# v22 Secondary Decoder-Transfer Protocol

**Frozen:** 2026-07-14T03:29:00Z  
**Status:** secondary confirmatory serving experiment; not an amendment to the v22.0 checkpoint promotion gate  
**Owner:** Codex  

## Question

The primary v22.0 experiment changes training exposure by stopping the sealed v21.2 trajectory at optimizer step
3120. Its decoder was selected from v22.0 development data only and locked at 2026-07-14T03:16:28Z as:

- `num_beams=1`
- `no_repeat_ngram_size=4`
- `repetition_penalty=1.10`
- `length_penalty=1.0`

The primary checkpoint gate has already failed because greedy v22.0 produced 37 repeated-segment rows on the tagged
synthetic test, above the preregistered maximum of 25. That gate remains failed. This separate experiment asks a
different operational question: **does the decoder chosen without v21.2 development or test output transfer to the
existing v21.2 final checkpoint and improve output safety without materially degrading held-out quality?**

## Frozen inputs

- Model: `v21.2-claude-balanced-replay-gvn-3epoch-lr2e-5`, final optimizer step 4155.
- Merged model SHA-256: `7f9d0fe325e9e4568e45f13179adb336b93bbd53e83ddab2826e999eba3c76f7`.
- Adapter SHA-256: `aaf155997d3c7b9ccea53d477a59a93f0460dd8d26226b50831233a42db9e5e6`.
- Base: the same frozen v12 `gvn_Latn`-extended NLLB model used in v21.2 and v22.0.
- Evaluation files: the exact checksum-sealed v21.2/v22.0 development, tagged test, untagged test, 43-row
  elder-shared diagnostic, 84-row dictionary-usage control, and two 325-row Bible controls.
- Baseline: archived v21.2 greedy predictions, generated with beams 1, no n-gram blocking, repetition penalty 1.0,
  and length penalty 1.0.

No decoder value may change after this document is frozen. No development selection is repeated on v21.2. Each
v21.2 transfer prediction set is generated once with the decoder above.

## Transfer gates

All checks must pass for a serving-recipe update:

| Set | Frozen requirement |
|---|---:|
| Every set | 0 empty outputs |
| Synthetic tagged test | chrF++ >= 52.8374 and repeated-segment>=10 rows <= 5 |
| Synthetic untagged test | chrF++ >= 52.8142 and repeated-segment>=10 rows <= 5 |
| Dictionary usage | chrF++ >= 46.1391 |
| Bible direct | chrF++ >= 42.7751 |
| Bible reference | chrF++ >= 42.9093 |

The chrF floors are the archived v21.2 greedy score minus one absolute point. The loop floors are deliberately much
stricter than v21.2 greedy's 37 tagged-test failures. Exact-match counts, length ratios, maximum repetition, BLEU,
and sentence-level paired deltas must be reported even though they are not tuning criteria.

The 43-row elder-shared set remains diagnostic only: report chrF++, exact match, length behavior, and examples, but
do not use it to claim speaker validity or tune the decoder.

## Decision language

- **PASS:** the same trained v21.2 model may receive a versioned guarded-decoding recipe. This is an inference
  safety improvement, not a new trained model, not speaker certification, and not approval for authoritative use.
- **FAIL:** retain the current greedy v21.2 serving recipe and preserve all outputs as negative evidence.

The original v22.0 checkpoint promotion result remains independent and cannot be changed by this transfer test.