Repository navigation
Closed
Conversation
Freezes spec §2–§5, the analysis code (stats.rs), the analysis plan and prompt v1 under prereg_sha ef51087dc79b… before any measurement (S-1). New crate aprender-review-experiment (publish = false); harness CLI is an example, not a [[bin]]. Planted §3 edit turns prereg-check and FALSIFY-REX-PREREG-001 RED; restore → GREEN. Receipt: docs/audits/rex-001/rex-00-receipt.md. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…#4356 Phase 0: no declared label admits apr-serve review jobs on any host; cells stay NotRun{NoDeclaredExecutor} until infra#1088 lands. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ation check — Refs #4357 corpus_version review-corpus-v1@787d2026256cc08b. P 50 (cargo-mutants listing, markers stripped), R 50 (revert-the-fix, comment-only hunks dropped), G 50 (green, unreverted, pre-cutoff). 45 dev / 105 test; 70 test defect items. Diffs are outside git (items-v1.tar sha256 70dffe56…). Contracts review-corpus-v1 (FALSIFY-RCV-001..003) and review-corpus-contamination-v1 (FALSIFY-RCC-001..003); planted train leak RED (rc=101) -> GREEN. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…FY-RXR-001..003) — Refs #4358 Harness drives apr serve /v1/chat/completions at greedy/seed 4354/512 tokens, never truncates (context refusal -> NotRun{ContextOverflow}), seeded run order + 10% rerun. Scorer denominators are the EXPECTED item set; raw outputs are sha-checked and the verdict re-parsed. Receipts verify with minisign when --pubkey is given. Three falsifiers planted RED and restored GREEN (docs/audits/rex-001/rex-03-receipt.md). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…03) — 6/6 NotRun{NoDeclaredExecutor}, S-7 — Refs #4359
Every §2.1 cell has exactly one admission row; a silent cell rejects the file, Admitted
needs a parity receipt at or above a cited threshold, and no admitted cell raises S-7
(`rex admission-check` exits 10). With infra#1088 open, all six cells are NotRun, so S-7
fires and the experiment STOPs. The §7 premise "apr parity vs llama.cpp" is false at
v0.69.3 (GPU-vs-CPU only); the oracle is an explicit field. The S-6 operator ruling is
recorded in the receipt.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`rex score --pilot --split dev` projects the test-split Wilson half-width on recall at the pilot estimate (design p = 0.5 when the pilot has no defect in its denominator) and grows the corpus above 0.12 or when the width cannot be computed. Runtime counts the 10 % re-run; no warm sample is None, not zero. The spec's ~0.117 is the Wald width; the frozen Wilson width is ~0.114. Contract rex-pilot-projection-v1 (FALSIFY-RXP-001..003), each planted RED at pilot_tests.rs:22/:29/:38 and green after restore. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 4B shadow lane already runs inside the quorum rail (paiml-implement #420, advisory-lane.sh); `rex ledger` reads what it records in each quorum receipt and holds it to REX-07: every quorum carries a shadow row (no advisory_lane, or lane `off`, is missing), a lane that did not answer is Unknown{LaneUnavailable}, and the shadow is uncounted and leaves the width at the counted lanes. On the live PMAT-405 receipt it emits a correct row (gx10-cuda PASS, width 3) and exits 0. Contract review-ledger-v1, FALSIFY-RXL-001..003 each planted RED at ledger_tests.rs:49/:66/:85 and green after restore. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
) `rex ladder REPORT_JSON` reads only a locked, confirmatory rex-001-report-v1 and returns the 4B lane's rung: shadow by default, tripwire on H4 (explicit holds + CI lower bound ≥ 0), vote on H4 ∧ H5, capped by the §4.5 hardware ruling. H5 alone stays shadow (R-11). Contract rex-promotion-ladder-v1; FALSIFY-RXG-001..003 each planted RED at ladder_tests.rs:31/:55/:72 and green after restore. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…g is RED (Refs #4365) review-lane-perf-ratchet-v1 (§5.1). `rex ratchet record` turns a tag's warm, non-rerun receipts on the primary cell into per-item wall times; `rex ratchet check` arms after 3 tags and compares each later tag with the best earlier one by a paired bootstrap of the Harrell–Davis p95 ratio (ALPHA from the prereg). Exit 10 on RED. Finding (FALSIFY-RXR-002): the §4 unpaired CI-overlap rule, and a paired nearest-rank p95, both miss a planted +10% at n=130; the HD paired rule catches it 40/40 with 0/40 false alarms at 3% per-item run noise. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…#4366) - champion.rs: five fail-closed §5.4 gates, 20 evaluations per sealed test version (refusals spend nothing, R-6), pin-bump promotion with rollback sha, two-consecutive-gold-escape demotion. - fewshot.rs: B1b retrieval from non-test gold only (R-2), never the query; guard() scans each fenced block alone. Finding: a gold `- defect at` line after a fence joins the last hunk, so a whole-text scan missed a sealed test diff in a rendered prompt (RXF-003 pins it). - `rex challenge` (exit 0 promoted / 10 rejected / 11 refused); score_cmd uses verify_signed. - contracts/review-champion-challenger-v1.yaml (RXC-001..003, RXF-001..003). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…pt (Refs #4367) - b2.rs: one review-b2-loop-v1 row per B2 tier from the refusal ledger (refusal-receipt-v1): ready only on an acceptance receipt, verb_refused on a ledger row, unledgered otherwise — an absent refusal is not acceptance. teacher_receipt(): sha, count, k, m=1, 0 test hits over the non-test pool; fails closed on an empty sealed index. - `rex b2 status|teacher-receipt` (status exits 11 while any row is NotRun). - contracts/review-b2-loop-v1.yaml (FALSIFY-RXB-001..003). - At HEAD all four rows read unledgered: distill/finetune/merge/quantize have no measured qwen35 refusal row (EPIC #4002 owns the verbs). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… logprob) (#4026) The session engine now ranks the raw logits at every decode step, before any repeat penalty, and records (token_id, logit, logprob) for the top K together with the chosen token. The records reach `apr run --json` as `logprobs: {top_k, prompt_token_ids, steps}` through a thread-local, the same way GENERATION_START already does, because dense_turn drops the Turn. Recording turns off the device-greedy shortcut, since that path never returns the logits to the host. A route that cannot record refuses the flag by name and never reports an empty list: --trace (the legacy instrumented loops), qwen3_moe, and every non-GGUF path. top_k_logprobs breaks ties the way ops::argmax does (lower id first) and never ranks NaN. Tests: a case table, logprob_of agreement and sum-to-one, a positive control (top-1 == chosen at every step), a negative control (moving the forward's argmax moves the record), pre-penalty ranking, the device-argmax bypass, and a refusal case table. Not yet run: the real-model controls on Qwen3.5-0.8B-Q4_K_M (#4004 ids). Parked to backlog per the cop's operator relay (0.71.0 staffing). Pmat-Ticket: PMAT-4026 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t dispatch time only (Refs #4354) Operator ruling 2026-09-25 (cop-inbox/handoff/ruling-r9-b.md, sha b6db6866c604). lambda-cuda may run the apr SHADOW lane only as rung 2: rung 1 (gx10-cuda) returned NotRun and the GPU lock is free at dispatch. No resident serve, no preloaded weights, no REX cells, no teacher/finetune jobs; other R-9 exclusions stand. A lock taken mid-round = NotRun{Busy}, ladder continues, no retry on the same cell. Fallback order gx10-cuda -> lambda-cuda -> intel-wgpu -> mini-metal -> intel-cpu -> NotRun{NoExecutor}. S-4, G7 and the lambda-minutes metric follow R-9. R-9 is §6, outside the §2-§5 prereg span: FALSIFY-REX-PREREG-001..003 pass, the lock is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ha carried; holds() needs identity_gaps == 0 (Refs #4354) - Shadow is now Verdict{verdict,cell,backend} | NotRun(NoExecutor|Busy| Timeout|ContextOverflow|TrainActive) | Refused{cell,removed_by: ladder|gpu-proof|parse}, deny_unknown_fields. Unknown::LaneUnavailable {why: free text} is gone: one spelling with receipt-lint. - The receipt's advisory_lane.row carries it; a missing or unparsable row is a violation, never a silent Unknown. - weights_sha256 is read from the receipt, and Coverage::holds() requires identity_gaps == 0. - FALSIFY-RXL-004 (no weights sha => holds() false) and RXL-005 (free-text reason fails to parse), each killed by its mutant (holds() clause dropped; deny_unknown_fields dropped). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…a row (Refs #4354) Operator ruling PRM-001 v2 items 3, 4 and 6 (cop-inbox/handoff/ruling-prm-v2.md). v1 was locked with 0 data rows, so nothing becomes exploratory (R-1). Locked (§3-§4, stats.rs, analysis plan): - H1 = VERDICT identity across cells (verdict_divergence); bytes descriptive. - H4 = one rule: paired bootstrap of p4B - min(pHaiku, pAgy), min taken inside each resample (bootstrap_vs_min_voter); Holm-adjusted one-sided p decides, CI report-only. ladder.rs reads p_holm (report schema v2). - H5 decided on class-R recall only; per-class P/R recall + precision descriptive (paired_parsed_class; Scored carries its class). - §4.1 admission oracle = the cell's parity.oracle; llama.cpp@d1d3c3396 where declared (#4444), else apr-parity-gpu-cpu + H1 verdict check. - §4.6 train-active falls through the dispatch ladder, never blocks; lane rows use the #436 Verdict|NotRun|Refused spelling. - κ on per-item errors, apr vs each voter (error_kappa, kappa_andon). Unlocked: §0.4/R-11 zero-weight shadow; R-9/S-4/§9.1 lambda accounting (experiment cells, teacher, resident serve: 0 min; shadow rung 2 measured). Falsifiers: min-inside-resample differs from pick-once on tied voters; byte-only diff does not reject H1; RXG-004 CI is report-only; prereg 001-003. 90/90 lib tests; pv validate both contracts; clippy -D warnings clean. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ith the cop's pending-instrument ruling; §9 report IN_PROGRESS Spec copied from the cop handoff (sha e8764b72adbe) and amended with §1.1: T0 found none of the §1 verify commands exist; G1/G7/G10/G11/G13 are [U]-pending-instrument tied to T1/T10/T1/T6/T16; G2/G4 are #436 scope; G12's command replaced (apr has no --model-arch). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…, fails closed Five fleet hosts, 2026-09-25: cleanupPeriodDays 36500 on all five; files older than 30 d lambda-labs 3278, intel 290, gx10 1, mini 0, yoga 0 (matches infra#1135 T0). An unreachable host prints UNREACHABLE and exits 1. Case table: unset/numeric/no-settings key, planted old file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ruling agent-trace item 6. FALSIFY-TPA-001 (planted AWS key quarantined) and TPA-002 (planted test-split diff refused) fail 4/4 against admit() admitting everything. Refs #4354 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…aled-test hit refuses Ruling agent-trace item 6. src/pool.rs: admit() gives each JSONL row one verdict. Admitted | Quarantined (secret hit, held byte-identical, never redacted) | Refused (sealed-test diff or sha, via contamination::Index, checked first) | Unchecked (empty sealed index admits nothing). The scanner has no regex dependency, scans JSON-unescaped strings, and uses whole-word prefixes. Contract training-pool-admission-v1 (TPA-001..004; pv validate: 0 errors). RED->GREEN: 4/4 fail on the stub at b177c3af8, 4/4 pass here; crate lib 87/87; clippy -D warnings clean. Mutants killed: no-unescape, no-refuse, no-quarantine, empty-index-admits, no-boundary. Refs #4354 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… reports nothing PRA-001 T6. Secret scan = union of builtin detectors and the gitleaks v8.30.1 ruleset vendored verbatim as data (MIT, sha256-pinned), interpreted in-process in Rust. 20 run-time-built canaries, per-rule ablation, AWS doc key (gitleaks-allowlisted) caught by builtin only. Contract renamed from training-pool-admission-v1. This commit's scan() clears its hits: TAS-001/003/004 and both case tables are RED (5 failed, 91 passed). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Refs #4354
scan() returns its hits. Mutation table (tests run: secret + pool): killed no-gitleaks, no-builtin, no-allowlists, no-entropy, no-unescape, no-translate, and-allows; survived no-keywords (equivalent: the keyword prefilter is a speed-up, every upstream regex already requires its keyword). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Refs #4354
…ash-only index PRA-001 T7. FALSIFY-RCC-004 plants a sealed diff with whitespace, blank-line, partial-rename and re-base perturbation into a candidate pool. The exact checks (diff sha, hunk sha) miss it; the Index stores sketches but does not scan them yet, so RCC-004 is red (110/111). cluster.rs (MOSS-normalised 12-token shingles, bottom-64 MinHash, 0.7 containment), RCC-005/006 and the contract v1.1 land here with it. Refs #4354 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…item is refused PRA-001 T7. Index::scan now checks each text's shingles against the sealed sketches: >= 0.7 of an item's bottom-64 sketch present is a `cluster` hit, so the perturbed sealed item is Refused at pool admission (111/111). False-positive probe, 391 real .rs commit diffs from origin/main, all 152,490 ordered pairs: 7 clustered at k=12 (0.005%), all one family of templated beat test files (copies with renamed identifiers); k=5 gave 469. Mutation table: whitespace skip, rename, header skip, marker, min-shingles, blank-line skip, threshold and scan removal are all killed. Refs #4354 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…pend-only outcome joins; v1 rows still read (Refs #4354) v1 kept each counted lane's findings as a count, so every quorum lost its (input, output, findings) triple. v2 keeps the text verbatim and the lane's agent-trace-v1 input/output shas. Text that was never captured is None, a trace gap, never an empty list that would train as 'no issues'. - read_row dispatches on schema; both v1 shapes written under that name (402aff8 tagged shadow, 89d994b typed #436 shadow) still parse. - v1::Row::migrate keeps the count, leaves text/shas None, and refuses a legacy shadow (no cell/backend, so no typed #436 form). - OutcomeJoin: the latest join by at wins; the row is never rewritten. - contracts/review-ledger-v2.yaml carries FALSIFY-RXL-001..008; v1's stale Unknown{LaneUnavailable} equation now points at the typed form. Mutants killed (3/3): migrate invents Some(vec![]), trace_gap ignores missing text, an unknown schema read as v1. Crate lib 103/103, clippy -D warnings clean. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Top-k=20 teacher-logit blob per PRA-001 §2.7: u32 ids, T=1 log-softmax (f16), logsumexp_full (f32), retained mass M (f16) so 1−M is the true tail. Closed JSON header, zstd LE columns, one shared validator for encode and decode. Raw 130 B/token at k=20; encoded M6 needs real apr serve rows (top-k capture not yet in apr serve). FALSIFY-SPL-001..004 in contracts/sparse-logits-v1.yaml. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…te in a valid frame) (Refs #4354) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…independence-v1) New module kappa_probe.rs (stats.rs, spec §2–§5 and analysis-plan.md are untouched: prereg-locked). κ_err = the frozen stats::error_kappa on per-item error indicators vs GOLD, shadow lane vs every counted lane, over matured gold rows of one split. Unparsed verdicts, silver/pending/quarantined rows and pending/closed_unmerged outcomes are excluded, never counted as errors. A pair under min_n, or with κ undefined, is `insufficient` and prints no number. `rex lane-kappa` prints the scoreboard; with --candidate it gates: δ and min_n come only from the run manifest's lane_independence block, registered strictly before training started. Exit 12 = S-14 andon (κ rise > δ on any counted lane, a lane missing or insufficient on either side, an unusable manifest, or --candidate without --manifest). FALSIFY-LIND-001..004 were planted RED against a stub (4/4 failed), then GREEN. Four mutants each kill the falsifier they target: gate `>`→`>=` (003), no gold filter (001), registered_at `>=`→`>` (004), gate ignores the manifest's min_n (002; it first survived and got its own assertion). CLI smoke: exit 0 scoreboard, 12 on a voter-copying candidate, 12 with no manifest, 2 usage. `make lane-kappa ROWS=F [SPLIT=val] [MIN_N | MANIFEST CANDIDATE]` wraps it. Refs #4354 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…g_sha f0b4e4ac) (Refs #4354) - docs/specifications/PRM-001-prometheus.md = handoff v3, sha256 c9b52d08d7eb. - rex-prereg-v2 now locks v3 §2–§5 + stats.rs + analysis plan v2 + prompt v1 in docs/audits/prm-001/prereg.lock. The 17:09Z draft lock (1f4acb3d) is void. - v1 lock docs/audits/rex-001/prereg.lock (ef51087d) restored byte-exact from 5f0ae10 and never edited again (operator). The v1 spec §2–§5 and v1 plan are restored byte-exact too; FALSIFY-REX-PREREG-004 keeps them verifying. - v1 spec gains a superseded banner above §0 (outside §2–§5). - stats.rs: H7 max-pair gate `h7_holds` (δ-free, undecided on any undefined κ); the v2 silver-step `kappa_andon` is removed (v3 has no silver pool). - A planted §3 edit on disk: `rex prereg-check` rc=1 (DRIFT spec_s2_s5). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts: # README.md # contracts/census.json
… own cluster (Refs #4354) FALSIFY-TDD-001..004 fail 4/4 against the one-cluster-per-item stub. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s written, not Some("lambda-cuda")
Refs #4354
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…, H4+H5+H7 → tie-breaker, lambda primary → refused (S-4); S-4 reason names the cell Follow-ups named, not wired: pi#428 (quorum gate), infra#1095 (arbiter). Refs #4354 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…5/5 FALSIFY-TOJ fail G-MAT: gold only after the 14 d outcome window; pending is never a negative. TOJ-001 plants the v3 row's own check (a 3-day-old outcome refused as gold); TOJ-005 the weekly matured count. split_guard gains date(), the inverse of day(). Refs #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…the 14 d window (Refs #4354) resolve(): merged/closed/reverted/escape events per repo#PR give the v3 outcome. A revert <= 14 d after the merge is reverted_le14d; a later revert or any escape is regression_escape. Anything dated before the merge does not count. Until 14 UTC dates have passed the row reads pending with no matured date. gold() emits matured outcomes plus the latest HRQ ruling per round, and refuses forged immature rows and rulings with no ruler. matured_between() gives the weekly count. FALSIFY-TOJ-001..005 GREEN; 170/170 crate tests; clippy -D warnings clean. Mutation table: 19/19 killed (after adding before-merge revert/escape cases). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y (Refs #4354, #4447) Frozen 4-stratum set (2k/8k/16k/32k, 50 each) of real quorum diffs, sealed items excluded by sha/hunk/cluster, one per repo#PR, sha256(seed||sha) order, short strata refused. Summary refuses mixed sets, GGUF/cell/tag mismatch, unequal coverage, not-run rows; ratios are apr/llama_cpp; queue budget is the slowest voter's p95. Contract review-replay-v1 (FALSIFY-REPLAY-001..006). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… tests (Refs #4354, #4447) examples/replay.rs: build counts composed tokens via llama-server /apply-template + /tokenize, requires the sealed manifest AND sketch file; run replays each item through apr serve or llama-server (same GGUF/cell); voters extracts consultations.<lane>.duration_seconds from quorum receipts; summary prints the §9 speed block. Mutants on replay.rs: 20/20 killed after these tests (15/20 before). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ssion (#4026) The step record now carries chosen_logprob and the full-vocab logsumexp_full, both taken from the logits the top-k is ranked from, before any penalty. A sampled token outside the top k keeps a defined logprob, and sparse-logits-v1 can derive the tail mass (PRM C11). Chat requests take logprobs:true plus top_logprobs (0..=20). The Qwen3.5 buffered path returns choices[0].logprobs.content[] with token, token_id, logprob, top_logprobs[] and logsumexp_full. These are refused with 400, never ignored: top_logprobs without logprobs, more than 20, stream with logprobs, and any backend other than the Qwen3.5 session (the refusal names the architecture). Not yet: streamed logprobs, /v1/completions logprobs, and live controls on a real Qwen3.5 GGUF. Refs #4026 #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…gits-v1 sparse_logits::from_chat_completion reads choices[0].logprobs.content from a #4026 apr serve chat body and returns one TokenLogits per output token, ready for encode(). It refuses, by name, anything it cannot record honestly: no logprobs, no logsumexp_full (a hosted or foreign body, whose tail mass is unknowable), fewer than k alternatives, logprobs that don't descend, a logprob above 0, or a missing token_id. Tests: a serve-shaped body captures to the same record as TokenLogits::from_full_logits on the full row (70k vocab, so u32 ids are exercised), and it round-trips through encode/decode. There is a planted-case table for every refusal, with an unplanted control. Refs #4026 #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rt faithful apr serve (6f909ca, CPU, backend=cpu) chat top_logprobs=20 was compared with llama-server (60b06ab9a, -ngl 0) on Qwen3.5-0.8B Q4_K_M. The oracle ran the official template (enable_thinking false), took ids, and was teacher-forced on apr's generated ids. 6 prompts, 29 steps. Results: top-1 28/29 and overlap ≥16 28/29 (both misses are p1 step 0, " Paris" vs "Paris"). max |Δlogprob| was 0.19–1.65, and the pre-registered 0.25 bar fails on 25/29 steps. Controls: the oracle against itself passes (Δ 0); a one-step shift and a cross-prompt comparison both fail; three refusals return 400. Findings for the cop: a forward-parity gap between apr and llama.cpp on CPU Q4_K, and POST /tokenize not parsing special tokens. Refs #4026 #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lit, <|im_start|> parsed Refs #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…cial Refs #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s, hash-verified (Refs #4354, #4447) corpus v1 (REX-02) predates the T7 sketch step, so docs/audits/review-corpus had a manifest and no sketch file, and `replay build` refuses without one: with no sketches the contamination scan is exact-hash only and a near-dup of a sealed item could enter the replay set. `replay sketch --items DIR` writes the file from items-v1.tar, and only when every test-split diff hashes to its sealed diff_sha256. Run on /mnt/nvme-raid0/rex-corpus/v1/items-v1.tar: 105/105 test items verified and sketched. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…g cap The /tokenize reply for a large PRM-S1 candidate diff exceeded 10 MB and replay build aborted with 'response too big for into_string'. Refs #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…hat path 200/200 replay rows have null prompt_ms/prompt_tps/decode_tps for apr on gx10; llama.cpp d1d3c3396 returns timings on the same set and cell. Refs #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
§13.11 rung 1 — quorum shadow verdict Shadow mode: this records a verdict and merges nothing. A refusal |
The receipt asserted 'oracle vs itself: pass, max Δ 0' with no script or file behind it (sonnet-5 quorum seat, FAIL INV-4/5). positive_control.sh extracts compare() from c4026_live.sh unchanged and runs it on p2/p3/p5 oracle data reshaped into apr's response shape: pass, max Δ 0 on 7/8/8 steps. A 0.3-nat plant in one alternative fails it (Δ 0.30). Refs #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… limit (#4459 quorum r3) cap.saturating_add(1) left no headroom at cap = u64::MAX (agy r3). Read exactly cap bytes, then probe the underlying reader: Ok now means EOF for every cap. Invalid UTF-8 is refused via String::from_utf8. aprender-review-experiment --lib in a private target dir: 192 passed. Refs #4354 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Member
Author
|
Closed to bring open PRs under the cap of 10 (18 open). The branch is kept (not deleted), and so is its head. Reopen this PR, or re-cut it from the branch as a PR of at most 4 rows, after the 0.70.0 FINAL, when merge-main is allowed again. Why it was not landed or folded: the required checks Head |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PROMETHEUS (PRM-001 v3) branch
rex/001: the rows landed since the reviewed PRM-L headfb367bf7f(#4451, which this PR contains and which shrinks out of the diff once it merges).Closes #4446 (PRM-C12 terms/provider tagging, RED→GREEN at d59de74/e7e80cc43). Refs #4354 #4447 #4026.
Rows
/v1/chat/completionslogprobs+top_logprobs≤ 20 on the Qwen3.5 session, each token carryinglogprobof the chosen id and full-vocablogsumexp_full. Named 400 refusals: k > 20,top_logprobswithoutlogprobs, stream, non-Qwen3.5 backend.sparse_logits::from_chat_completionturns that body intosparse-logits-v1and refuses what it cannot record honestly.Live control (#4026, CPU, Qwen3.5-0.8B-Q4_K_M vs llama.cpp 60b06ab9a)
docs/audits/prm-001/c11-4026-live/RECEIPT.md. Top-1 28/29, top-20 overlap ≥16 28/29. The positive control and 2 planted controls behave, and the 3 refusals return 400. The pre-registered |Δlogprob| ≤ 0.25 bar is RED on 25/29. That is filed as a forward-parity finding, not a defect in the report. The bar is unchanged.Findings
docs/findings/2026-09-26-aprender-84.jsonl: CPU forward-parity gap (P2);/tokenizesplits<|im_end|>(P3).Tests
Targeted
aprender-serve(session, logprobs, chat_logprobs, openai_handlers) andaprender-review-experimenttests pass; clippy and fmt are clean. The full qwen35 serve suite is left to CI (it loads a real model in debug).🤖 Generated with Claude Code