployed through the immutable release script. The full web suite passed 775 tests across 93 files, focused Wajarri routing passed 14/14 tests, TypeScript passed, the production build completed under the service environment, the candidate smoke test passed, and runtime checksums remained unchanged before and after public smoke testing.

The homepage route is abstracted through the same multi-language hybrid registry as Kuku Yalanji and Mi'kmaq. Wajarri adds a controlled task contract rather than language-specific endpoint code. The registry binds model ID, model version, dictionary code `wbv`, task contract, six construction IDs, templates, and the immutable dataset revision.

#### Supported live request

```json
{
  "text": "The bilby is running.",
  "mode": "translate",
  "stage": "draft",
  "direction": "to_language"
}
```

The public endpoint `POST https://mobtranslate.com/api/translate/wbv` returned HTTP 200 and exact `marruwa jamarnimanha.` through route `huggingface_draft`. The response identified model `mobtranslate-wajarri-v3-controlled`, the exact v3 version, contract `wajarri-subject-slot-v3-s17-step40-20260802`, construction `running`, model template `<copy> jamarnimanha.`, and dictionary headword `marruwa`.

The first request took 7.726 seconds and wrote a model-bound PostgreSQL draft cache row. The repeat took 90 ms and reported `draft: hit`. The complete source-check request preserved the translation exactly, returned route `huggingface_grammar_review`, and exposed the supporting dictionary records and scope limit under progressive disclosure. Its first request took 4.278 seconds; the repeat took 85 ms with draft, evidence, review, and resolver all reporting cache hits.

#### Unsupported live request

`The bilby is flying.` is outside the six-construction contract. It did not invoke the v3 adapter. The public endpoint returned through `dictionary_prompt` using `gpt-5.4-mini`, then cached that result separately. The first request took 2.800 seconds and the repeat took 72 ms. This proves that one model's narrow pass does not authorize unsupported Wajarri sentence generation.

#### Cache identity

The controlled draft cache key binds:

```text
language
HMAC source fingerprint and source length
dictionary revision
model ID and immutable version
draft contract
construction ID
model template
dictionary bindings fingerprint
```

Evidence, review, and resolution add the evidence fingerprint, reviewer version, and resolver version. The live database contains separate ready rows for `hybrid_hf_controlled_draft`, `hybrid_controlled_dictionary_evidence`, `hybrid_controlled_llm_review`, and `hybrid_controlled_resolved_translation`. Unsupported prompt output is stored under separate stage `dictionary_prompt_translation`. A dictionary update, model update, task-contract change, construction change, binding change, reviewer change, or resolver change produces a new identity rather than reusing stale output.

#### Browser and layout evidence

Real Chromium acceptance selected Wajarri on the public homepage, entered `The bilby is running.`, and observed `marruwa jamarnimanha.`. The interface shows the translation first. “Why this result?” then reveals the dictionary records, automated source check, limits, model identity, and cache state. The text is ordinary public-facing language rather than training-pipeline jargon.

| Browser evidence | Viewport | Result | SHA-256 |
| --- | --- | --- | --- |
| Desktop first viewport | 1440x1000 | exact result visible; no horizontal overflow or overlap | `687b1cc8b05212667ad59d5ab3f828c5e93f5a1e5332e64abfc52175faec196c` |
| Mobile full page | 390x844 | exact result present after normal scroll; no horizontal overflow | `1eb53e4fd772f89d3f36dd7f48c0323a2f16a318a11b79ea0c26d8bef3a45bb0` |
| Mobile expanded evidence | 390x844 | progressive disclosure opens and remains readable | `b5c0c9b44299a98100e5c4523d9ece88729332880f58f919e98ed695c81b269e` |

The machine-readable homepage and public-registry acceptance record is `experiments/results/wbv-v3-controlled-homepage-live-acceptance-20260802.json`, SHA-256 `7bf05e5b78ae1c05a57589aeb5a1ee320c1cc208af8a1f7fcae563d6f8516b54`.

### 2026-08-02: public model registry reconciled

The final cross-system audit found one publication inconsistency after the model, dataset, Space, and homepage route were already live. `GET /api/v1/models` listed only Kuku Yalanji and Mi'kmaq, while `GET /api/v1/models/mobtranslate-wajarri-v3-controlled` returned HTTP 404. This did not break homepage translation, but it meant a third-party host could not discover the released Wajarri adapter through MobTranslate's model-distribution API.

The fix added Wajarri to the existing generic registry rather than creating a Wajarri-specific API. The entry binds the exact controlled task, immutable version, base revision, adapter hash, bounded metrics, noncommercial terms, model and dataset repositories, checksum ledger, CPU parity report, Space endpoint, and this work log. It states `free_form_translation_authorized: NO`. Public serialization strips the source-only `localPath`, and a focused test fails if any `/mnt/` path becomes visible.

The first registry deployment attempt stopped before cutover because an unrelated one-hour spaced-repetition assertion read the live clock twice and occasionally observed 59 minutes 59 milliseconds. The test now pins time with Vitest's fake clock, matching the deterministic seven-day boundary test already in the suite. Its file passes 83/83 tests. Release `20260802T100136Z-72668fe89dc5` then exposed the correct API, but real-browser inspection found that the catalog rendered the custom status `published-controlled` as both “published controlled” and “Not published.” A shared status predicate was added so controlled publication remains a distinct state but is displayed and selected as published. Release `20260802T101815Z-72668fe89dc5` fixed that contradiction. Its mobile check then exposed clipped immutable base and dataset IDs, so the model card gained flex shrinking and `overflow-wrap:anywhere` for long identities.

The final immutable deployment passed all gates and published production release `20260802T103148Z-72668fe89dc5` with source SHA-256 `e434ad550ee20b438939dcdf4cbbeca569dff8bb0fccca8c1a0e176a6330790f`. Release integrity covers 8,640 files and 37 symlinks. The service is active and `/api/health` reports that exact release.

| Final registry acceptance | Result |
| --- | --- |
| Catalog `GET /api/v1/models` | HTTP 200; model count 3; Wajarri present |
| Wajarri model endpoint | HTTP 200 |
| `versions/latest` | HTTP 200; resolves exact v3 version |
| Exact immutable-version endpoint | HTTP 200 |
| Published status | `published-controlled` |
| Task | `controlled_subject_slot_translation` |
| Model downloads | 2: immutable repository and adapter weights |
| Dataset downloads | 1 |
| Documentation downloads | 1 |
| Available artifact links | 8 / 8 returned HTTP 200 |
| Private path leaks | 0 |
| Model-distribution tests | 8 / 8 passed |
| Full web suite | 777 / 777 passed across 93 files |
| TypeScript, release-integrity, build, candidate smoke | pass |
| Catalog publication label | `Published: controlled`; contradictory `Not published` absent |
| Mobile base and dataset identity wrapping | full text retained; client width equals scroll width |
| Catalog horizontal overflow | none at 1440x1000 or 390x844 |
| Immutable model-version page horizontal overflow | none at 390x844 |

The registry source SHA-256 is `79887d73de776d53ea1c8a728083cf4d2df2187f55be58f29e7eb2a36e4f3c20`; its focused test source SHA-256 is `4aa1355039700f16f4723fb8abe84e9a42347fe830d97857757c05b09363b4fc`; the shared publication-status helper SHA-256 is `dfd0c6304f081722eccd4498cfffa5e8f2203e9d2b146f097608bc8c7c3f9ac7`; the final catalog-page source SHA-256 is `183de1561b632a5654da80cc8c90e670f7505c3e3efe33cba653497e329239e1`. Final live translation smoke repeated the bounded route: `The bilby is running.` returned exact `marruwa jamarnimanha.` through `huggingface_draft`; complete review preserved the exact translation; unsupported `The bilby is flying.` remained on `dictionary_prompt` and did not invoke the adapter.

| Final catalog browser evidence | Viewport | SHA-256 |
| --- | --- | --- |
| Wajarri catalog card, controlled status | 1440x1000 | `6f56fe749741b9542f6ab62021882daeef9d4f9c317768e76c2bc538fd04d877` |
| Wajarri catalog card, wrapped immutable identities | 390x844 | `07d840248ee0e6ad80d2eb7881365f5985e723675c2413c86c8627474f292ee8` |
| Immutable model-version page | 390x844 | `893beb8e52f39f4dafca9b9755831e20d4fd5dfb555d880ec3b77159a20862c2` |

The earlier homepage release `20260802T091947Z-72668fe89dc5` remains the first successful controlled-route deployment and is preserved in the chronological record above. The final release supersedes it only to expose the same immutable Wajarri artifact through the public registry and present that record consistently on mobile and desktop. Deployment retention also reported one pre-existing incomplete release from 2026-07-21 that could not be verified for pruning; it did not alter the active release or current runtime integrity.

### Final v3 release ruling

The Wajarri v3 program is complete for the bounded release objective:

| Required outcome | Final state |
| --- | --- |
| Source-governed model-visible pairs | complete and published |
| Independent seed confirmation | pass, 3/3 distinct adapters |
| Exact controlled route | pass, 35/35 for every seed |
| Batch-size invariance | pass at batch 1 and 16 |
| GPU-to-CPU parity | pass, exact prediction hashes |
| Full result recovery and checksums | pass |
| Wajarri RunPod shutdown | pass; worker deleted |
| Public Hugging Face dataset | pass |
| Public Hugging Face model and immutable tag | pass |
| Shared-Space API | pass with fail-closed refusals |
| Public model registry and immutable version API | pass |
| Public model catalog desktop/mobile presentation | pass |
| Homepage supported/fallback routing | pass |
| PostgreSQL model-bound caches | pass |
| Desktop/mobile browser acceptance | pass |
| Free-form Wajarri translation | **not authorized** |
| Productive morphology or natural-language reliability | **not established** |

The technically important result is architectural. Sparse, source-bound evidence did not justify manufacturing thousands of novel Wajarri clauses. Direct sentence fine-tuning repeatedly traded sentence fitting against dictionary retention. The successful release instead isolates exactly what the evidence supports: deterministic dictionary retrieval for the subject, neural selection/generation of one complete governed predicate under a forced task prefix, deterministic rendering, and refusal outside the frozen contract.

Any later free-form Wajarri model is a new research program. It must begin with additional independently attested clauses or explicit speaker-reviewed morphological paradigms, family-level held-out evaluation, and a separate sentence-generation gate. It must not inherit v3's narrow pass as evidence for unrestricted translation.