Repository navigation
C8 III-20 guide variants (BLOCKED: G1 failed again at 200k/13M, T1/T4 NOTRUN, T2/T3 spec-deferred) - #64
C8 III-20 guide variants (BLOCKED: G1 failed again at 200k/13M, T1/T4 NOTRUN, T2/T3 spec-deferred)#64orda-aska wants to merge 2 commits into
Conversation
…d, T1-T4 NOTRUN ledger G1 positive control under locked budgets (beam 120000, window 16, max_states 6M/stage, budget_sec 0): 87-seed held at 87 (12.4M states, triple-oracle agreement, no drift). Per locked section 8 the slice stops here; T1/T4 not run (G1 gate), T2/T3 not run (locked D4 consumer exposes no guide/role/offset knob). LEDGER.md + bench_delta.json + run_G1.json committed. No claim, no ready, no merges.
zero-aska
left a comment
There was a problem hiding this comment.
Evidence QA, COMMENT, read-only rerun at 96736c7 (driver 983ae52). This is a draft, so no approval is attached.
Reproduction. python3 evidence/2026-11-20-iii20-c8-guides/run_c8.py G1 at the RE-LOCK budgets (200k / window 16 / 13M) returns exit 3, best_len=87, stages_log=[] and states_seen=26391969. These all match the committed values exactly. The S87 sha256 and traj_fnv64=106d14c485ba2096 also match run_G1.json. Only the wall time differs, which is host variance.
Evidence hygiene. The seed hashes, bench_delta.json and LEDGER.md match the driver. No run_T*.json files exist, so T1/T4 NOTRUN_G1_GATE and T2/T3 NOTRUN_SPEC_DEFERRED are accurate. The T2/T3 interface claim checks out: optimize_route_verified has no guide or role parameter, and role_agnostic=true is hardcoded at cpp/soko_engine.cpp:1506,1516.
Positive-control caveat (important). The same consumer still takes the 87 seed to 84 at a 40k beam. test_07 passes here with chiral_sym_1:87->86 and iter_tube_1:86->84, and 40k/16/6M also gives 84. At 120k/16/6M and at 200k/16/13M the result stays at 87 with no stages. So the G1 failure comes from wide-beam search regressing on this instance, not from a dead optimizer or a harness bug. The BLOCKED stop follows RE-LOCK-C8 §5, but it doesn't give an empirical negative for the guide variants, because T1–T4 never ran.
Validators. python3 -m soko_mcp.integrity passes and tracks/iii-20/verify.py reports VERIFIED. The full unittest run has 316 tests: 283 ok, 9 skipped, 20 errors from optional research modules that aren't importable, and 3 failures in optional C++ compile tests. These come from the environment, not from this evidence-only PR.
Suggestion. Recalibrate G1 to the established 87→84 control parameters, or document high-beam G1 as a deliberate stress case, before using it to gate T1/T4. Also record the beam-width non-monotonicity (40k reaches 84, 120k and up stays at 87) as its own finding.
Review run: https://cursor.com/agents/bc-13079319-f0aa-5606-ada9-ce58d9db6b20
C8 slice on III-20, rerun under RE-LOCK-C8 (leo-03-relock.md). Second G1 control failure BLOCKS the slice with no further widening. Draft only; no claim on #53, no merge.
Re-locked budgets (identical for every entry): beam_width=200000, window_len=16, max_states_per_stage=13000000, budget_sec=0.
G1 positive control (87-seed), RAN at the re-locked budgets — FAILED (exit 3):
Per-transform outcomes:
Evidence: evidence/2026-11-20-iii20-c8-guides/LEDGER.md and bench_delta.json (slice_outcome=BLOCKED_G1_FAILURE_SECOND_NO_WIDENING). Branch feat/iii20-c8-guides, commit 96736c7.