# Kuku Yalanji v21.2 Model, Training, Data, and Hosting Guide

**Published:** 2026-07-13 UTC  
**Updated:** 2026-07-14 UTC  
**Model ID:** `kuku-yalanji-nllb-lora`  
**Current downloadable release:** `v21.2-claude-balanced-replay-guarded-20260714`  
**Immutable weights release:** `v21.2-claude-balanced-replay-v2-candidate`  
**Task:** English (`eng_Latn`) -> Kuku Yalanji (`gvn_Latn`) translation  
**Status:** noncommercial research only; not speaker-certified; not approved for authoritative translation

## 1. Read this first

This is the public, end-to-end handoff for downloading, loading, hosting, auditing, and understanding the current
Kuku Yalanji model candidate. The guarded release preserves the exact v21.2 weights and changes only the validated
decoding policy. It remains downloadable, but local custom-model CPU inference is stopped and disabled because
v21.2 fails the separate 80% dictionary deployment gate. It accompanies the versioned model API and immutable model
and dataset bundles.

The release is a sequence-to-sequence translation model, not a chat LLM. It is a PEFT LoRA continuation of a
project-specific NLLB checkpoint derived from `facebook/nllb-200-distilled-1.3B`. The exact frozen base contains the
project's `gvn_Latn` token at vocabulary ID 256204. The v21.2 adapter is therefore **not compatible with an untouched
stock NLLB checkpoint**, even if the upstream model name is the same.

The release decision is deliberately narrow:

| Question | Evidence-based answer |
|---|---|
| Best current Kuku Yalanji custom research candidate | v21.2 balanced replay |
| Current inference policy | Beam 1, no-repeat 4, repetition penalty 1.10, length penalty 1.0 |
| New model weights in the guarded release | No; weights are byte-identical to v21.2 |
| Safe unrestricted translator | No |
| Speaker or elder certified | No |
| Appropriate role | Labelled research draft after exact-resource lookup and generation guards |
| Exact known sentence and dictionary resources | Continue to serve by retrieval, not generation |
| Commercial use | Not granted; the NLLB base is CC-BY-NC-4.0 and project data has additional restrictions |

The model learned the project-authored synthetic treatment and retained Bible and dictionary-usage domains much
better than the v21.1 synthetic-only ablation. It did not establish general natural Kuku Yalanji competence. Both
v21 models scored 0/43 exact on the frozen elder-shared control. The archived v21.2 greedy decoder exhibited severe
under-translation and morpheme-loop failures; the guarded policy reduced tagged-test repeated-segment rows from 37
to 2 without materially degrading the Bible controls. Natural-register transfer remains unresolved. A successful
checksum or smoke test establishes artifact compatibility only.

## 2. Distribution API

The public API is read-only, requires no API key, permits cross-origin `GET` requests, and returns absolute URLs.
No server filesystem paths are exposed.

| Purpose | URL |
|---|---|
| List every model and version | <https://mobtranslate.com/api/v1/models> |
| Kuku Yalanji model record | <https://mobtranslate.com/api/v1/models/kuku-yalanji-nllb-lora> |
| Current downloadable Kuku Yalanji release | <https://mobtranslate.com/api/v1/models/kuku-yalanji-nllb-lora/versions/latest> |
| Current guarded downloadable release | <https://mobtranslate.com/api/v1/models/kuku-yalanji-nllb-lora/versions/v21.2-claude-balanced-replay-guarded-20260714> |
| Historical weights/greedy release | <https://mobtranslate.com/api/v1/models/kuku-yalanji-nllb-lora/versions/v21.2-claude-balanced-replay-v2-candidate> |
| Mi'gmaq model record | <https://mobtranslate.com/api/v1/models/migmaq-nllb-lora> |
| OpenAPI 3.1 document | <https://mobtranslate.com/api/v1/models/openapi.json> |
| Human model browser | <https://mobtranslate.com/models> |

Discover model downloads programmatically:

```bash
curl -fsS \
  https://mobtranslate.com/api/v1/models/kuku-yalanji-nllb-lora/versions/latest \
  | jq '.release | {version, rights, downloads: .downloads.models}'
```

Resolve a URL and continue an interrupted large download:

```bash
API=https://mobtranslate.com/api/v1/models/kuku-yalanji-nllb-lora/versions/latest
URL="$(curl -fsS "$API" | jq -r '.release.downloads.models[] | select(.kind == "bundle") | .url')"
curl --fail --location --continue-at - --output kuku-yalanji-v21.2-guarded-atlas-lora-bundle.tar.gz "$URL"
```

`latest` is a convenience alias and may resolve to a different version after a future release. Hosts must record
and pin the immutable version and SHA-256 value returned by the API. The literal version endpoint is the
reproducible contract.

## 3. Release downloads

### 3.1 Models and hosting metadata

| Artifact | Bytes | SHA-256 | Direct URL |
|---|---:|---|---|
| Atlas dynamic LoRA bundle: exact base + adapter + guarded policy | 2,530,623,924 | `012335cfb393ca386c7817bd3850dd52f1a751e0e1a58e80e9e5bc9bae9c89d5` | [tar.gz](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/kuku-yalanji-v21.2-guarded-atlas-lora-bundle.tar.gz) |
| Standalone merged model (weights archive unchanged) | 2,200,740,249 | `920e179d8b4fe45ce75bdd1e9cc2784be51ff8d55d1a190cd45e19cf01a5f490` | [tar.gz](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/kuku-yalanji-v21.2-guarded-merged.tar.gz) |
| Machine hosting manifest | 3,487 | `e40fd2586a3d4eabfdf6052bd65e4f3a92532b1141ea446bc81feb03a7a01c95` | [JSON](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/HOSTING-MANIFEST.json) |
| Release metadata | 1,915 | `a5ded1fba823e2596bfaa966298961d81c91087c6a424fd1abd8744a371b33df` | [JSON](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/release.json) |
| Release checksum file | 1,302 | `81e992b0e07ca37d99422ddcf65449fca88bf831a18d4ab1cfec08b1d37c15a1` | [SHA256SUMS](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/SHA256SUMS) |

Use the **Atlas bundle** when the host supports dynamic encoder-decoder PEFT adapters. It contains `base/`,
`adapter/`, `README.md`, and `HOSTING-MANIFEST.json`. The adapter configuration is portable and points to
`../base`.

Use the **merged bundle** when the serving stack cannot dynamically attach a LoRA adapter. It extracts directly
into one Transformers-compatible model directory. Both modes implement the same frozen release and must pass the
same smoke tests.

### 3.2 Training data and process evidence

| Artifact | Bytes | SHA-256 | Direct URL |
|---|---:|---|---|
| Exact v21.2 train, validation, test, gate, manifest, and quarantine package | 5,969,287 | `91fc18f539b863f9a0e87ed3c372b5f040f359e9bcd427dc506875e1b4f6a467` | [tar.gz](/datasets/kuku-yalanji-v21.2-training-20260711.tar.gz) |
| Complete 20,047-sentence synthetic process corpus v2 | 22,069,344 | `5db11f1f10f60b1c6df0ea00f0bd4b2f8a313dfc487ae11145c495c96d880e73` | [zip](/datasets/kuku-yalanji-synthetic-research-corpus-v2.0.0-20260711.zip) |
| Complete process corpus, alternate format | 19,438,024 | `f3d671921944a065e83c11d028471fd7bf0127e6665cae4ded699cd1b23e56ba` | [tar.gz](/datasets/kuku-yalanji-synthetic-research-corpus-v2.0.0-20260711.tar.gz) |
| All 20,047 sentences | 17,606,394 | `b06ec5f5bbc1f4e39cf20a30fefe04d496c42f1408c73e75616ebbd531a714f0` | [JSONL](/datasets/kuku-yalanji-synthetic-research-corpus-v2.0.0-20260711/data/sentences.jsonl) |
| Source-preserving dictionary | 843,498 | `d1929dab4cb5cd9b8399f2a63e2567a836a6cb40463ad70afefa4cc6f698e074` | [JSONL](/datasets/kuku-yalanji-synthetic-research-corpus-v2.0.0-20260711/data/dictionary.jsonl) |
| Complete corpus and process database | 35,815,424 | `b9db281cccfa50ced087747d812b5712ceac22dc7b0cd1f3227d7c2e36988703` | [SQLite](/datasets/kuku-yalanji-synthetic-research-corpus-v2.0.0-20260711/data/synthetic.db) |

The model-specific package is the shortest path to the exact final v21.2 treatment. The larger process release is
the audit path: it contains the sentence corpus, dictionary, canonical splits, governed v21.1 splits, reviews,
revisions, lessons, event logs, audit code, and living books.

### 3.3 Integrity verification

Verify before extraction or loading:

```bash
sha256sum kuku-yalanji-v21.2-guarded-atlas-lora-bundle.tar.gz
sha256sum kuku-yalanji-v21.2-guarded-merged.tar.gz
sha256sum kuku-yalanji-v21.2-training-20260711.tar.gz
```

After extracting the Atlas bundle:

```bash
sha256sum base/model.safetensors base/tokenizer.json base/config.json
sha256sum adapter/adapter_model.safetensors adapter/tokenizer.json
```

The expected internal identities are recorded in section 5 and in `HOSTING-MANIFEST.json`.

## 4. Loading and serving

### 4.1 Proven software versions

The completed RunPod lane used:

```text
Python       3.11
PyTorch      2.4.1+cu124
Transformers 4.48.3
PEFT         0.19.1
Accelerate   1.14.0
Datasets     4.8.5
SacreBLEU    2.6.0
GPU          NVIDIA A40
```

The training requirements allowed `transformers>=4.46,<4.49`, `peft>=0.18,<1`, `torch>=2.4`, and
`sentencepiece>=0.2,<1`. A host may use a newer compatible stack, but it must pin its environment and rerun both
smoke tests before accepting the deployment.

### 4.2 Dynamic LoRA loading

Extract the Atlas bundle into an empty directory. The archive has `base/` and `adapter/` at its root.

```python
import torch
from peft import PeftModel
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained(
    "./adapter",
    src_lang="eng_Latn",
    tgt_lang="gvn_Latn",
)
base = AutoModelForSeq2SeqLM.from_pretrained("./base").to(device)
model = PeftModel.from_pretrained(base, "./adapter").to(device)
model.eval()

text = "<translate> The woman saw the water."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=200).to(device)
target_id = tokenizer.convert_tokens_to_ids("gvn_Latn")

with torch.inference_mode():
    output = model.generate(
        **inputs,
        forced_bos_token_id=target_id,
        num_beams=1,
        max_new_tokens=208,
        no_repeat_ngram_size=4,
        repetition_penalty=1.10,
        length_penalty=1.0,
    )

print(tokenizer.batch_decode(output, skip_special_tokens=True)[0])
```

Expected output:

```text
Jalbu-ngku bana nyajin.
```

Atlas or another dynamic host must support:

- encoder-decoder `M2M100ForConditionalGeneration` models;
- PEFT task type `SEQ_2_SEQ_LM`;
- custom tokenizer files and the added `gvn_Latn` token;
- LoRA rank 64 on attention and feed-forward modules;
- forced target BOS configuration;
- visible research-only and non-speaker-certified labeling.

If a platform accepts only a stock upstream model ID plus an adapter, do not point it at stock NLLB. Upload or
mount the included frozen `base/` as the base model.

### 4.3 Merged-model loading

The merged archive extracts with model files at its root:

```python
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained(
    "./kuku-v21.2-merged",
    src_lang="eng_Latn",
    tgt_lang="gvn_Latn",
)
model = AutoModelForSeq2SeqLM.from_pretrained("./kuku-v21.2-merged").to(device)
model.eval()

inputs = tokenizer(
    "<translate> The children are sitting here.",
    return_tensors="pt",
    truncation=True,
    max_length=200,
).to(device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        forced_bos_token_id=tokenizer.convert_tokens_to_ids("gvn_Latn"),
        num_beams=1,
        max_new_tokens=208,
        no_repeat_ngram_size=4,
        repetition_penalty=1.10,
        length_penalty=1.0,
    )

print(tokenizer.batch_decode(output, skip_special_tokens=True)[0])
```

Expected output:

```text
Karrkay-karrkay yaluy bundanday.
```

### 4.4 Minimal HTTP contract

A host-facing translation service should accept a UTF-8 JSON request and identify the immutable model version in
every response. A minimal contract is:

```json
{
  "text": "The woman saw the water.",
  "source_lang": "eng_Latn",
  "target_lang": "gvn_Latn",
  "model_version": "v21.2-claude-balanced-replay-guarded-20260714"
}
```

The service must prepend `<translate> ` exactly once, reject empty or over-limit input, force `gvn_Latn` as BOS,
serialize or bound concurrent generation, and return a retryable error when capacity is exhausted. It should also
record input/output hashes, model version, decoding configuration, latency, and guard decisions. Do not log private
user text indefinitely without an explicit retention policy.

## 5. Model identity and lineage

### 5.1 Architecture

```text
Upstream family       facebook/nllb-200-distilled-1.3B
Transformers class    M2M100ForConditionalGeneration
Task                  conditional sequence-to-sequence translation
Frozen project base   v12.0 tagged direct+reference Bible continuation
Source language       eng_Latn
Target language       gvn_Latn
Vocabulary size       256,205
gvn_Latn token ID     256,204
Input prefix          <translate> 
Adapter type          PEFT LoRA / SEQ_2_SEQ_LM
Merged weight bytes   2,741,395,552
```

Earlier MobTranslate experiments did use the 600M NLLB model. The v21 family did not: v21.1 and v21.2 both start
from the frozen v12 continuation derived from the **1.3B** upstream checkpoint. This distinction matters for
compatibility, memory planning, attribution, and license records.

### 5.2 Frozen base hashes

| File | SHA-256 |
|---|---|
| `base/model.safetensors` | `26bb5fc5ca75eca215081699e5d8a1dc64f69153431a4e78d6e87ebaee131a0f` |
| `base/tokenizer.json` | `357daed13ffd789678235cb2e8e25a1e57800f569fa59b99339e165e8860db0e` |
| `base/config.json` | `846f3f6eeed33aeb5c6260fb0a1316cc19b9888ae6c3a8da5ec61c9239ff1534` |

### 5.3 v21.2 adapter and merge hashes

| File | SHA-256 |
|---|---|
| `adapter/adapter_model.safetensors` | `aaf155997d3c7b9ccea53d477a59a93f0460dd8d26226b50831233a42db9e5e6` |
| `adapter/tokenizer.json` | `9041375d4d92d6b87628b57b64103d0ce1974559b7ccf146e871656b754fc8ed` |
| merged `model.safetensors` | `7f9d0fe325e9e4568e45f13179adb336b93bbd53e83ddab2826e999eba3c76f7` |

### 5.4 LoRA configuration

```text
rank                     64
alpha                    128
dropout                  0.0
target modules           q_proj,k_proj,v_proj,out_proj,fc1,fc2
bias                     none
weight tying requested   true
task type                SEQ_2_SEQ_LM
PEFT version             0.19.1
```

## 6. Training corpus

### 6.1 The 20,047-sentence synthetic corpus

The governed synthetic corpus contains 20,047 English-Kuku Yalanji pairs, 135,547 Kuku Yalanji word tokens, and no
open review rows at closure. Its final disposition was 19,136 accepted and 911 revised. Whole-corpus lint passed
20,047/20,047 rows. Normalized Kuku-side duplicate count and normalized pair duplicate count were both zero.

The complete process release also preserves:

| Record family | Count |
|---|---:|
| Dictionary entries | 2,688 |
| Working lexemes | 1,870 |
| Reviews | 101,235 |
| Revisions | 1,555 |
| Lessons | 8,035 |
| Process events | 4,745 |

Rows were authored and reviewed against the project's source-preserving dictionary, Elisabeth Patz's grammatical
description, explicit morphology and frame records, batch review, lint, phantom-lexeme discipline, revision logs,
and closure audits. Read the [Operator's Guide](/docs/operator-guide.html), [Grammar Cheatsheet](/docs/grammar-cheatsheet.html),
and [Dictionary Errata](/docs/dictionary-errata.html) for the living linguistic method.

This is process verification, not community certification. The label
`project_approved_synthetic_pending_elder_verification` must not be shortened to `speaker verified`.

### 6.2 Canonical split and deterministic quarantine

The canonical corpus split was:

```text
train       16,820
development  1,609
test         1,618
total       20,047
```

The shared v21 leakage treatment preserved 190 excluded rows in a quarantine ledger rather than deleting them:

```text
effective synthetic train       16,642
validation                       1,609
effective synthetic test         1,606
quarantine                         190
```

Source keys use NFKC, case folding, normalized quotation marks, and collapsed whitespace. Target-surface keys use
NFKC, case folding, and Unicode letters only. After quarantine, train/validation/test intersections were zero for
normalized source, target surface, and exact pair.

### 6.3 v21.2 balanced-replay treatment

The final training mixture per epoch was:

| Component | Rows |
|---|---:|
| Leakage-audited synthetic train | 16,642 |
| Bible direct replay | 2,047 |
| Bible reference-conditioned replay | 2,047 |
| Dictionary-usage replay | 1,428 (357 base rows x4) |
| Elder evaluation rows | 0 |
| **Total** | **22,164** |

The intended canonical replay mix had 22,198 rows. Ten replay collision records were found before training and
preserved in `quarantine/replay_overlap.jsonl`: one synoptic Bible surface and eight dictionary-usage base rows,
corresponding to 34 would-be training rows after paired Bible and x4 dictionary expansion. The effective replay
mixture had zero normalized source, target-surface, or exact-pair intersections with validation, synthetic test,
or the locked external evaluations. Bible canonical references and dictionary word IDs were also checked.

### 6.4 Public and locked data

The public model-specific archive contains the exact 22,164-row train file, 1,609-row validation file, 1,606-row
synthetic test, 128-row pipeline gate, both quarantine ledgers, checksums, and the complete construction manifest.

The 43-row elder-shared control and other locked external evaluation files are deliberately excluded. They had
zero training rows and remain unavailable as training material so that future evaluation is not invalidated. The
full internal evaluation battery also included 84 held-out dictionary-usage rows and two 325-row Bible conditions.

## 7. How v21.2 was trained on RunPod

### 7.1 Experimental ownership

`Claude` in the run ID means the balanced-replay lane was prepared and operated by Claude under the frozen paired
experiment. It does not mean Claude itself is the translation model. The trained network is the NLLB-derived model
described above. v21.1 was the Codex-operated synthetic-only lane; v21.2 changed the training mixture.

### 7.2 Data builder

The fail-closed builder was:

```text
training/translation/build_v21_2_claude_balanced_replay.py
SHA-256 542b1209e814966a4488620d3d86f297bb5d2c47b1374c6b6688cf2b0cee2930
```

It refused a nonempty output directory, verified the final corpus audit, pinned input and frozen-base hashes,
reproduced the v21.1 quarantine, checked replay against every held-out family, stable-sorted by a recorded seed,
asserted row counts and unique IDs, and emitted JSONL plus SHA-256 manifests.

The builder invocation was:

```bash
python3 training/translation/build_v21_2_claude_balanced_replay.py \
  --output-dir /workspace/v21.2-claude/data
```

For external reproduction, download the already built exact training archive. Rebuilding from raw sources requires
the governed internal source roots and locked controls; it is intentionally not equivalent to concatenating the
public 20,047-row file with arbitrary replay data.

### 7.3 Mandatory gate

The full run could not start until a deterministic 128-row mixed-domain gate passed:

```text
gate composition       96 synthetic, 13 Bible direct, 12 Bible ref, 7 dictionary usage
gate epochs            100
gate learning rate     1e-4
acceptance              128 rows, 0 empty, >=115 exact, chrF >=97, length ratio 0.90-1.10
observed                128/128 exact, chrF 100.0, mean length ratio 1.0
```

The gate establishes pipeline and optimization viability. It is deliberate memorization and is not a
generalization score.

### 7.4 Full optimization contract

```text
train / validation / test    22,164 / 1,609 / 1,606
epochs                       3
learning rate                2e-5
batch size                   8
gradient accumulation        2
effective batch              16
warmup                       0
weight decay                 0
source / target cap          192 / 208 tokens during training
LoRA rank / alpha / dropout  64 / 128 / 0.0
seed                         42
optimizer steps              4,155
selection                    validation chrF
archived evaluation decoding beam 1, no repetition constraints
```

The completed A40 run recorded:

```text
training runtime             3,431.87 seconds (57.20 minutes)
training samples/second      19.375
training loss                1.199454
recorded training FLOPs      2.609e16
peak GPU memory              30,383 MiB
```

### 7.5 Full driver

The completed orchestration driver was:

```text
training/translation/run_v21_2_claude_balanced_replay.sh
SHA-256 81beaf25d64b72cfce1a1aef14ff1566a04ffaa0b1df634587505b4bd071ef67
```

Its logical invocation was:

```bash
PYTHON=/opt/mobtranslate-mt-venv/bin/python \
DATA_ROOT=/workspace/v21.2-claude/data \
BASE_MODEL=/workspace/v21.2-claude/base_model/merged \
OUT_ROOT=/workspace/v21.2-claude/models \
LOG_ROOT=/workspace/v21.2-claude/logs \
REPORT_ROOT=/workspace/v21.2-claude/reports \
bash training/translation/run_v21_2_claude_balanced_replay.sh
```

The driver verified checksums, ran tokenizer-length preflight, executed the mandatory gate, trained, merged the
adapter, evaluated all frozen sets, retained row-level predictions, sampled GPU/CPU/RAM/disk resources every five
seconds, and wrote `RUN_COMPLETE` only after every stage succeeded.

### 7.6 Reproducibility boundary

The sealed record includes the exact data manifest, data checksums, model manifests, adapter, merged model,
row-level predictions, metric outputs, resource observations, and final archive checks. It did **not** copy and hash
every Python helper into the v21.2 run directory. The builder and driver above have pre-run timestamps and recorded
hashes; shared trainer/evaluator files were modified later for subsequent work. Therefore this guide does not claim
that the current repository contains a byte-identical snapshot of every executed helper. This is a provenance gap,
not a reason to alter the historical run. Future runs must seal all executed source files and environment locks in
their archive before pod deletion.

## 8. Evaluation

### 8.1 Frozen automatic results

| Set | Rows | v21.2 BLEU | v21.2 chrF++ | Exact | Interpretation |
|---|---:|---:|---:|---:|---|
| Synthetic development | 1,609 | 26.20 | 52.85 | 396 | Same authored process as training |
| Synthetic test, tagged | 1,606 | 26.84 | 53.84 | 426 | Primary synthetic held-out |
| Synthetic test, untagged | 1,606 | 26.77 | 53.81 | 429 | Prefix-dependence control |
| Elder-shared sentence pairs | 43 | 6.52 | 29.36 | 0 | Natural-register transfer remains weak |
| Dictionary usage | 84 | 11.58 | 47.14 | 3 | Replayed-domain retention |
| Bible direct | 325 | 12.36 | 43.78 | 0 | Replayed-domain retention |
| Bible reference-conditioned | 325 | 12.41 | 43.91 | 0 | Replayed-domain retention |

All fourteen model/set combinations in the v21.1-v21.2 comparison had zero empty outputs. BLEU and chrF++ measure
surface overlap; they do not establish adequacy, grammaticality, morphology, pragmatics, dialect fit, cultural
appropriateness, or community acceptance.

### 8.2 v21.2 versus v21.1

Relative to the synthetic-only v21.1 ablation, v21.2 changed corpus chrF++ by:

```text
synthetic tagged test       -0.87
synthetic untagged test     -0.58
elder-shared                +0.56
dictionary usage           +12.74
Bible direct               +15.23
Bible reference            +15.15
```

On the primary tagged synthetic set, paired mean sentence chrF++ was +0.082 for v21.2 with 95% bootstrap interval
[-0.225, +0.396]. That is synthetic parity, not a clean win. The dictionary and Bible retention gains were broad
and large, which is why v21.2 is the better practical recipe. v21.2 had 4,155 optimizer steps versus v21.1's 3,120,
so the comparison does not isolate replay composition from update budget.

Read the [independent paired comparison](/docs/model-comparison.html) and the broader
[translation-systems benchmark](/docs/translation-systems-benchmark.html) for bootstrap methods, linguistic
slices, LLM post-edit comparisons, and the routing decision.

### 8.3 Guarded decoder transfer

The v22 step-matched run did not produce a promotable checkpoint, but its development-only decoder selection yielded
a transferable serving policy. A separate protocol froze that policy before any v21.2 transfer output was
generated. The exact published v21.2 merged model hash was verified before inference.

| Set | Greedy chrF | Guarded chrF | Exact greedy / guarded | Repeated-segment rows greedy / guarded |
|---|---:|---:|---:|---:|
| Synthetic dev | 52.8471 | 55.1631 | 396 / 394 | 39 / 1 |
| Synthetic tagged | 53.8374 | 55.7097 | 426 / 426 | 37 / 2 |
| Synthetic untagged | 53.8142 | 55.7278 | 429 / 422 | 36 / 2 |
| Elder diagnostic | 29.3558 | 29.3565 | 0 / 0 | 0 / 0 |
| Dictionary usage | 47.1391 | 47.9648 | 3 / 1 | 0 / 0 |
| Bible direct | 43.7751 | 43.7197 | 0 / 0 | 2 / 0 |
| Bible reference | 43.9093 | 43.8353 | 0 / 0 | 1 / 0 |

All seven sets had zero empty outputs and all 14 frozen transfer checks passed. In 50,000 paired bootstraps, the
mean per-sentence chrF++ interval was positive on tagged synthetic test `[+0.097, +0.560]`; intervals for untagged,
elder, usage, and both Bible controls included zero. The exact-match decreases are visible in the table and were not
used as tuning criteria.

The policy is an inference intervention on unchanged weights. Kuku Yalanji has linguistically meaningful
reduplication, so `no_repeat_ngram_size=4` remains subject to speaker review for false suppression of valid
morphology. Read the [full v22 and decoder-transfer report](/docs/kuku-v22-experiment.html).

### 8.4 Isolated lexicon and elder version probe

A separately preregistered diagnostic compared the reconstructed step-2,770 checkpoint, the failed v22.0
step-3,120 checkpoint, and the current step-4,155 v21.2 weights. All three models used the same guarded decoder:
one beam, four-token no-repeat guard, repetition penalty 1.10, length penalty 1.0, and maximum 208 new tokens.
There were no empty outputs in any of the 1,020 model-row evaluations.

The lexicon builder began with 1,640 JSON-LD entries, retained 395 lowercase single-gloss lexical entries, and
grouped synonymous headwords into 297 unique English prompts. A response counted as normalized exact when it
matched any canonical Kuku Yalanji headword attached to that gloss. None of the 297 English prompts occurred as an
isolated training input, but the dictionary informed earlier corpus work, so this is a coverage probe rather than
an untouched test set.

| Model checkpoint | Accepted exact | Mean maximum sentence chrF++ | Mean minimum CER | Empty |
|---|---:|---:|---:|---:|
| Step 2,770 | 42/297 (14.14%) | 28.06 | 0.807 | 0 |
| Step 3,120 | 43/297 (14.48%) | 28.83 | 0.811 | 0 |
| Step 4,155 guarded | 48/297 (16.16%) | 30.38 | 0.782 | 0 |

On identical rows, step 4,155 exceeded step 3,120 by 1.68 exact-match percentage points; the 10,000-sample paired
bootstrap interval was [+0.34, +3.37]. Its mean sentence chrF++ difference was +1.55 with interval
[+0.37, +2.95]. This is evidence that continued exposure improved canonical lexical retrieval on this probe. It
does not establish that step 4,155 generates better free translations.

The exposure split is the more important result. Step 4,155 retrieved 46/192 (23.96%) accepted forms that appeared
as full tokens somewhere in training, but only 2/105 (1.90%) forms not observed that way. Excluding 14
identity/loanword cases, exact retrieval was 43/283 (15.19%). A lookup-first dictionary route is therefore not just
an optimization: it is required for reliable known-word answers.

The same run evaluated the 43 rights-cleared elder sentence pairs, which were excluded from training but had been
observed in earlier evaluations:

| Model checkpoint | Exact | Corpus chrF++ | Mean sentence chrF++ | Empty |
|---|---:|---:|---:|---:|
| Step 2,770 | 0/43 | 29.18 | 29.63 | 0 |
| Step 3,120 | 0/43 | 29.26 | 29.66 | 0 |
| Step 4,155 guarded | 0/43 | 29.43 | 29.89 | 0 |

Every paired elder interval included zero. For step 4,155 minus step 3,120, the mean sentence chrF++ difference was
+0.23 with interval [-0.42, +0.99]. There is no evidence here of improved natural elder register.

The public audit package includes the [preregistration](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/lexicon-elder-benchmark-20260714/PREREGISTRATION.md),
[run contract](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/lexicon-elder-benchmark-20260714/RUN-CONTRACT.md),
[340-row combined input](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/lexicon-elder-benchmark-20260714/combined_lexicon_elder_probe.eng-gvn.jsonl),
[row-level predictions and paired statistics](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/lexicon-elder-benchmark-20260714/version_probe_analysis.json),
[concise results](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/lexicon-elder-benchmark-20260714/VERSION-PROBE-RESULTS.md),
and [package checksums](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/lexicon-elder-benchmark-20260714/SHA256SUMS).

### 8.5 Sealed RunPod reproduction and fast feedback loop

The complete three-checkpoint probe was rerun twice on an exact NVIDIA A40 runtime with deterministic float32
inference, sampling and TF32 disabled, and all behavior-affecting model files pinned. The final A40 outputs matched
the frozen local CPU baselines on all 1,020/1,020 rows. The two A40 executions also matched each other on all
1,020/1,020 rows. The final run's output inventory passed remotely and again after download.

The portable kit verifies input/model hashes and runtime versions, evaluates all manifest entries, computes the
paired analysis, compares GPU rows with local baselines, records real GPU/process resources, and writes
`RUN_COMPLETE` only after its own checksum inventory verifies. All three inference passes took about 3.5 minutes;
a new checkpoint already present on a training pod should return a sealed result in roughly one to two minutes.

Read the [complete RunPod evaluation and fast-loop manual](/docs/kuku-runpod-evaluation-loop.html), or download the
[sealed kit](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/lexicon-elder-benchmark-20260714/runpod-evaluation-kit-v1.tar.gz)
and [A40 result bundle](/model-artifacts/kuku-yalanji-nllb-lora/v21.2-claude-balanced-replay-guarded-20260714/lexicon-elder-benchmark-20260714/kuku-yalanji-runpod-a40-benchmark-results-20260714.tar.gz).

### 8.6 Known failure modes

- Both v21 models are 0/43 exact on the elder-shared control.
- The current model retrieves only 48/297 accepted dictionary forms on the isolated-gloss probe, falling to
  2/105 when the target form did not appear as a full training token.
- v21.2's mean elder output length is about 68% of reference length; multi-clause content is often compressed.
- The archived greedy policy produced 37 tagged-test repeated-segment rows; the guarded policy reduced this to 2
  but does not guarantee that every novel output is loop-free.
- The no-repeat guard can in principle suppress valid reduplication and needs fluent-speaker error analysis.
- Replay improves adjacent-domain retention but does not prove unseen-domain or natural-register transfer.
- The synthetic test reflects the same authoring process and has strong internal difficulty stratification.
- One seed and one run per treatment do not estimate run-to-run variance.
- No reverse-direction, calibrated uncertainty, arbitrary-input robustness, or complete speaker evaluation is
  claimed for this release.

## 9. Production safeguards

A responsible integration must use the model as one component in a routed translation system:

1. Return an exact rights-cleared sentence/resource match when one exists.
2. Return a unique exact dictionary match for a clearly lexical query.
3. Generate only for novel text within the supported direction and length.
4. Reject empty output, target-token failure, length-limit termination, severe under-translation, severe
   over-generation, and repeated-segment loops.
5. Label every accepted generation as a research draft and expose model version and provenance.
6. Preserve the source, draft, guard result, and any speaker correction as separate versioned records.
7. Never promote a generated form into the dictionary or training corpus without adjudication, rights metadata,
   dialect/register metadata, and a supersession trail.

LLM post-editing remains experimental. On the frozen 43-row control, Codex and Claude post-editors improved point
scores, but neither route was speaker-reviewed and their mutual difference did not cleanly exclude zero. Do not
silently add a hosted LLM post-editor and present the result as verified language.

## 10. Rights, attribution, and cultural responsibility

The upstream NLLB checkpoint is distributed under
[CC-BY-NC-4.0](https://huggingface.co/facebook/nllb-200-distilled-1.3B). This release is therefore for
noncommercial research use. The complete process corpus does not assert a blanket Creative Commons or public-domain
license over community language knowledge, dictionary material, grammatical descriptions, or third-party sources.
The download does not relicense those sources.

Each training row carries its own provenance and rights fields. Synthetic rows are labelled
`project_approved_synthetic_pending_elder_verification`; replay rows retain their source rights metadata. Public
availability is not speaker endorsement and not an unrestricted commercial grant. Before redistribution,
production deployment, or derived commercial work, conduct an independent rights review and consult the relevant
Kuku Yalanji community and source rights holders.

Do not use this model for high-stakes communication, legal or medical interpretation, cultural authority, language
teaching presented as certified, or replacement of fluent-speaker consultation.

## 11. Versioning and update procedure

The immutable version string, manifests, checksums, and archives define this release. Do not overwrite an archive
in place. A future model must receive a new version, new artifact URLs, new checksums, a new model card/guide, and a
registry entry that keeps v21.2 available as historical evidence.

For each future run:

1. Freeze and checksum data before training.
2. Seal every source file, dependency lock, container image ID, environment variable contract, and command.
3. Keep train, validation, test, locked external controls, and quarantine explicitly separated.
4. Record failed gates and provisioning failures as evidence.
5. Retain adapter, merged model, row-level predictions, metrics, resources, and all checksums.
6. Run speaker evaluation before any promotion claim.
7. Build immutable host bundles and verify full archive reads.
8. Add the release to `apps/web/public/models/registry.json`.
9. Rerun `render-docs.sh` so the live books and raw Markdown update together.
10. Verify API JSON, CORS, complete downloads, byte ranges, checksum values, smoke outputs, and responsive pages.

## 12. Related records

- [Model registry and downloads](/models)
- [v22 step-matched replay and decoder-transfer experiment](/docs/kuku-v22-experiment.html)
- [Independent v21.1-v21.2 model comparison](/docs/model-comparison.html)
- [Translation systems research and benchmark](/docs/translation-systems-benchmark.html)
- [Synthetic corpus operator guide](/docs/operator-guide.html)
- [Mega grammar cheatsheet](/docs/grammar-cheatsheet.html)
- [Dictionary errata and phantom watchlist](/docs/dictionary-errata.html)
- [Sample production pages](/docs/sample-pages.html)
- [Mi'gmaq model training handoff](/docs/migmaq-model-training.html)
- [Mi'gmaq v1 model card](/docs/migmaq-model-card.html)
- [Mi'gmaq evaluation and promotion report](/docs/migmaq-evaluation.html)
- [RunPod evaluation and fast feedback loop](/docs/kuku-runpod-evaluation-loop.html)
- [MobTranslate Next 16 build incident and deployment runbook](/docs/mobtranslate-next16-build-incident.html)
- [Raw Markdown for this guide](/docs/kuku-v21-2-model-guide.md)

## 13. Live distribution deployment verification

The public distribution, inference, documentation, and activity-logging surfaces were rebuilt and verified end to
end on 2026-07-14 UTC. This is an operational verification of delivery and metadata integrity; it is not an
additional linguistic evaluation or a promotion of the model beyond `research-only`.

### 13.1 Build and service gate

- Next.js 16.2.9 completed an optimized **Turbopack** build for all 361 static and dynamic routes. The final build
  compiled in 74 seconds, completed strict TypeScript in 3.1 minutes, and generated 361/361 pages in 55 seconds.
  End-to-end wall time was 402 seconds while the shared data disk was under independently measured contention.
- The immutable deployed build ID is `ZS7-SuKDwqjovoz_tBnyE`. It was built into the isolated
  `.next.release-turbopack-final-20260714T1347Z` tree, started on port 3310, and became ready in 500 ms before the
  atomic production swap. The immediately preceding `LUCL6-Xz-yvnyYPvmgJJa` tree remains at
  `/mnt/donto-data/workspace/.build-archive/mobtranslate.com/20260714/.next.rollback-LUCL6-Xz-yvnyYPvmgJJa-20260714T135544Z`.
- Thirteen focused model-distribution, bilingual Discord, saved-result, and contact-link tests passed across four
  files. The production build supplied the strict TypeScript and prerender gates.
- `mobtranslate-web.service` is active on Caddy's configured upstream `127.0.0.1:3300`, has zero restarts, and became
  ready in 299 ms after the final swap. Both custom CPU inference services, `kuku-v2-infer.service` and
  `mobtranslate-migmaq-v1-infer.service`, are inactive and disabled; ports 7955 and 8765 have no listeners.
- The homepage translation route now uses the complete selected dictionary with `gpt-5.4-mini` and a required,
  Zod-validated tool call. The custom-model lab and versioned inference routes fail closed with explicit 503
  `not_configured` responses while model serving is paused.
- The final browser has no application console errors. Vercel Analytics is now rendered only when `VERCEL` is set,
  avoiding a failed `/_vercel/insights/script.js` request on this self-hosted deployment.
- The only retained notices are Next's deprecated `middleware.ts` convention and Node's experimental SQLite API.

The build incident was not a single compiler defect. The previous command forced Webpack on a Next 16 application,
an obsolete custom webpack alias duplicated native monorepo support, Tailwind v4 could discover retained `.next.*`
rollback trees inside the source repository, and broad runtime filesystem access inflated Next file traces. Retained
builds now live outside the repository, `.next.*/` is ignored, the app uses native `transpilePackages` plus a pinned
Turbopack root, and runtime filesystem boundaries are explicit. All 207 final route trace manifests were audited;
none retained an unscoped project or prior-`.next` trace. The full evidence and recovery procedure are in the
[build incident runbook](/docs/mobtranslate-next16-build-incident.html).

### 13.2 API contract gate

The live assertions covered all of the following:

| Check | Observed result |
|---|---|
| `GET /api/v1/models` | 200; two model families |
| Registered releases | 21 total; 20 Kuku Yalanji and one Mi'gmaq |
| Kuku Yalanji `versions/latest` | `v21.2-claude-balanced-replay-guarded-20260714` |
| Mi'gmaq `versions/latest` | `1.0.0-rc1` |
| Rejected v22.0 record | Public negative result; no hostable model download |
| Rejected v23.0 record | Public attested-adaptation report; all candidate weights deleted |
| Historical v21.2 record | Original greedy serving-policy release retained immutably |
| Version benchmark metadata | 48/297 exact; 16.1616%; frozen row-level audit artifacts |
| OpenAPI | 200; OpenAPI 3.1.0; canonical server `https://mobtranslate.com` |
| `OPTIONS` | 204 with `Access-Control-Allow-Origin: *` |
| `HEAD` | 200 with the public CORS contract |
| Unknown model and version | structured 404 responses with distinct error codes |
| Shared-cache policy | `public, s-maxage=300, stale-while-revalidate=3600` |
| Private-path audit | no `localPath`, `/mnt/`, `localhost`, `127.0.0.1`, or port 7955 in public JSON |
| Homepage prompt translation | 200 with nonempty structured output and English gloss |
| Versioned custom inference | Explicit 503 `not_configured`; no private endpoint exposed |

The route resolves public links from the configured `NEXT_PUBLIC_APP_URL=https://mobtranslate.com` and validates
every fallback origin. Public model JSON and custom-inference responses expose neither reverse-proxy links nor a
private inference service location.

### 13.3 Artifact and metadata gate

- Every registered release was fetched through its literal version endpoint.
- The registry contains 147 artifact records across 21 releases. Of these, 129 records are marked available and
  resolve to 125 unique public URLs. The current release contributes 33 artifacts: 16 evaluation, six
  documentation, six metadata, three dataset, one bundle, and one standalone-model record.
- Every benchmark package member reverified against `SHA256SUMS`; failures: zero. The public result Markdown,
  analysis JSON, frozen input, manifest, preregistration, run contract, and prediction hashes match the registry.
- The sealed RunPod A40 package evaluated 340 rows at each of three checkpoints, producing 1,020 predictions. Both
  A40 executions matched each other and the frozen local CPU baselines on all 1,020 rows.
- The v21.2 Atlas and merged archives passed full local archive traversal. Their live endpoints expose the exact
  advertised `Content-Length`; the Atlas endpoint also returned 206 with a correct 1,024-byte range and total size.
- The exact training archive passed full traversal and its live endpoint reported the advertised size. The raw
  Markdown guide and experiment report downloaded byte-identically to their rendered source copies.
- Safe live Discord webhook probes showed English and Indigenous-language fields for translation and TTS. Unit
  coverage verifies the same bilingual labelling for recording uploads; chat and translation routes pass both
  sides into the shared logger.

### 13.4 Browser gate

The homepage, registry, translation lab, documentation index, and this guide were inspected in real Chromium
sessions, including the final live build at 1280x720 and 390x844:

- `/models` had no horizontal overflow at either width; long model/version identifiers wrapped without overlap.
- The model API, hosting guide, v22 negative-result report, OpenAPI, model archive, benchmark inputs, predictions,
  checksums, dataset, and evaluation links were present with canonical public URLs.
- Light and dark themes were legible on the registry and guide.
- Responsive checks found no horizontal overflow or incoherent overlap on the tested registry, lab, or docs pages.
- At 1280 px the compact navigation remained available without clipping. At 390 px the menu exposed all nine
  destinations -- About, Atlas, Dictionaries, Map, Spread, Translate, Models, Learn, and Community -- without
  horizontal overflow or obscured controls.
- `/docs/` linked the rendered guide and v22 report together with their raw Markdown sources.

## 14. Citation record

When reporting this release, identify at least the model ID, immutable version, direction, upstream lineage,
training-mixture count, evaluation scope, status, and access date:

```text
MobTranslate. 2026. Kuku Yalanji NLLB-LoRA,
v21.2-claude-balanced-replay-guarded-20260714 (byte-identical weights from
v21.2-claude-balanced-replay-v2-candidate). English (eng_Latn) to Kuku Yalanji
(gvn_Latn). Noncommercial research release, not speaker-certified.
https://mobtranslate.com/api/v1/models/kuku-yalanji-nllb-lora/versions/
v21.2-claude-balanced-replay-guarded-20260714
```