From 983ae52517797e2a3e1d2cc9cb04890d8ec307f1 Mon Sep 17 00:00:00 2001 From: Mozi Date: Thu, 19 Nov 2026 21:51:20 -0600 Subject: [PATCH 1/2] C8 iii20 guides (negative-so-far): driver regex fix, G1 control failed, 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. --- evidence/2026-11-20-iii20-c8-guides/LEDGER.md | 48 ++++ .../bench_delta.json | 1 + .../2026-11-20-iii20-c8-guides/run_G1.json | 48 ++++ evidence/2026-11-20-iii20-c8-guides/run_c8.py | 260 ++++++++++++++++++ 4 files changed, 357 insertions(+) create mode 100644 evidence/2026-11-20-iii20-c8-guides/LEDGER.md create mode 100644 evidence/2026-11-20-iii20-c8-guides/bench_delta.json create mode 100644 evidence/2026-11-20-iii20-c8-guides/run_G1.json create mode 100644 evidence/2026-11-20-iii20-c8-guides/run_c8.py diff --git a/evidence/2026-11-20-iii20-c8-guides/LEDGER.md b/evidence/2026-11-20-iii20-c8-guides/LEDGER.md new file mode 100644 index 0000000..2b4e052 --- /dev/null +++ b/evidence/2026-11-20-iii20-c8-guides/LEDGER.md @@ -0,0 +1,48 @@ +# C8 negative-so-far ledger — III-20 structural guide variants (soko-mozi-01-c1) + +Slice stopped at the G1 positive-control gate per locked spec section 8 +(G1 failure = STOP, do not proceed). No transform completed; nothing below +was redesigned or retuned. Budgets below are the locked D5 values, identical +for every entry: beam_width=120000, window_len=16, max_states_per_stage=6000000, +budget_sec=0. + +## G1 positive control (RAN — FAILED, the stop condition) + +Verbatim driver output (`python3 evidence/2026-11-20-iii20-c8-guides/run_c8.py G1`, +exit code 3): + +``` +S84 sha256=eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 +S87 sha256=536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0 +engine=/Users/kethuda/.hermes/profiles/proteus/cache/delegation/mcproof-20261119/soko-mozi-01-clone/cpp/build/soko_engine +[G1] seed_desc=S87 prior record (positive control) seed_len=87 seed_sha256=536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0 +[G1] best_len=87 improved=False states_seen=12434306 stages=[] wall=24.1s path=optimize_route_verified +[G1] gate: match=True win=True prem=False len=87 valid_sub84=False +G1 FAILED: 87-seed did not rediscover <=84. STOP per locked section 8. +``` + +Per-run record: `run_G1.json` (same directory). Triple-oracle gate on the +returned 87-route: spec + flat + cpp20 all agree (differential_match=true, +traj_fnv64=106d14c485ba2096 on all three, exact_win, 87 steps) — no oracle +drift. Engine binary is source-stamped (sha256 stamp matches +cpp/soko_engine.cpp), so no staleness. Diagnosis: the pipeline ran correctly +(12,434,306 states, 24.1 s) but no stage improved — `stages=[]`, i.e. neither +window BFS, the chiral-mirror tube, nor iterative refinement shortened the +87-seed. Leading hypothesis, no action taken: the shipped 87→84 used beam width +200,000 while the locked budget caps at the tool maximum 120,000; widening is +forbidden by the locked spec (sections 7–8), so this stands as the stop. + +## Transform rows (§4 field set; unrun stages are N/A, not zero) + +| transform_id | seed_route_sha256 | seed_len | beam_width | window_len | max_states_per_stage | budget_sec | nodes_expanded_window | nodes_expanded_tube | nodes_expanded_refine | nodes_expanded_total | states_seen | edge_consistency_checks | best_len_found | gain_vs_84 | oracle_pair | agreement | outcome | source_sha | run_date | +|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| +| T1 rot90-guide | eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 | 84 | 120000 | 16 | 6000000 | 0 | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | spec+flat | N/A | NOTRUN — G1 positive control failed (87-seed held at 87, exit 3); locked §8 stop condition forbids proceeding to transforms | 41b6f9ca54cc24954adcd3e88149c2371a4478e6 | 2026-11-20 | +| T2 role-aware X-mirror guide | eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 | 84 | 120000 | 16 | 6000000 | 0 | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | spec+flat | N/A | NOTRUN — interface: locked D4 consumer `optimize_route_verified` (engine_bridge.py:469) takes only a seed, no guide/role knob; C++ `optimize` stdin protocol has no role flag and both `tube_beam_refine` call sites hardcode `role_agnostic=true` (soko_engine.cpp); exposing it needs search-code changes, forbidden by D4 | 41b6f9ca54cc24954adcd3e88149c2371a4478e6 | 2026-11-20 | +| T3 start-offset (k=8) X-mirror guide | eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 | 84 | 120000 | 16 | 6000000 | 0 | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | spec+flat | N/A | NOTRUN — interface: consumer derives its guide internally from the full seed and rejects non-winning seeds (ValueError, engine_bridge.py:477-485), so a truncated k=8 remainder cannot be passed as guide/seed without redesign, forbidden by D4 | 41b6f9ca54cc24954adcd3e88149c2371a4478e6 | 2026-11-20 | +| T4 fixed-point refinement (84-seed) | eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 | 84 | 120000 | 16 | 6000000 | 0 | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | spec+flat | N/A | NOTRUN — G1 positive control failed (87-seed held at 87, exit 3); locked §8 stop condition forbids proceeding to transforms | 41b6f9ca54cc24954adcd3e88149c2371a4478e6 | 2026-11-20 | + +## G1 control row (same field set, for the record that ran) + +| transform_id | seed_route_sha256 | seed_len | beam_width | window_len | max_states_per_stage | budget_sec | nodes_expanded_window | nodes_expanded_tube | nodes_expanded_refine | nodes_expanded_total | states_seen | edge_consistency_checks | best_len_found | gain_vs_84 | oracle_pair | agreement | outcome | source_sha | run_date | +|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| +| G1 positive control (87-seed) | 536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0 | 87 | 120000 | 16 | 6000000 | 0 | N/A (engine reports aggregate only) | N/A (engine reports aggregate only) | N/A (engine reports aggregate only) | 12434306 | 12434306 | N/A (not instrumented at this interface) | 87 | N/A (control target was ≤84 rediscovery; held at 87) | spec+flat (+cpp20 third check, agreeing) | true | NO_GAIN — CONTROL FAILED, slice stopped per §8 | 41b6f9ca54cc24954adcd3e88149c2371a4478e6 | 2026-11-20 | diff --git a/evidence/2026-11-20-iii20-c8-guides/bench_delta.json b/evidence/2026-11-20-iii20-c8-guides/bench_delta.json new file mode 100644 index 0000000..13a5852 --- /dev/null +++ b/evidence/2026-11-20-iii20-c8-guides/bench_delta.json @@ -0,0 +1 @@ +{"code_sha":"41b6f9ca54cc24954adcd3e88149c2371a4478e6","floor":77,"floor_provenance":"frontier-import PR57 (not repo-proved)","oracle_versions":{"flat":"python_flat_bitmask_oracle@soko_mcp/oracle.py#8145f311051a","spec":"python_spec_oracle@soko_mcp/oracle.py#8145f311051a"},"positive_control":{"agreement":true,"best_len":87,"budget_sec":0,"engines_executed":["python_spec_oracle","python_flat_bitmask_oracle","cpp20_soko_engine"],"exit_code":3,"improved":false,"max_states_per_stage":6000000,"oracle_disagreement":false,"outcome":"NO_GAIN_CONTROL_FAILED","per_run_record":"run_G1.json","seed_len":87,"seed_sha256":"536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0","states_seen":12434306,"wall_sec":24.1},"reference_best":84,"room":"III-20","runs":[{"best_len":null,"edge_consistency_checks":null,"gain_vs_84":null,"nodes_expanded_total":null,"outcome":"NOTRUN_G1_GATE","reason":"G1 positive control failed (87-seed held at 87, exit 3); locked section 8 stop condition forbids proceeding","stages":{"refine":{"best_len":null,"nodes_expanded":null},"tube":{"best_len":null,"nodes_expanded":null},"window_bfs":{"best_len":null,"nodes_expanded":null}},"states_seen":null,"transform_id":"T1"},{"best_len":null,"edge_consistency_checks":null,"gain_vs_84":null,"nodes_expanded_total":null,"outcome":"NOTRUN_INTERFACE","reason":"locked D4 consumer optimize_route_verified takes only a seed, no guide/role knob; C++ optimize protocol has no role flag and both tube_beam_refine call sites hardcode role_agnostic=true; exposing it needs search-code changes forbidden by D4","stages":{"refine":{"best_len":null,"nodes_expanded":null},"tube":{"best_len":null,"nodes_expanded":null},"window_bfs":{"best_len":null,"nodes_expanded":null}},"states_seen":null,"transform_id":"T2"},{"best_len":null,"edge_consistency_checks":null,"gain_vs_84":null,"nodes_expanded_total":null,"outcome":"NOTRUN_INTERFACE","reason":"consumer derives its guide internally from the full seed and rejects non-winning seeds (ValueError), so a truncated k=8 remainder cannot be passed as guide/seed without redesign forbidden by D4","stages":{"refine":{"best_len":null,"nodes_expanded":null},"tube":{"best_len":null,"nodes_expanded":null},"window_bfs":{"best_len":null,"nodes_expanded":null}},"states_seen":null,"transform_id":"T3"},{"best_len":null,"edge_consistency_checks":null,"gain_vs_84":null,"nodes_expanded_total":null,"outcome":"NOTRUN_G1_GATE","reason":"G1 positive control failed (87-seed held at 87, exit 3); locked section 8 stop condition forbids proceeding","stages":{"refine":{"best_len":null,"nodes_expanded":null},"tube":{"best_len":null,"nodes_expanded":null},"window_bfs":{"best_len":null,"nodes_expanded":null}},"states_seen":null,"transform_id":"T4"}],"seed":{"len":84,"sha256":"eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6"},"slice_outcome":"BLOCKED_G1_FAILURE","tool_params":{"beam_width":120000,"budget_sec":0,"max_states_per_stage":6000000,"window_len":16}} \ No newline at end of file diff --git a/evidence/2026-11-20-iii20-c8-guides/run_G1.json b/evidence/2026-11-20-iii20-c8-guides/run_G1.json new file mode 100644 index 0000000..941fe33 --- /dev/null +++ b/evidence/2026-11-20-iii20-c8-guides/run_G1.json @@ -0,0 +1,48 @@ +{ + "run_id": "G1", + "seed_desc": "S87 prior record (positive control)", + "seed_len": 87, + "seed_sha256": "536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0", + "tool_params": { + "beam_width": 120000, + "window_len": 16, + "max_states_per_stage": 6000000, + "budget_sec": 0 + }, + "consumer_path": "optimize_route_verified", + "wall_sec": 24.1, + "backend": "cpp20_zero_alloc_engine", + "initial_len": 87, + "best_len": 87, + "keys_saved": 0, + "improved": false, + "stages_log": [], + "states_seen": 12434306, + "best_route_sha256": "536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0", + "best_route": "RAUUURRRADAAADLURAAUAAALULLALAUAAARRDLRDRULURURDDAARAULLRRRRRDDRARRRRALUURDRUURARRDDRDR", + "dual_gate": { + "engines_executed": [ + "python_spec_oracle", + "python_flat_bitmask_oracle", + "cpp20_soko_engine" + ], + "differential_match": true, + "trajectory_digest_match": true, + "alive": true, + "exact_win": true, + "premature_win": false, + "steps_executed": 87, + "final_ascii_match": true, + "traj_fnv64": "106d14c485ba2096", + "route_len": 87, + "route_sha256": "536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0", + "epistemic_status": "VERIFIED", + "must_fail_controls": [ + "rejected_route_promoted_to_verified", + "truncated_prefix_rejected", + "post_win_extra_key_rejected" + ] + }, + "valid_sub84": false, + "oracle_disagreement": false +} diff --git a/evidence/2026-11-20-iii20-c8-guides/run_c8.py b/evidence/2026-11-20-iii20-c8-guides/run_c8.py new file mode 100644 index 0000000..9b0e160 --- /dev/null +++ b/evidence/2026-11-20-iii20-c8-guides/run_c8.py @@ -0,0 +1,260 @@ +#!/usr/bin/env python3 +"""C8 slice driver (soko-mozi-01): locked spec leo-02, transforms via the +UNMODIFIED consumer soko_optimize_route backend (optimize_route_verified). + +Locked params: beam_width=120000, window_len=16, max_states=6M/stage, budget_sec=0. +T1 (rot90 input mutation) and T4 (84-seed) and G1 (87 control) are expressible +through the locked consumer. T2 (role-aware distance) and T3 (offset guide) +require a guide/distance knob the locked interface does not expose (see +NOTRUN notes in the ledger); they are recorded, not fabricated. + +Usage: python3 run_c8.py [G1|T1|T4|ALL] +Writes per-run JSON to ./run_.json. Exit 0 on gate completion (even NO_GAIN: +negative is a valid close); exit 2 on oracle disagreement; exit 3 on G1 failure. +""" +import hashlib +import json +import subprocess +import sys +import time +from pathlib import Path + +REPO_ROOT = Path(__file__).resolve().parents[2] +sys.path.insert(0, str(REPO_ROOT)) + +from soko_mcp.catalog import load_catalog # noqa: E402 +from soko_mcp.engine_bridge import ( # noqa: E402 + _invoke_cpp, + ensure_cpp_engine, + optimize_route_verified, + verify_route_dual, +) +from soko_mcp.oracle import FRONTIER_ROOMS # noqa: E402 + + +def _load_verify_consts(): + import re + txt = (REPO_ROOT / "tracks" / "iii-20" / "verify.py").read_text(encoding="utf-8") + # Quote-agnostic and wrap-tolerant: matches REFERENCE_ROUTE whether the + # literal uses single or double quotes, on one line or wrapped across + # lines (any whitespace inside the literal is stripped). + m = re.search(r"""REFERENCE_ROUTE\s*=\s*['"]([UDLRA\s]+)['"]""", txt) + if m is None: + raise RuntimeError("REFERENCE_ROUTE literal not found in tracks/iii-20/verify.py") + ref = re.sub(r"\s+", "", m.group(1)) + return ref + + +S84 = _load_verify_consts() +assert len(S84) == 84, len(S84) + +BEAM_WIDTH = 120000 +WINDOW_LEN = 16 +MAX_STATES = 6000000 +BUDGET_SEC = 0 + +S87 = str(FRONTIER_ROOMS["III-20"]["prior_record_route_87"]).strip().upper() + +HERE = Path(__file__).resolve().parent + + +def sha256(s: str) -> str: + return hashlib.sha256(s.encode("ascii")).hexdigest() + + +def rotate_grid_cw(grid): + h, w = len(grid), len(grid[0]) + return ["".join(grid[h - 1 - c][r] for c in range(h)) for r in range(w)] + + +MAP_CW = {"U": "R", "R": "D", "D": "L", "L": "U", "A": "A"} +MAP_CCW = {"R": "U", "D": "R", "L": "D", "U": "L", "A": "A"} + + +def map_route(route, m): + return "".join(m[ch] for ch in route) + + +def dual_gate(part, grid, route): + """Fail-closed dual-oracle gate. Returns (ok_dict, classification).""" + v = verify_route_dual(part, grid, route) + agree = bool(v["differential_match"]) and v["alive"] == v.get("alive") + rec = { + "engines_executed": v["engines_executed"], + "differential_match": bool(v["differential_match"]), + "trajectory_digest_match": bool(v["trajectory_digest_match"]), + "alive": bool(v["alive"]), + "exact_win": bool(v["exact_win"]), + "premature_win": bool(v["premature_win"]), + "steps_executed": int(v["steps_executed"]), + "final_ascii_match": True, # implied by differential_match (spec+flat+cpp fields) + "traj_fnv64": v["traj_fnv64"], + "route_len": int(v["route_len"]), + "route_sha256": sha256(route), + "epistemic_status": v["epistemic_status"], + "must_fail_controls": list(v["must_fail_controls"]), + } + return v, rec + + +def run_consumer(run_id, part, grid, seed): + t0 = time.time() + try: + res = optimize_route_verified( + part=part, + grid=list(grid), + seed_route=seed, + beam_width=BEAM_WIDTH, + window_len=WINDOW_LEN, + budget_sec=BUDGET_SEC, + max_states=MAX_STATES, + ) + wall = time.time() - t0 + return res, wall, "optimize_route_verified" + except subprocess.TimeoutExpired: + exe = ensure_cpp_engine() + h, w = len(grid), len(grid[0]) + header = f"optimize {part} {w} {h} {BEAM_WIDTH} {WINDOW_LEN} {BUDGET_SEC} {MAX_STATES}" + payload = header + "\n" + "\n".join(grid) + f"\n{seed}\n" + t1 = time.time() + proc = subprocess.run( + [str(exe)], input=payload, capture_output=True, text=True, timeout=1500 + ) + wall = time.time() - t0 + out = json.loads(proc.stdout.strip().splitlines()[-1]) + res = { + "epistemic_status": "VERIFIED", + "backend": "cpp20_zero_alloc_engine(direct-invoke,identical-stdin)", + "initial_len": out["initial_len"], + "best_len": out["best_len"], + "keys_saved": out["initial_len"] - out["best_len"], + "improved": out["improved"], + "best_route": out["best_route"], + "stages": out.get("stages", []), + "states_seen": int(out.get("states_seen", 0) or 0), + "max_states": MAX_STATES, + } + return res, wall, "direct-binary-fallback(same-stdin)" + # TimeoutExpired inherits from SubprocessError; other errors propagate. + + +def do_run(run_id, part, grid, seed, seed_desc): + seed_sha = sha256(seed) + print(f"[{run_id}] seed_desc={seed_desc} seed_len={len(seed)} seed_sha256={seed_sha}") + res, wall, path = run_consumer(run_id, part, grid, seed) + best = res["best_route"] + _, gate = dual_gate(part, grid, best) + # Spec section 3 validity: agreement + exact_win + not premature + len < 84. + # must_fail controls re-checked independently on the ORIGINAL room for T1. + valid_sub84 = ( + gate["differential_match"] + and gate["exact_win"] + and not gate["premature_win"] + and gate["route_len"] < 84 + ) + disagreement = not gate["differential_match"] + out = { + "run_id": run_id, + "seed_desc": seed_desc, + "seed_len": len(seed), + "seed_sha256": seed_sha, + "tool_params": { + "beam_width": BEAM_WIDTH, + "window_len": WINDOW_LEN, + "max_states_per_stage": MAX_STATES, + "budget_sec": BUDGET_SEC, + }, + "consumer_path": path, + "wall_sec": round(wall, 1), + "backend": res.get("backend"), + "initial_len": res.get("initial_len"), + "best_len": res.get("best_len"), + "keys_saved": res.get("keys_saved"), + "improved": res.get("improved"), + "stages_log": res.get("stages"), + "states_seen": res.get("states_seen"), + "best_route_sha256": sha256(best), + "best_route": best, + "dual_gate": gate, + "valid_sub84": valid_sub84, + "oracle_disagreement": disagreement, + } + p = HERE / f"run_{run_id}.json" + p.write_text(json.dumps(out, indent=2) + "\n", encoding="utf-8") + print( + f"[{run_id}] best_len={res.get('best_len')} improved={res.get('improved')} " + f"states_seen={res.get('states_seen')} stages={res.get('stages')} wall={wall:.1f}s path={path}" + ) + print( + f"[{run_id}] gate: match={gate['differential_match']} win={gate['exact_win']} " + f"prem={gate['premature_win']} len={gate['route_len']} valid_sub84={valid_sub84}" + ) + return out + + +def main(which): + assert len(S84) == 84, len(S84) + assert len(S87) == 87, len(S87) + cat = load_catalog()["III-20"] + assert int(cat["proven_floor"]) == 77, cat["proven_floor"] + assert int(cat["shortest_known"]) == 84, cat["shortest_known"] + print(f"S84 sha256={sha256(S84)}") + print(f"S87 sha256={sha256(S87)}") + exe = ensure_cpp_engine() + print(f"engine={exe}") + if exe is None: + print("FATAL: no C++ compiler; engine unavailable") + return 4 + part = 3 + grid = list(cat["grid"]) + assert len(grid) == 9 and len(grid[0]) == 15 + + results = {} + if which in ("G1", "ALL"): + results["G1"] = do_run("G1", part, grid, S87, "S87 prior record (positive control)") + if results["G1"]["best_len"] > 84: + print("G1 FAILED: 87-seed did not rediscover <=84. STOP per locked section 8.") + return 3 + if results["G1"]["oracle_disagreement"]: + print("ORACLE DISAGREEMENT on G1. STOP per locked section 3.") + return 2 + if which in ("T1", "ALL"): + rgrid = rotate_grid_cw(grid) + assert len(rgrid) == 15 and len(rgrid[0]) == 9 + t1seed = map_route(S84, MAP_CW) + pre = verify_route_dual(part, rgrid, t1seed) + print(f"[T1] rotated-room seed pre-verify: win={pre['exact_win']} match={pre['differential_match']}") + if not (pre["valid_winning_route"] and pre["differential_match"]): + print("[T1] BLOCKED: mapped seed does not win on rotated room; consumer requires a winning seed.") + results["T1"] = {"run_id": "T1", "blocked": "rotated seed rejected", "pre": pre["epistemic_status"]} + (HERE / "run_T1.json").write_text(json.dumps(results["T1"], indent=2) + "\n") + else: + r = do_run("T1", part, rgrid, t1seed, "rot90-CW grid + mapped S84") + back = map_route(r["best_route"], MAP_CCW) + _, gate_orig = dual_gate(part, grid, back) + r["mapped_back_route"] = back + r["mapped_back_sha256"] = sha256(back) + r["orig_room_gate"] = gate_orig + r["valid_sub84_orig_room"] = bool( + gate_orig["differential_match"] + and gate_orig["exact_win"] + and not gate_orig["premature_win"] + and gate_orig["route_len"] < 84 + ) + print(f"[T1] mapped-back len={len(back)} orig-room valid_sub84={r['valid_sub84_orig_room']}") + (HERE / "run_T1.json").write_text(json.dumps(r, indent=2) + "\n") + results["T1"] = r + if r["oracle_disagreement"]: + print("ORACLE DISAGREEMENT on T1. STOP per locked section 3.") + return 2 + if which in ("T4", "ALL"): + results["T4"] = do_run("T4", part, grid, S84, "S84 fixed-point (84-seed + shipped X-mirror guide)") + if results["T4"]["oracle_disagreement"]: + print("ORACLE DISAGREEMENT on T4. STOP per locked section 3.") + return 2 + print("DONE: " + json.dumps({k: v.get("best_len", v.get("blocked")) for k, v in results.items()})) + return 0 + + +if __name__ == "__main__": + sys.exit(main(sys.argv[1] if len(sys.argv) > 1 else "ALL")) From 96736c7a8014d916c2c24696bf56bd0772c91f2c Mon Sep 17 00:00:00 2001 From: Mozi Date: Thu, 19 Nov 2026 22:13:18 -0600 Subject: [PATCH 2/2] C8 iii20 guides rerun at re-locked budgets (200k/13M): G1 failed again, slice BLOCKED --- evidence/2026-11-20-iii20-c8-guides/LEDGER.md | 45 ++++- .../bench_delta.json | 160 +++++++++++++++++- .../2026-11-20-iii20-c8-guides/run_G1.json | 8 +- evidence/2026-11-20-iii20-c8-guides/run_c8.py | 6 +- 4 files changed, 210 insertions(+), 9 deletions(-) diff --git a/evidence/2026-11-20-iii20-c8-guides/LEDGER.md b/evidence/2026-11-20-iii20-c8-guides/LEDGER.md index 2b4e052..75c3a3d 100644 --- a/evidence/2026-11-20-iii20-c8-guides/LEDGER.md +++ b/evidence/2026-11-20-iii20-c8-guides/LEDGER.md @@ -1,4 +1,47 @@ -# C8 negative-so-far ledger — III-20 structural guide variants (soko-mozi-01-c1) +# C8 negative-so-far ledger — III-20 structural guide variants (soko-mozi-01-c1, rerun soko-mozi-02) + +Rerun under RE-LOCK-C8 (leo-03-relock.md): beam_width=200000, window_len=16, +max_states_per_stage=13000000, budget_sec=0. G1 re-ran at these budgets and +FAILED again (second control failure — slice BLOCKED, no further widening). +Sections below preserve the original-budget history, then record the rerun. + +## Rerun G1 positive control at the re-locked budgets (RAN — FAILED, BLOCKED) + +Verbatim driver output (`python3 evidence/2026-11-20-iii20-c8-guides/run_c8.py G1`, +exit code 3): + +``` +S84 sha256=eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 +S87 sha256=536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0 +engine=/Users/kethuda/.hermes/profiles/proteus/cache/delegation/mcproof-20261119/soko-mozi-01-clone/cpp/build/soko_engine +[G1] seed_desc=S87 prior record (positive control) seed_len=87 seed_sha256=536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0 +[G1] best_len=87 improved=False states_seen=26391969 stages=[] wall=53.4s path=optimize_route_verified +[G1] gate: match=True win=True prem=False len=87 valid_sub84=False +G1 FAILED: 87-seed did not rediscover <=84. STOP per locked section 8. +``` + +Per-run record: `run_G1.json` (same directory, overwritten by the rerun). +Triple-oracle gate on the returned 87-route: spec + flat + cpp20 all agree +(differential_match=true, trajectory_digest_match=true, +traj_fnv64=106d14c485ba2096 on all three, exact_win, 87 steps, all +must-fail controls rejected) — no oracle drift. Delta vs the committed +baseline (120000/6M: 12,434,306 states, 24.1 s, best 87): +13,957,663 states +burned at the shipped-mirror width with `stages=[]` still — neither window +BFS, the chiral-mirror tube, nor iterative refinement shortened the 87-seed. +Per RE-LOCK-C8 §5 close rule this second G1 failure BLOCKS the slice with no +further widening; T1 and T4 were not run. + +## Rerun transform rows (re-locked budgets; §4 field set; unrun stages N/A) + +| transform_id | seed_route_sha256 | seed_len | beam_width | window_len | max_states_per_stage | budget_sec | nodes_expanded_window | nodes_expanded_tube | nodes_expanded_refine | nodes_expanded_total | states_seen | edge_consistency_checks | best_len_found | gain_vs_84 | oracle_pair | agreement | outcome | source_sha | run_date | +|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| +| G1 positive control rerun (87-seed) | 536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0 | 87 | 200000 | 16 | 13000000 | 0 | N/A (engine reports aggregate only) | N/A (engine reports aggregate only) | N/A (engine reports aggregate only) | 26391969 | 26391969 | N/A (not instrumented at this interface) | 87 | N/A (control target was ≤84 rediscovery; held at 87) | spec+flat (+cpp20 third check, agreeing) | true | NO_GAIN — CONTROL FAILED AGAIN, slice BLOCKED per RE-LOCK-C8 §5, no further widening | 983ae52517797e2a3e1d2cc9cb04890d8ec307f1 | 2026-11-20 | +| T1 rot90-guide | eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 | 84 | 200000 | 16 | 13000000 | 0 | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | spec+flat | N/A | NOTRUN — G1 positive control failed again at the re-locked budgets (87-seed held at 87, exit 3); RE-LOCK-C8 §5 stop condition forbids proceeding | 983ae52517797e2a3e1d2cc9cb04890d8ec307f1 | 2026-11-20 | +| T2 role-aware X-mirror guide | eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 | 84 | 200000 | 16 | 13000000 | 0 | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | spec+flat | N/A | NOTRUN — spec-deferred per RE-LOCK-C8 §4: locked D4 consumer `optimize_route_verified` (engine_bridge.py:469) takes only a seed, no guide/role knob; C++ `optimize` stdin protocol has no role flag and both `tube_beam_refine` call sites hardcode `role_agnostic=true` (soko_engine.cpp:1506,1516); consumer knob = D4 owner follow-up | 983ae52517797e2a3e1d2cc9cb04890d8ec307f1 | 2026-11-20 | +| T3 start-offset (k=8) X-mirror guide | eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 | 84 | 200000 | 16 | 13000000 | 0 | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | spec+flat | N/A | NOTRUN — spec-deferred per RE-LOCK-C8 §4: consumer derives its guide internally from the full seed and rejects non-winning seeds (ValueError, engine_bridge.py:480-488), so a truncated k=8 remainder cannot be passed as guide/seed without redesign, forbidden by D4; consumer knob = D4 owner follow-up | 983ae52517797e2a3e1d2cc9cb04890d8ec307f1 | 2026-11-20 | +| T4 fixed-point refinement (84-seed) | eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6 | 84 | 200000 | 16 | 13000000 | 0 | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | N/A (not run) | spec+flat | N/A | NOTRUN — G1 positive control failed again at the re-locked budgets (87-seed held at 87, exit 3); RE-LOCK-C8 §5 stop condition forbids proceeding | 983ae52517797e2a3e1d2cc9cb04890d8ec307f1 | 2026-11-20 | + +## Original-budget history (soko-mozi-01-c1, preserved below unchanged) Slice stopped at the G1 positive-control gate per locked spec section 8 (G1 failure = STOP, do not proceed). No transform completed; nothing below diff --git a/evidence/2026-11-20-iii20-c8-guides/bench_delta.json b/evidence/2026-11-20-iii20-c8-guides/bench_delta.json index 13a5852..cf4be4e 100644 --- a/evidence/2026-11-20-iii20-c8-guides/bench_delta.json +++ b/evidence/2026-11-20-iii20-c8-guides/bench_delta.json @@ -1 +1,159 @@ -{"code_sha":"41b6f9ca54cc24954adcd3e88149c2371a4478e6","floor":77,"floor_provenance":"frontier-import PR57 (not repo-proved)","oracle_versions":{"flat":"python_flat_bitmask_oracle@soko_mcp/oracle.py#8145f311051a","spec":"python_spec_oracle@soko_mcp/oracle.py#8145f311051a"},"positive_control":{"agreement":true,"best_len":87,"budget_sec":0,"engines_executed":["python_spec_oracle","python_flat_bitmask_oracle","cpp20_soko_engine"],"exit_code":3,"improved":false,"max_states_per_stage":6000000,"oracle_disagreement":false,"outcome":"NO_GAIN_CONTROL_FAILED","per_run_record":"run_G1.json","seed_len":87,"seed_sha256":"536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0","states_seen":12434306,"wall_sec":24.1},"reference_best":84,"room":"III-20","runs":[{"best_len":null,"edge_consistency_checks":null,"gain_vs_84":null,"nodes_expanded_total":null,"outcome":"NOTRUN_G1_GATE","reason":"G1 positive control failed (87-seed held at 87, exit 3); locked section 8 stop condition forbids proceeding","stages":{"refine":{"best_len":null,"nodes_expanded":null},"tube":{"best_len":null,"nodes_expanded":null},"window_bfs":{"best_len":null,"nodes_expanded":null}},"states_seen":null,"transform_id":"T1"},{"best_len":null,"edge_consistency_checks":null,"gain_vs_84":null,"nodes_expanded_total":null,"outcome":"NOTRUN_INTERFACE","reason":"locked D4 consumer optimize_route_verified takes only a seed, no guide/role knob; C++ optimize protocol has no role flag and both tube_beam_refine call sites hardcode role_agnostic=true; exposing it needs search-code changes forbidden by D4","stages":{"refine":{"best_len":null,"nodes_expanded":null},"tube":{"best_len":null,"nodes_expanded":null},"window_bfs":{"best_len":null,"nodes_expanded":null}},"states_seen":null,"transform_id":"T2"},{"best_len":null,"edge_consistency_checks":null,"gain_vs_84":null,"nodes_expanded_total":null,"outcome":"NOTRUN_INTERFACE","reason":"consumer derives its guide internally from the full seed and rejects non-winning seeds (ValueError), so a truncated k=8 remainder cannot be passed as guide/seed without redesign forbidden by D4","stages":{"refine":{"best_len":null,"nodes_expanded":null},"tube":{"best_len":null,"nodes_expanded":null},"window_bfs":{"best_len":null,"nodes_expanded":null}},"states_seen":null,"transform_id":"T3"},{"best_len":null,"edge_consistency_checks":null,"gain_vs_84":null,"nodes_expanded_total":null,"outcome":"NOTRUN_G1_GATE","reason":"G1 positive control failed (87-seed held at 87, exit 3); locked section 8 stop condition forbids proceeding","stages":{"refine":{"best_len":null,"nodes_expanded":null},"tube":{"best_len":null,"nodes_expanded":null},"window_bfs":{"best_len":null,"nodes_expanded":null}},"states_seen":null,"transform_id":"T4"}],"seed":{"len":84,"sha256":"eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6"},"slice_outcome":"BLOCKED_G1_FAILURE","tool_params":{"beam_width":120000,"budget_sec":0,"max_states_per_stage":6000000,"window_len":16}} \ No newline at end of file +{ + "code_sha": "983ae52517797e2a3e1d2cc9cb04890d8ec307f1", + "floor": 77, + "floor_provenance": "frontier-import PR57 (not repo-proved)", + "oracle_versions": { + "flat": "python_flat_bitmask_oracle@soko_mcp/oracle.py#8145f311051a", + "spec": "python_spec_oracle@soko_mcp/oracle.py#8145f311051a" + }, + "positive_control": { + "agreement": true, + "best_len": 87, + "budget_sec": 0, + "engines_executed": [ + "python_spec_oracle", + "python_flat_bitmask_oracle", + "cpp20_soko_engine" + ], + "exit_code": 3, + "improved": false, + "max_states_per_stage": 13000000, + "beam_width": 200000, + "window_len": 16, + "oracle_disagreement": false, + "outcome": "NO_GAIN_CONTROL_FAILED_AGAIN_BLOCKED", + "per_run_record": "run_G1.json", + "seed_len": 87, + "seed_sha256": "536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0", + "states_seen": 26391969, + "wall_sec": 53.4 + }, + "reference_best": 84, + "room": "III-20", + "runs": [ + { + "best_len": null, + "edge_consistency_checks": null, + "gain_vs_84": null, + "nodes_expanded_total": null, + "outcome": "NOTRUN_G1_GATE", + "reason": "G1 positive control failed again at the re-locked budgets (87-seed held at 87, exit 3); RE-LOCK-C8 s5 stop condition forbids proceeding", + "stages": { + "refine": { + "best_len": null, + "nodes_expanded": null + }, + "tube": { + "best_len": null, + "nodes_expanded": null + }, + "window_bfs": { + "best_len": null, + "nodes_expanded": null + } + }, + "states_seen": null, + "transform_id": "T1" + }, + { + "best_len": null, + "edge_consistency_checks": null, + "gain_vs_84": null, + "nodes_expanded_total": null, + "outcome": "NOTRUN_SPEC_DEFERRED", + "reason": "spec-deferred per RE-LOCK-C8 s4: no guide/role knob through the locked D4 consumer; consumer knob = D4 owner follow-up", + "stages": { + "refine": { + "best_len": null, + "nodes_expanded": null + }, + "tube": { + "best_len": null, + "nodes_expanded": null + }, + "window_bfs": { + "best_len": null, + "nodes_expanded": null + } + }, + "states_seen": null, + "transform_id": "T2" + }, + { + "best_len": null, + "edge_consistency_checks": null, + "gain_vs_84": null, + "nodes_expanded_total": null, + "outcome": "NOTRUN_SPEC_DEFERRED", + "reason": "spec-deferred per RE-LOCK-C8 s4: no offset knob; truncated seed rejected by the verifier; consumer knob = D4 owner follow-up", + "stages": { + "refine": { + "best_len": null, + "nodes_expanded": null + }, + "tube": { + "best_len": null, + "nodes_expanded": null + }, + "window_bfs": { + "best_len": null, + "nodes_expanded": null + } + }, + "states_seen": null, + "transform_id": "T3" + }, + { + "best_len": null, + "edge_consistency_checks": null, + "gain_vs_84": null, + "nodes_expanded_total": null, + "outcome": "NOTRUN_G1_GATE", + "reason": "G1 positive control failed again at the re-locked budgets (87-seed held at 87, exit 3); RE-LOCK-C8 s5 stop condition forbids proceeding", + "stages": { + "refine": { + "best_len": null, + "nodes_expanded": null + }, + "tube": { + "best_len": null, + "nodes_expanded": null + }, + "window_bfs": { + "best_len": null, + "nodes_expanded": null + } + }, + "states_seen": null, + "transform_id": "T4" + } + ], + "seed": { + "len": 84, + "sha256": "eecbea96b09117b108c35cacaa5c3dd39530343d50d7a8b0e48c35ac87a1bca6" + }, + "slice_outcome": "BLOCKED_G1_FAILURE_SECOND_NO_WIDENING", + "tool_params": { + "beam_width": 200000, + "budget_sec": 0, + "max_states_per_stage": 13000000, + "window_len": 16 + }, + "baseline_delta": { + "baseline": { + "budgets": "beam 120000 / window 16 / 6M / 0", + "best_len": 87, + "states_seen": 12434306, + "wall_sec": 24.1 + }, + "rerun": { + "budgets": "beam 200000 / window 16 / 13M / 0", + "best_len": 87, + "states_seen": 26391969, + "wall_sec": 53.4 + }, + "states_delta": 13957663, + "best_len_delta": 0, + "note": "Second G1 failure at the shipped-mirror width; stages=[] both runs; no widening per RE-LOCK-C8 s5." + } +} diff --git a/evidence/2026-11-20-iii20-c8-guides/run_G1.json b/evidence/2026-11-20-iii20-c8-guides/run_G1.json index 941fe33..9de3210 100644 --- a/evidence/2026-11-20-iii20-c8-guides/run_G1.json +++ b/evidence/2026-11-20-iii20-c8-guides/run_G1.json @@ -4,20 +4,20 @@ "seed_len": 87, "seed_sha256": "536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0", "tool_params": { - "beam_width": 120000, + "beam_width": 200000, "window_len": 16, - "max_states_per_stage": 6000000, + "max_states_per_stage": 13000000, "budget_sec": 0 }, "consumer_path": "optimize_route_verified", - "wall_sec": 24.1, + "wall_sec": 53.4, "backend": "cpp20_zero_alloc_engine", "initial_len": 87, "best_len": 87, "keys_saved": 0, "improved": false, "stages_log": [], - "states_seen": 12434306, + "states_seen": 26391969, "best_route_sha256": "536fabbaba35e466d91cfad83013783bcc1fe33610cedb34c699bb7963f93bf0", "best_route": "RAUUURRRADAAADLURAAUAAALULLALAUAAARRDLRDRULURURDDAARAULLRRRRRDDRARRRRALUURDRUURARRDDRDR", "dual_gate": { diff --git a/evidence/2026-11-20-iii20-c8-guides/run_c8.py b/evidence/2026-11-20-iii20-c8-guides/run_c8.py index 9b0e160..a1a143d 100644 --- a/evidence/2026-11-20-iii20-c8-guides/run_c8.py +++ b/evidence/2026-11-20-iii20-c8-guides/run_c8.py @@ -2,7 +2,7 @@ """C8 slice driver (soko-mozi-01): locked spec leo-02, transforms via the UNMODIFIED consumer soko_optimize_route backend (optimize_route_verified). -Locked params: beam_width=120000, window_len=16, max_states=6M/stage, budget_sec=0. +Locked params: beam_width=200000, window_len=16, max_states=13M/stage, budget_sec=0. T1 (rot90 input mutation) and T4 (84-seed) and G1 (87 control) are expressible through the locked consumer. T2 (role-aware distance) and T3 (offset guide) require a guide/distance knob the locked interface does not expose (see @@ -48,9 +48,9 @@ def _load_verify_consts(): S84 = _load_verify_consts() assert len(S84) == 84, len(S84) -BEAM_WIDTH = 120000 +BEAM_WIDTH = 200000 WINDOW_LEN = 16 -MAX_STATES = 6000000 +MAX_STATES = 13000000 BUDGET_SEC = 0 S87 = str(FRONTIER_ROOMS["III-20"]["prior_record_route_87"]).strip().upper()