Results and model lineage

The deployed model: jepa-asl-b3

jepa-asl-b3 (2 October) is one student with the phonological head, distilled from twelve fine-tunes whose encoders first learned American Sign Language: Google's Isolated Sign Language Recognition corpus — 94,290 landmark clips of 250 ASL signs by 21 Deaf signers. It replaced jepa-x12 as the deployed 260-sign model by one to three clips on identical inputs, and its confidence is better calibrated (expected calibration error 0.039 against 0.056). What the ASL data did and did not do is below.

What changed on 2 October — ASL transfer

Twenty-one Deaf ASL signers are more signer variety than every Indian corpus this program has combined, so the question was whether learning ASL first makes the encoder more signer-invariant for ISL. All arms ran on the v4 recipe, four seeds each, against a same-harness control:

Two follow-ups on 3 October, both negative and kept here:

• A small ensemble in the browser. Two or three of the int8 models run side by side and combined with the app's own mirror rule, on sets named before scoring: none beat jepa-asl-b3 alone on the unseen signer (best 0.5091 against 0.5136); the best sets add about 0.8 pp on the accuracy test for three times the download and inference. Not shipped.
• Self-supervised pre-training on the ASL clips. The Phase B encoder trained further on the unlabelled pool with the 94,290 ASL clips added, against the same training on the pool alone, each then fine-tuned four times: unseen signer 0.4273 and 0.4250 against 0.4296 for Phase B itself. The ASL clips help only when their labels are used.

Before it: jepa-x12

jepa-x12 (26 September) is twelve fine-tunes distilled into one Phase B student of the same size as jepa-rr4, trained on the reviewed v4 dataset pack. It was the deployed default from 27 September (the model frozen for Review 2) to 2 October. Most of its gain over jepa-rr4 is on CISLR (+13.5 pp); the unseen signer moved +1.8 pp. The measurements, the seed tables and that reading are on the Review 2 results page.

Deployed until 27 September: jepa-rr4

jepa-rr4 is a distilled student of a four-seed ensemble, fine-tuned with the all+synth:full@2 + rr recipe on the Phase B encoder cut at layer 4, exported to int8. How it is made is the training program; how it is scored is the gate.

Model lineage

All numbers are cross-signer RKMVU top-1 unless the column says otherwise. "Gate" is the two-corpus number that decides.

What changed on 3 September

A night and a day, in the order it happened.

The recipe screen on the wrong encoder

Ten candidate recipe changes were screened, two seeds each, against an in-harness control — and nothing was positive on the gate:

The re-screen on the real encoder

On the Phase B encoder, the in-harness control reproduces the champion exactly (per-seed gates .4140 / .3923 / .3773 / .3940). Against it, rr is the one winner: four seeds at .4057 / .4174 / .4124 / .4107, +1.7 pp gate, +2.5 pp RKMVU, three of four seeds up, and half the seed spread. swad −0.3, angles +0.1 (mixed), lw8 −1.4. rr drops up to 25% off either end of a clip and stretches it back to 64 frames, with probability 0.5.

Ensemble, distil, ship

The four rr seeds were ensembled (gate 0.4574; 0.4641 with mirror TTA) and distilled into jepa-rr4 at 14:25 — the current best, exported and dropped into the Lab.

The Phase C pilot

In parallel, four pretraining objectives were compared at 12 epochs on a random 30k pool subset: the all-layer teacher gives +4.8 pp gate at equal budget; motion targets hurt. Phase C proper was launched on the full pool. The table and the reasoning are on the JEPA page.

Also fixed on 3 September

• Demo-table leak. Stale windows from a previous capture epoch could be voted into the next row ("husband 77%" on the "dead" row). Fixed with capture epochs in LandmarkRing, a capture-start message to every worker, and a per-row window vote — runtime page.
• Five training-side bugs: the LP-FT encoder learning rate stuck at 0 (cosine scheduler recursion); a head-slice bias in pool subsets; Colab cell interrupts killing child processes; ONNX missing on fresh VMs; a gate-scorer name mismatch.

Phase C proper — a wash at full budget

Target standard deviation pinned at 0.999 (no collapse); checkpoints to Drive every five epochs; auto-resume proven at the epoch-11 idle-timeout kill. Fine-tuned, jepa_c4 scored 0.4524 on the gate against jepa-rr4's 0.4591 — a wash — and Phase C students lose on the unseen signer. The pilot's +4.8 pp did not survive the full budget. Phase C encoders still serve as eight of the twelve teachers behind jepa-x12.

What changed between 4 and 26 September

The work moved from the model to the labels. A review of the training data for sense conflation — clips filed under one word that are really different signs, and the reverse — produced a reviewed dataset contract (the v4 pack). The same recipe trained on it scored +7.0 pp on the gate over four seeds, all four up; ensembling twelve fine-tunes and distilling them gave jepa-x12. The loop, the seed tables and the negative results are on the Review 2 pages.

What this means for the review targets

Review 1 pre-registered a Phase B target of ≥ 0.50. Phase B alone missed it (0.3273). The fine-tuning program on the Phase B encoder reached 0.4818 on the unseen signer by 4 September, and jepa-x12 stood at 0.5509 on the gate — met — and 0.5000 on the unseen signer (0.4909 int8) — borderline. Its successor jepa-asl-b3 reads 0.5526 and 0.5136 (0.5091 int8), which clears the unseen-signer target by a few clips. The targets are scored in full on the Review 2 outcomes page.