# Kuku Yalanji v22.0 Step-Matched Replay Preregistration

Frozen: 2026-07-14T02:15:40Z  
Owner: Codex  
Status at freeze: no v22 model output has been generated or inspected

## 1. Research question

Does stopping the sealed v21.2 balanced-replay training trajectory at exactly 3,120 optimizer updates reduce
degenerate generation while retaining the adjacent-domain gains of the 4,155-update v21.2 checkpoint?

This is an update-exposure experiment. It is not a new-data experiment, a new-seed replication, or evidence of
speaker-judged translation quality.

## 2. Falsifiable hypothesis

The v21.2 step-4,155 model over-trained on its mixture. A continuation reconstructed from its sealed step-2,770
checkpoint and stopped at step 3,120 will:

1. reduce synthetic-test rows with any orthographic word/morpheme segment repeated at least ten times from 37 to at
   most 25;
2. retain synthetic-test corpus chrF++ within 1.0 point of v21.2 step 4,155;
3. retain dictionary-usage and each Bible control within 2.0 chrF++ points of v21.2 step 4,155; and
4. emit no empty strings on any frozen evaluation set.

Failure of any condition blocks promotion. A negative result is retained and documented rather than hidden.

## 3. Invariants

- Base: the exact frozen v12 merged NLLB-1.3B checkpoint used by v21.2.
- Training data: exact v21.2 `train.eng-gvn.jsonl`, 22,164 rows per epoch.
- Development data: exact v21.2 `validation.eng-gvn.jsonl`, 1,609 rows.
- Test and controls: exact v21.2 synthetic test (1,606), usage (84), Bible direct (325), Bible
  reference-conditioned (325), and elder diagnostic (43).
- Optimizer, scheduler, RNG, and adapter state: restored from the sealed v21.2 `checkpoint-2770`.
- Seed: 42.
- Hyperparameters: unchanged from v21.2: LoRA r=64, alpha=128, dropout=0; target modules
  `q_proj,k_proj,v_proj,out_proj,fc1,fc2`; effective batch 16; learning rate schedule originally planned for 4,155
  steps; zero warmup and weight decay.
- Stop: immediately after optimizer update 3,120. Best-checkpoint reload is disabled so the serialized candidate is the
  actual step-3,120 state.
- No elder row enters training or model/decoder selection.
- The training process does not receive the synthetic test or external controls.

## 4. Reconstruction limitation

The continuation uses the saved optimizer, scheduler, RNG, and trainer state. Its step-2,770 adapter hash must match
the sealed archive before launch. Exact bitwise identity to the unobserved original step-3,120 trajectory cannot be
asserted because that checkpoint was not saved in the original run. The run will report package versions, hashes,
GPU, and the observed final global step.

## 5. Decoder selection

Only the 1,609-row development set is used to select among these predeclared deterministic decoders:

| ID | Beams | no-repeat n-gram | repetition penalty | length penalty |
|---|---:|---:|---:|---:|
| `greedy` | 1 | 0 | 1.00 | 1.00 |
| `nr4` | 1 | 4 | 1.00 | 1.00 |
| `rp110` | 1 | 0 | 1.10 | 1.00 |
| `nr4-rp110` | 1 | 4 | 1.10 | 1.00 |

A decoder is eligible when it has zero empty outputs and development corpus chrF++ is no more than 1.0 point below
the best development configuration. Eligible configurations are selected lexicographically by:

1. fewest rows with a segment repeated at least ten times;
2. fewest severe shape failures (token ratio below 0.5 plus character ratio above 2.0);
3. lowest maximum repeated-segment count;
4. highest corpus chrF++;
5. the table order above, as the deterministic final tie-break.

This prioritizes catastrophic-output control while bounding average reference-score loss. Kuku Yalanji has meaningful
reduplication, so the no-repeat treatment is applied only at four tokenizer tokens and remains subject to later speaker
review; the orthographic detector is a safety diagnostic, not a claim that linguistic reduplication is erroneous.

The selected decoder JSON must be written before any v22 test/control inference begins.

## 6. Promotion gates

Baseline-decoded step 3,120 is compared with baseline-decoded step 4,155 to isolate training exposure. It must satisfy:

| Frozen set | Gate |
|---|---|
| synthetic tagged 1,606 | chrF++ >= 52.8374; repeated-segment rows <= 25; empty = 0 |
| dictionary usage 84 | chrF++ >= 45.1391; empty = 0 |
| Bible direct 325 | chrF++ >= 41.7751; empty = 0 |
| Bible reference 325 | chrF++ >= 41.9093; empty = 0 |

The selected decoder must have zero empty outputs on every set and cannot increase synthetic-test repeated-segment
rows relative to the candidate's baseline decoder. Paired sentence-level chrF++ differences and 95% paired-bootstrap
intervals are reported; corpus thresholds alone do not imply statistical superiority.

The elder 43-row set is an adaptively observed diagnostic, not a blind promotion endpoint. Report exact match, chrF++,
length ratio, and examples, but do not tune on it or claim community validity from it.

## 7. Interpretation boundaries

- All automatic references are imperfect proxies; chrF++ and BLEU do not establish grammaticality, semantic
  equivalence, dialect fit, register, or cultural acceptability.
- The synthetic test measures agreement with the authored corpus, not independent natural-language competence.
- Bible and dictionary controls measure retention in adjacent known domains, not broad transfer.
- This single reconstructed seed does not estimate training variance.
- No checkpoint is called production-ready or speaker-certified without blind Kuku Yalanji speaker review.
- Existing v21 tests have already been observed. They are frozen regression controls, not untouched confirmatory data.

## 8. Required archive

The immutable run archive must contain this preregistration, all exact input/code hashes, copied source code, RunPod
pod metadata, dependency versions, GPU and resource traces, complete stdout/stderr, trainer state, adapter and merged
weights, every prediction row for every decoder/set, selection JSON, paired analyses, checksums, cost/runtime record,
and the post-deletion empty pod-list check. The living books and public model guide must record positive and negative
results alike.