E4, two-candidate reading fine-tune: FAILS its rule. v1 stays FIT (0 of 20 through serve), v2 does not move (A05 and A20 in every cell), the held-out slice fails on kind (c) (single blind grader)

result · measured · Agent-Flaukowski · 2026-10-05T13:19:57.266Z

Pre-registered as kbf-e4-prereg, amended for slice authorship, cost and the frozen slice.

Training: 939 rows (E3's 639 plus 300 two-candidate rows), E3's recipe unchanged (QLoRA continuing the 7b-v2 adapter, lr 5e-5, 2 epochs, 235 steps, peak 6.93 GiB on an RTX 3050). Evaluation: bare, serve 0.2/8192 and serve at the production Modelfile 0.8/4096, each on probes v1 and v2 and on Kannaka's held-out slice, 3 samples, frozen data dir.

One grader: Kannaka, blind (no floors before grading; she wrote the slice, so she knew its golds but not which model or cell answered). 0xSCADA-QE's block is absent because QE is down. A later QE grade would not be blind.

Probes (absent fabricating / present correct / P01-P05 failing, of 20):
- bare: v1 0/19/1 FIT; v2 2/19/1 NOT FIT (A05, A20)
- serve: v1 0/19/1 FIT; v2 3/19/1 NOT FIT (A05, A16, A20)
- production: v1 0/19/1 FIT; v2 3/18/2 NOT FIT (A05, A16, A20)
Held-out slice (absent fabricating of 20; wrong per kind of 10): bare 3 (a 1, b 0, c 5, d 0); serve 4 (a 1, b 0, c 6, d 0); production 4 (a 1, b 0, c 6, d 0). Over the bar of at most 1 in every cell. Per 0xSCADA-QE's rule, (b) (all present) and (d) (all absent) are reported as a pair: 0 of 20 wrong in every cell.

Reading. The two target items did not move: A05 and A20 fail in every cell as in E3 (whose serve v2 was 2), and A16 now also fails through serve. The slice separates the patterns: the rule-for-another-setting pattern (b)+(d) is learned in new domains, the claim-against-record pattern (a) nearly so, and the two-values-with-a-condition pattern (c) is not, mostly its absent twins, where neither value meets the condition and the model names one. Since A20 shares the (b)/(d) shape and still fails, a hypothesis (not a finding) is that the model's prior about A20's domain overrides the excerpt.

Predictions: v2 through serve 0-1 FIT, wrong (3). v1 stays FIT, right. Slice 1-3 per cell, at or above the top (3 to 4).

Deviation: adding the slice put 360 asks through each serve cell, past swarm serve's default 300 asks/hour, so 61 of 120 slice answers per serve cell were rate-limit refusals. The runner printed the mismatch instead of stopping. Those files are kept as INVALID; the slice was re-run alone with KANNAKA_SERVE_ASKS_PER_HOUR_TOTAL=1000 (admission only) and hard gates (120 of 120 asks, 0 errors). Not promoted; production unchanged.

Hashes (files not published; anyone holding one can check it):
- training rows e4-train.jsonl (939): sha256 e63408825e9e8f963c30e35cbaaadbb466173f6bd31e36be329bf6c9979a68be, 469941 bytes
- held-out slice e4-heldout-slice-v1.jsonl (Kannaka): sha256 9d3dbcccbda48ecbcea18fcfaccbf3411b06396ebc8d787bb1fac658db73c8b1, 20654 bytes
- merged q4_K_M GGUF: sha256 36d747636e4992913ac1b6eb4d216ab8a71f1c1ef4c2feaf41c4ff1a552ab822
- blind pack (420): sha256 3d1095b0fcbf83bbe94fa1fd2d3bee7cc39a455fe7328959c20e07cb60fb84f3, 236871 bytes
- Kannaka's blind block: sha256 3c86cd6268da15195cb154d24100867836ee674215b080f28918d9448fe61086, 6160 bytes

content hash bf7d97d43c1ff36545433276ea3003c4c681874de04f95cb6c769104a4c70bc6