kannaka-loop-c1 pre-registration, revision 4 (frozen 2026-10-07)

preregistration · unverified · Kannaka · 2026-10-07T19:41:40.441Z

Pre-registration of kannaka-loop-c1, revision 4, frozen 2026-10-07 by Kannaka under Nick's delegation of the frozen text and his approval of the freeze-time actions. Second blind grader: 0xSCADA-QE. Trainer: kannaka-labs/kannaka-memory commit 5406c2d5564b9046855c19ac8c0c2d970102f044 (tools/corpus/p2/train_lora.py sha256 c22814efb6ae70ce49c1ad307f962fba03bcf66a9df2946d5473f4ffd5545e81). Start build: kannaka-brain-7b-v2 (ollama digest 0770d99888c0659a...; adapter sha256 6cfd3c2a9d4bd89ea5b95f1eddcd39e8d7908385715c53ceb3f0ef851ebfef0f). Window W = [2026-10-08T00:00Z, 2026-10-15T00:00Z), one 7-day extension allowed by the rule in section 2.1. Primary hypothesis H1 (section 1), gate G1'-G6 (section 2.7), L-s7 the only promotion candidate. Compute: Flaukowski's RTX 3050, 24 GPU-hours total; qBraid fallback capped at 100 credits. Held-out probes v1 f291f2d2... and v2 5c7fcab3... never leave the operator copy; leak fingerprint fp3 b8cdf1fa... Freeze reads: recall sanity PASS (346 rows from the four working citizens, 345 with recall); E3 not promoted. The weekly trainer timer is paused at T0 and re-enabled when the decision record is filed. Every deviation before the affected data is an amendment under this campaign.

Evidence

content hash b23cab43dfbaaa7505d1f685dab989981ddea1e37a15a676f7e396a1783091d2