# v8 = arm A of the vipafix-20260910T175025Z campaign

Installed 14 September 2026. `phase3_best.pt` here is a COPY of
`/ix1/ishi/models/phonetic/checkpoints/vipafix-20260910T175025Z/phase3_best.pt`,
verified equal by md5 at copy time:

    51d0f7f8e8c4e0c72454803c154bb419  phase3_best.pt

Vocabulary that MUST be paired with it (`/ix1/ishi/models/phonetic/data/v8/vocab`):

    16a4d41e868fecc16e1b701363e1bda8  char_vocab.json   (114,845 chars)
    17ce7fe69a836ddd32c9577cdd60f389  lang_vocab.json   (2,438 langs)
    18fd9410e1c543429ec8327a29ee3225  script_vocab.json (37 scripts)

Pairing this checkpoint with the v7 vocabulary does NOT raise — it produces
plausible garbage. v7 carries 113,280 / 20 / 1,944.

## Why arm A and not one of the other three

Four candidates were trained and scored on one frozen test set
(`developer/baselines/v8-report-v8{A,B,G,M}.json`, and v7 for the baseline).
Decision recorded in `developer/plan-symphonym-v8.md` sections 80 and 81 under a
rule fixed BEFORE any candidate finished: better only if recall_at_gate does not
fall AND separability improves.

                    v7      A       B       G       M
  recall_at_gate    87.5    95.2    95.5    95.3    94.2
  separability   -0.2127 -0.0574 -0.0510 -0.0593 -0.0758
  permutation       70.5     5.0     6.6     6.0     4.4
  historic_hard   0.5179  0.5179  0.4995  0.5081  0.5081
  latin_latin     0.8527  0.8495  0.8370  0.8464  0.8370

B and G lead the headline by noise (0.3pp ~ 12 pairs in 4,000); A leads the
collateral measures by margins that are not noise, and is the only candidate
with NO historic-Latin regression at all.

## Do not select by training loss

Arm M had the best val_loss at all three phases (0.0051 / 0.0467 / 0.0206,
beating A at every one) and is the worst of the four models. Phase loss orders
candidates within a fixed batch size and does not compare across recipes.

## final_model.pt is NOT this file

The gateway loads `<base>/models/phonetic/checkpoints/v<version>/final_model.pt`,
and that name already denotes DIFFERENT CONTENT on /vast and /ix1 for v7
(see developer/baselines/v7-artefacts-witness.md). Export deliberately and
verify by hash; never copy name-to-name across the two filesystems.