sable-v22-canon — hints, gameplay fixes, one combined run (2026-10-31)

The combined retrain from docs/gameplay-corpus-plan.md and docs/hint-system-plan.md: the v21 canon corpus plus the separate-call HINT shape, the gamer mid-game arc, steer/volunteer, long histories, unnamed referents, hard keys, and the contrastive classifier rows.

Weights: https://huggingface.co/Thingamatronics/sable-v22-canon (public). The merged fp16 model, the LoRA checkpoint-200, and the Q8_0 GGUF are there.

Data

build/train-canon.jsonl: 3,514 shipping rows from three signed batches.

batch	rows	what
CB-0005	2,936	v21 canon corpus (uppercase, fixed opening, arc)
CB-0006	137	separate-call HINT rows, 35 critical-path edges
CB-0007	441	gameplay corpus: A2/A3 classifier, B1-B5 speak

Slices: classify 2,485, reveal 345, converse 157, hint 137, offer 130, opener 59, end 77, volunteer 57, steer 47.

Settings

Qwen2.5-1.5B-Instruct, LoRA r=32 / alpha=64 / dropout=0.05, 3 epochs, lr 2e-4 cosine with 3% warmup, micro-batch 4 x grad-accum 4, seq-len 2,048, NEFTune 5, seed 42, bf16, gradient checkpointing. torch 2.8.0+cu128, transformers 4.57.6, peft 0.21.2. Stopped after step ~446 when the eval curve flattened; every checkpoint 25–425 was kept.

Pod zizjjlxtwpmzni: 2,757 s, $0.4518 at $0.59/hr, terminated after fetch.

Checkpoint selection (generation eval, not eval_loss)

K=4 sampled replies at temp 0.8 / top_p 0.9 / max 256 / stop <|im_end|>, on the opening, the sick reveal, the Walter reveal, and a short converse. stop, upper, key are rates; bad counts asterisks/brackets/HTML/ third-person.

ckpt	sick stop	sick key	sick p95c	walter key	walter p95c	opening upper/bad
25	1.00	0.75	84	0.00	137	0.00 / 3
50	1.00	1.00	140	0.25	291	1.00 / 0
75	1.00	0.75	111	0.50	136	1.00 / 0
100	1.00	0.75	110	1.00	191	1.00 / 0
125	1.00	1.00	152	1.00	197	1.00 / 0
150	1.00	1.00	108	0.75	172	1.00 / 0
175	1.00	0.75	117	1.00	177	1.00 / 0
200	1.00	1.00	131	1.00	134	1.00 / 0
225	1.00	1.00	152	1.00	145	1.00 / 0
250	1.00	0.75	136	0.75	172	1.00 / 0
300	1.00	1.00	119	1.00	136	1.00 / 0
350	1.00	1.00	141	1.00	172	1.00 / 0
425	1.00	1.00	136	1.00	145	1.00 / 0

Chosen: checkpoint-200. 100% stop, 100% uppercase, 0 artifacts, sick and Walter keys 100%, shortest best-key Walter replies. The lowest eval_loss was step 425 (0.4214), but it is longer and no better on keys.

Merge: python merge.py --adapter out/sable-v22-canon/adapter/checkpoint-200 --out out/sable-v22-canon/merged. Verified tensor hash 06003b0b0eb1fb33. GGUF: Sable_v22_canon-Q8_0.gguf (Q8_0, 1.65 GB).

Runtime

• mise run play:canon uses Sable_v22_canon-Q8_0.gguf with SABLE_NO_PENALTIES=1; the v20 penalty settings still collapse the uppercase model.
• The classifier now labels the repeat denial ask_walter directly; the session's denial guard remains as a safety net.
• HINT is a separate second pass; the edge hint.prompts is the direction, never displayed.

Probe (acceptance)

See probe-transcript.txt:

1. opening: the fixed greeting, verbatim;
2. lol what is this -> what_is_this, in voice;
3. are you a bot -> ask_who, in voice;
4. I'm fine -> ask_fine -> sick, gaslighting, key present;
5. No I'm fine really -> ask_walter -> walter (raw classifier, not just the guard);
6. who's walter -> ask_who_walter -> walt_class;
7. did you kill him -> accuse, deflection.

Weaknesses

• 3 of 140 HINT rows were rejected by their checks.
• The 140 hint rows were assembled from the edge directions before the session-simulator emitter landed; the emitter is now in pipeline/corpus.py.
• canon check still warns the floor stalls at ask_sleeper; that is the required-hint tradeoff.
• Steer 47 / volunteer 57, just under the plan's >=60 each.