run 差异、锚点与评分限制

findings.md — research findings log (paper/distill-fingerprint)

2026-09-02 — result-to-claim gate: PARTIAL (Codex gpt-5.6-sol, ultra)

• Verdicts: H1 partial, H2 partial, H3 partial, H4 partial, S1 partial, S2 partial, S3 partial, S4 yes, S5 no; overall partial; confidence high. Evidence pre-check 344/344 verified. Details and accepted wordings: CLAIMS_FROM_RESULTS.md.
• Gaps recorded: (1) no held-out same-pool anchor yet — pool confound larger than residual dose effect; (2) marginal-CI overlap is not equivalence — need difference CIs; (3) no multiplicity correction over 111 paired tests; MDE derived from a 237-task subsample; (4) behavioural reliability asserted, not measured; (5) model-only annotation of endings; (6) S1 cannot bound memorisation (MDE ≈ 15 pp at n = 88); (7) S5 contradicted by the bare-MoE direction and by incomplete teacher side-chain capture → dropped as a claim.
• Bookkeeping errors found and fixed in the ledger: corpus is 8,137 rows (not "8.6k"); teacher groups sum to 32,583 of 32,830 captured; Python-reference endpoints (s256 0.0195, s381 0.0178); s256 Pro composite 0.0215; Wald-CI vs exact-p statements for s256–n27 and s256–s128; dense-only 1.52–1.80 M range; 110 vs 109 regex-negative samples; 140 → 92 is −34 %.
• Constraint honoured: the twelve arms are final; no reruns. Only new experiments (held-out anchor, v12 on Qwen3.8-27B, MoE checkpoints, mini-SWE-agent check) may be added.

2026-09-02 — novelty-check: PROCEED WITH CAUTION, 7/10 (Codex gpt-5.6-sol, xhigh)

• Closest: arXiv:2604.21255 (ACL 2026 main) and 2606.16988; no paper contains the headline conjunction → no ABANDON. Adopted framing (b): "What agentic SFT buys: rapid behavioural and input-token convergence, while capability gains depend on the starting model." Caution items: held-out anchor, distance CIs + second distance, behavioural reliability, Holm, bare-vs-post-trained qualification central. Details: NOVELTY_REPORT.md; literature record lit/.

Presentation rule (user, 2026-09-02)

• No confidence intervals in the paper body. Report paired exact McNemar p (Holm-corrected) with significance markers and pp effect sizes; express "above noise" for fingerprint distances via the same-weights (temperature-twin) and split-half floors; one sentence on power (≈4 pp detectable at 80 % power with one rollout on 731 tasks). Bootstrap intervals computed in analysis/fp/ are internal checks (appendix at most).

Framing update (user, 2026-09-02)

• Significance statistics are de-emphasised (appendix only); the paper's two innovation lines are (1) behavioural analysis of what SFT transfers and (2) the SFT data-curation pipeline and what its rules change in the student (v12 vs v13 as the curation ablation; corpus thinking share → student overshoot; write-without-read as a corpus-taught habit → candidate curation rule). Source material being compiled in analysis/data/data_pipeline.md.

References supplied by the user

• Feishu wiki C4f2wlLpXi40qDk68Y7cCXhgn2e = repo file 经验/对外分享/sft20260901训练-评测报告-35B-A3B.md (external report on the bare-base A3B run; framing "the corpus performs the base→instruct step"; tables 1–12 reusable for Data/Cost sections). lark-cli is installed (~/.local/bin/lark-cli) if other Feishu docs are needed.

2026-09-02 — analysis agents (score statistics; data pipeline)

• Score statistics (analysis/stats/): 111 valid paired tests reproduce report.json; Holm over the 13 baseline-relative corpus tests leaves only a3b_2ep vs a3b_base; over all 111, 74 raw → 49 Holm → 71 BH (q38 vs n27 Verified and both dense-vs-sparse tests fall under Holm-111 but hold within their 6-test family). MDE (k = 1) median 4.03 pp Pro / 5.04 pp Verified; the 4.6 pp figure was N = 237. Twins: 97/83 flips (13.3 / 11.4 %). Pro minus teleport (655): s256 +2.60 (p 0.086), s381 −0.15, v13 +1.98, a3b_v12 +3.05 (p 0.054), a3b_2ep +9.62; teleport-only lift +3.9 to +10.5 pp for every v12/v13 checkpoint (repo-level exposure; 76 tasks).
• Data pipeline (analysis/data/data_pipeline.md): v12 funnel 12,257 → 12,036 → 8,601 (drop 28.5 %) → 8,562 → 8,137/425; raw-thinking floor 0.98 (not 0.90) explains v13's 9,880; masks 11,713 text / 16,856 call segments; v13 failed_conclusion 3,104/6,578; SUCCESS_CLAIM fires on 46.4 % of successful teacher endings; no single-rule ablation corpus was ever trained (v12→v13 changes floor, rules, inclusion, dedup, epochs together); 23 interrupted trajectories leaked into v13; three failed_conclusion rates circulate (cite decisions.tsv: 47.2 %).

2026-09-02 — data_pipeline.md open item #4 (answer_source_fetch 128 → 346), partial resolution

• git log on train/scripts/dataprep/case_audit.py between the v12 build (fb28fdc0, 08-06) and the v13 build (192b3b2a, 08-20): 59c10549 (08-13) completed the package-manager verb table and added a 200-char argument window (net 117 → 120 hits, audit column only); 192b3b2a rewrote the read layer for the v5 delivery format, changed fact_has_vendor_marks to "issues" in quality, moved patch sizes to code-only, added git_ref_content and pkg_upstream_source as decisions, and added HTTP_CLIENT patterns (requests/httpx/axios/fetch) to the network criterion. The 128 → 346 growth is therefore a rule-set change (broader fetch detection + cross-bucket guilt-by-association on 32,830 rows), not a data change; an exact per-row attribution would need a diff of the two decisions.tsv files (not done).

2026-09-02 — second (cross-model) annotation of the failure endings

• Codex gpt-5.6-sol relabelled the 240 endings + 68 RATIONALIZE windows blind: 0/110 admissions, precision 69/70, verification-habit shares n27 2/8 / s256 13/15 / teacher 11/15 — identical to annotator 1 on the quantities used by H4 (analysis/annot/second_annotator_agreement.md). Human κ pass still planned.

2026-09-02 — analysis agent (fingerprint level, analysis/fp/)

• Cache rebuilt (28 groups match fp_summary exactly; JS reproduced to 1e-15; gate 2 has 2 trials with extra session content vs the ledger snapshot, immaterial). Difference checks: epoch 1 moves bash/bigram/thinking-position/composite above noise, epoch 2 only bash (7× floor), epoch 3 nothing. Hellinger preserves all orderings; total variation shows SFT arms overshooting the teacher's Bash share (50.7 % → 60–62 %), so tool-mix convergence is ruler-dependent. Same-weights floor composite ≈ 0.001, bash 0.002–0.005; the published 0.002–0.003 "teacher floor" is a sub-population distance. Absolute post-error thinking rises after SFT (373 → 580 chars) while the ratio falls (ordinary-step thinking triples). Anchor: s256 reads fewer files (7 vs 11) but precision vs the teacher's set 0.50 → 0.80; negative-control Jaccard 0.001; the two p-values (0.073 / 0.099) differ by one task's void trial; Δsolve vs Δimitation ρ = −0.12.

2026-09-02 — paper-plan cross-review (Codex xhigh): overall 6/10 before fixes

• Weakest: page budget (4) and claim wording (5). Fixes applied in PAPER_PLAN.md v1 (see its Reviewer Feedback section). New requirement: an offline rule-impact table for §3 (per rule family: sole/marginal removals, overlap, removed-vs-kept behavioural statistics, masked loss mass, language/repo composition before/after) — commissioned to an agent. Held-out anchor is now the claim gate for the word "task-independent".

2026-09-02 — Fig 1 delivered (figures/, pgfplots agent)

• fig1.tex / fig1.pdf (6.36 × 2.59 in, fonts embedded) / fig1.png / fig1_data.json (every plotted number with source S1–S6) / caption.md (113-word caption + alt text). (a) epochs × bash-JS (log) for the dense, post-trained-MoE and bare-MoE lines, Qwen3.8 as a dashed reference, split-half floor band 0.003–0.005, right-axis Pro pass@1 as open markers; (b) median cumulative input (log) × Pro pass@1 with the three arrows (corpus on post-trained base: −53 % input, +2.7 pp; corpus on bare base: +8.9 pp, −28 % input; generation: +5.6 pp at 2.9× the input of +v12×2). No detectable-effect band, no panel titles, no funnel ribbon (height budget).
• Coordinator edits after delivery: the "(n.s.)" tag on the corpus arrow removed (no significance markers in the body); floor-band provenance corrected to the split-half means in analysis/fp/fp_ci.md (bash-JS n27 0.0045 / s256 0.0039 / teacher-sub725 0.0024 / q38 0.0013), not the teacher sub-population distances. Bare-base cumulative-input values (3.40 M / 2.44 M) are on no HTML page; they come from the A3B external report table 6 and were reproduced by 算_行为分解.py on the Pro job directories (a3b_q36 3.27 M and a3b_v12 2.05 M reproduced exactly as well). The 2.9× on the generation arrow is relative to the 1.52 M arm; relative to native 27B it is 1.35×, and the label says so.

2026-09-02 — ACL LaTeX scaffold (tex/, agent) and Table 1 fill

• tex/: acl-style-files @ d5adc823 (acl.sty / acl_natbib.bst copied), main.tex with [review], 8 section stubs + Limitations + Ethics + appendix A–C, macros.tex (arm macros, \todo), generated Table 1 (tables/make_tab1.py), hand-written funnel Table 2, Makefile. Placeholder build 10 pages A4, 0 errors. Float numbering differs from the plan (funnel = Table 1, arms = Table 2) — labels are stable. The style clone, build artefacts and main.pdf are git-ignored; make rebuilds.
• Table 1 completed: cumulative input for s128 1.69 / s381 1.75 (27B page COST array), a3b_base 3.40 / a3b_2ep 2.44 / a3b_q36t06 3.13 / a3b_v12t06 2.01 M (算_行为分解.py on the Pro job directories, 2026-09-02); Bash-JS now read from fp_summary.json (v13 0.055, twins 0.200 / 0.069); teleport-excluded Δ for q38 vs n27 +5.04 pp ((371 − 338)/655, stats_fixes §e1) and for the twin pair +1.83 (§e2). n27 and a3b_q36 both at 3.27 M is a verified coincidence. The T = 0.6 rows stay in the table as a temperature-perturbation reference (PAPER_PLAN v1 says 12-arm table; NARRATIVE_REPORT's "DROP as table rows" is superseded).
• References: SWE-bench (2310.06770, ICLR 2024) and the Qwen3 technical report (2505.09388) added to the arXiv-verified record (refs.bib now 44); SWE-bench Pro (2509.16941) was already there. lit/refs_manual.bib holds the vendor/blog items (SWE-bench Verified, Claude Opus 4.8 system card, Claude Code, Qwen3.5/3.6/3.8), all marked [VERIFY] for citation-audit.

2026-09-02 — Fig 3 delivered (figures/fig3.*, agent)

• Double-column 6.21 × 2.36 in, six panels on Pro along native → +v12 ×1/2/3 → +v13 → native Qwen3.8, MoE pair dodged, teacher dashed: (a) protocol-error recovery@3 0.299 → 0.58–0.63 (teacher 0.603, q38 0.338); (b) thinking chars per step after an error vs other steps (373/359 → 580/1,199; teacher 248/726, pattern copied, level overshoots); (c) thinking share 0.858 → 0.93–0.95 (teacher 0.893, corpus 0.881); (d) truncation pipes 0.432 → 0.64–0.65 (teacher 0.672); (e) write-without-read flat 0.09–0.11 (teacher 0.110, q38 0.177); (f) task-management calls 5.2 % → 0.3–0.5 % (teacher 3.2 %). Left out with values recorded in fig3_data.json: thinking-position JS, Bash head shares, sub-agent dispatch (S5 caveat), bare-MoE pair, Verified pool.
• Definitions to carry into the text: recovery@3 = same tool, same target, success within 3 calls (not "any means"); the "corpus 0.672" truncation-pipe reference in the scaffold is really T_hp; panel (b) means are over all steps (teacher thinks on 43 %/50 % of them), so the caption claims the pattern, not the level.

2026-09-02 — Fig 4 and Fig 5 delivered (figures/fig4.*, figures/fig5.*, agent)

• Fig 4 (appendix B, single column 2.95 × 1.76 in): (a) ECDF of per-task read-set Jaccard to the teacher for n27 vs s256 (medians 0.44 → 0.60, closer on 66/88) with the two negative-control levels 0.001 / 0.03; (b) Δsolve vs Δread-Jaccard scatter (ρ = −0.12). Source anchor_diag.json, cross-checked against p8_train100.json.
• Fig 5 (appendix C, double column 6.23 × 2.53 in): (a) three cumulative-input bridges (n27 → s256 3.27 → 1.52 M: steps −32 %, context −29 %; a3b_q36 → a3b_v12 3.27 → 2.05; n27 → q38 3.27 → 4.41); (b) tool-output characters by call-parameter bucket (112 / 46 / 47 / 49 / 93 / 109 / 53 M) with the truncation-pipe and limited-read strips and teacher lines. Every number from one run of 算_行为分解.py on the ten Pro job directories; it reproduced all page values exactly.
• Corrections carried into the ledger: the narrative's "78 % of the drop is whole-file reads" is the MoE-line share, the dense pair gives 74 %; "pipes 47 → 69 %" is 47 → 68 % (fp_pooled 0.471 → 0.683), dense 43 → 65 %. The s256 step median is exactly 35.5; the page's integer rounding (36) gives the accepted −1.05 / −0.70 M attribution, the exact median −1.08 / −0.67 M (both in fig5_data.json). Appendix C's figure environment switched to figure* (double column).

2026-09-02 — rule-impact analysis delivered (analysis/data/rule_impact.*, agent; first attempt stalled on a single oversized Write and was restarted with a chunked brief)

• Stage 1 reconciles exactly with build_report.json and the funnel (12,036 → −3,435 → 8,601 → −39 → 8,562 = 8,137 + 425; v13-HP drop 2,377). Marginal removals: raw_thinking_low 2,450 (71 % of drops), vendor_sourcing_mark 493, leak family 45 (answer_source_fetch / upstream_clone / integrity_policy_violation / gold_test_fetch fully shadowed), sandbox_escape 31, multi_user_turn 27; demote tier removes nothing (1,050 demoted rows trained).
• Stage 2: raw_thinking_low leaves the tool mix untouched (JS 0.005) but removes the longest, think-heaviest rows (median 68 vs 44 steps; think share 0.897 vs 0.838); sourcing/leak rules and the 221-row vendor removal take a Bash-heavier (58–63 % vs 51 %), cd/git-heavier profile (JS heads 0.039–0.065 vs floors ≤ 0.002) with far worse recovery@3 (leak family 0.24 vs 0.60); sandbox_escape most distinct per row (JS heads 0.105) but 31 rows. Sidechain share 0 everywhere by construction.
• Stage 3: v12 masks 2.77 % of learnable characters (assistant 1.23 %, tool_call 4.77 %; segment counts match data_pipeline.md exactly); harness_error 0.78–2.32 % (bracket); v13 3.43 %, failed_conclusion 0.15–0.94 % over 3,088 rows, failed population 4.28 % masked.
• Stage 4: Go 37.0 → 38.1 %, TypeScript 13.7 → 15.1 %, Rust 6.4 → 4.3 %, Java 9.8 → 8.0 % (Rust/Java over-represented in drops); top-15 repos 40.7 → 37.7 %; cap inert (max repo share 6.66 → 6.69 %, top-5 22.4 → 22.4 %); v13 by-product inclusion is the only composition shift (Go 28.2 %), no cap, 1.82 rows/task (cap 2 would cost 4,320 rows).
• Caveats recorded in rule_impact.md §5: task counts 7,427 (base_task_id) vs 7,435 (instance); 59/65 rows with a mask kind but no masked segment; raw.pertask2's −3,985 is at floor 0.95 (different base); SUCCESS_CLAIM kept 48.2 % vs dropped 42.3 % (pooled 46.4 %). The 15 MB rule_impact_cache/ is git-ignored.

2026-09-02 — plan page published

• Artifact https://claude.ai/code/artifact/92aa45cd-e325-4198-9b3b-86d747e50079 (claims ledger, gate verdicts, Fig 1/3/4/5, page budget, deliverables, decisions pending, next steps). Source page built from PAPER_PLAN.md + figures/*.png; update by republishing to the same URL.

2026-09-02 — same-pool anchors commissioned (anchors/)

• User decision: run both. (A) val373: the v12 val split (425 rows → 373 base tasks, zero task overlap with train) replayed once by n27/s128/s256/s381/v13/q38, same harness/instruction/temperature as the twelve arms; template = the 88-task train replay (chain_train100_zhihu.sh, oracle gate, python3 pre-bake). (B) pro100: teacher Claude Opus 4.8 in Claude Code 2.1.202 with reasoning_effort=high on 100 Pro tasks (seed 20260902; teleport excluded; all 12 arms valid; repo × difficulty stratified: easy 44 / mid 19 / hard 37; Python 40, Go 32, TS 22, JS 6), harbor call copied from run_eval.sh:687 with only endpoint/key/model/effort changed; container proxy via 192.168.10.131:7890; cost gate after 3 tasks.
• Facts that shape the runs: no val373 image is resident on either machine (registry sotalab-cn-beijing.cr.volces.com, 373 pulls + pre-bake); gpu08's 8 GPUs are fully held by RL (ray + sglang), gpu23 has ~55 GB free per GPU beside areal-rl/slime-rl → placement is zly's call; v13 step-362 HF snapshot exists only on gpu23 (gpu08 has the trainer checkpoint); natives at /data1/models on both machines; s128/s256/s381 HF checkpoints under $SWE/models on gpu08. Teacher thinking will be summarised by the API, so only tool-based rulers apply to (B); temperature is API default vs students' 0.8.
• What they gate: "task-independent" in H1 (A), pool confound removed on the benchmark itself (B). The 88-task anchor stays in the appendix either way.

2026-09-02 — anchor briefs finalised after the harness survey (Explore agent, 153 tool calls)

• Placement: pro100 (teacher) on gpu08 — all 100 Pro images resident (280 GB), harbor + containers there, proxy to Anthropic through gpu08's mihomo (192.168.10.131:7890), no GPU needed; inst1 has 0 Pro images (only OCI archives), gpu23 225/731. val373 — pool built on gpu08 (delivery tree only there) and rsynced to gpu23; harbor, containers and images on gpu23 (99/373 delivery images already resident, 3.2 T free, 42 s per 2.4 GB pull with existing volces credentials); engine either inst1's eight GPUs (held by our own idle q38 engine container n27, last request 2026-09-01 13:32; weights Qwen3.6-27B/step256/step381/Qwen3.8-27B there, step128 + v13 to transfer at 19 MB/s) via remote_engine.sh, or gpu23 at reduced util beside areal-rl — user's call.
• Facts from the survey: the twelve arms share one harbor call (run_eval.sh:687-695), Claude Code 2.1.202, NET_POLICY off, 5400 s wall (1.8× of 3000 s; train replay 1.5× of 3600 s), no turn budget, extra instruction appended to the task prompt; the teacher corpus was captured by the vendor with claude-opus-4-8, thinking:{type:adaptive, display:summarized}, max_tokens 64000; no prior Anthropic-API run exists in the harness; _MIRROR_BASE reads INSTALLER_DOWNLOAD_URL from harbor's own process env (the 08-19 264/264 failure); Pro Dockerfile builds need the private DOCKER_CONFIG proxy; ops helpers exist for every val373 step (pull_delivery_images.sh, volc_cr_login.py, prebake_python3.sh, oracle_pool_check.sh, mk_oracle_pass_pool.py, oracle_pool_report.py).
• pro100.json now carries pool_dir (case-insensitive prefix match; NodeBB dirs are lower-case, report ids may end in -vnan) and image per task.

2026-09-02 — paper-write draft v1 (tex/)

• Full draft: abstract, §1–8, Limitations, Ethics, appendix A (7 tables + endings audit), B (8), C (5) — every appendix table generated from the analysis JSON/MD by tex/tables/make_tabA/B/C.py; references.bib filtered to the 45 cited entries (tables/make_refs.py). Body ends on page 8 of the ACL two-column build (28 pages with appendix), 0 errors, 0 undefined references, no figure fallback text in the PDF.
• Codex cross-review (xhigh) before fixes: calibration 6.5, clarity 7.0, readiness 4.5; all CRITICAL/MAJOR items applied (see traces/paper-write/2026-09-02_run01/README.md). Two blanks are deliberate: % DATA_NEEDED: ANCHOR_VAL373 and % DATA_NEEDED: TEACHER_PRO100 with visible \todo markers in §5.1, §5.2 and Limitations; conclusions are written per the plan and will be re-scoped when the anchors land.
• Remaining \todo markers: the two anchors (3 sites), the human annotation pass (Table 3 caption), per-arm distribution tables (Appendix B). Next: paper-claim-audit and citation-audit (refs_manual entries still [VERIFY]), then kill-argument / auto-review-loop.

2026-09-03 — val373 anchor running; two facts that constrain how the arms may be read

• The v12 "step series" is not one training run. Read from each checkpoint's own args.json:

s256 and s381 come from the same run (output_dir=/output/qwen36-27b-v12-3ep/v0-20260809-124327, max_length=196608, save_steps=64); s128 does not (output_dir=/v12/ckpt128-hf, max_length=262144, no save_steps). All three start from /models/Qwen3.6-27B on corpus v12, bfloat16, seed 42. So only s256→s381 is a same-run step pair; s128 is an early checkpoint of a *different* v12 run whose context window was 256k rather than 192k. The stored args.json leaves learning rate / batch / train_type as null, so the remaining hyper-parameters cannot be compared: we can show the two runs differ, we cannot show the difference is only the step count. 🔴 Do not draw 128→256→381 as a single training trajectory in the paper.

• reward=1 and AgentTimeoutError co-occur — the agent is cut at the 5400 s wall but the patch it

had already written passes the tests. Measured on this anchor: 11/136, 11/129 and 6/59 of the reward=1 trials in s381 / v13 / s128. The BRIEF's paper rule ("anomalies count as unsolved") therefore moves pass@1 by 4–6 points, so collect_val373_results.py now judges exception before reward and reports both the strict (paper) and lenient numbers. Any arm-to-arm comparison must use one of them.

• Placement actually used (three lines, 2 arms each, all six at concurrency 20, 5400 s wall):

A = gpu23 local engine (s381, n27); B = gpu08 local engine (s128, s256); C = inst1 engine via 12 SSH tunnels + an nginx round-robin on gpu23 (v13, q38). The v12 checkpoints all sit on local-engine lines; only the reference arms (v13, q38) take the cross-border tunnel.

• Landed so far: s381 345/345 strict pass@1 49.0 % (void 42, of which 11 passed anyway);

v13 345/345 strict 50.1 % (void 62, of which 17). Native n27 voids far less than the trained checkpoints (10/112 vs 12–18 %) — an endings-behaviour signal worth its own look.

2026-09-05 — val373 anchor complete, eight arms

All eight arms finished 345/345 (bench/results/anchors/val373/summary.md, regenerated 09-06):

Reading notes for the paper:

• 🔴 The v12 checkpoints do not order by step: 128 (51.9) > 256 (51.3) > 381 (49.0). Combined with the

earlier finding that s128 comes from a *different run*, the only same-run pair (s256→s381) goes down 1.3 points from 256 to 381 steps. Nothing here supports "more SFT steps → better".

• 🔴 q38 is the void story, not the capability story: strict 35.4 % vs lenient 53.0 % — 61 of its 147

anomalies had already written a passing patch. It is slow (median 4745 s against a 5400 s wall), so the wall, not the model, produces most of its gap. Quote the strict number only with this caveat attached.

• Native arms void far less than SFT'd ones on both bases (n27 35, a3bnat 32 vs a3b2ep 70, v13 62, s128 46).

The endings-behaviour signal survives the base change, which makes it a stronger candidate for the fingerprint claim than any single pass@1 delta.

• 🔴 The two MoE arms are not same-family with the six 27B dense arms, and not with each other

(Qwen3.5 vs 3.6 base): read each as its own control, never subtract them for an "SFT delta".

• Placement caveat added late: s256 moved from line B (gpu08 local engine) to line C (inst1 engine over

the tunnel) on 09-04 when gpu08's GPUs were handed to RL. So within the v12 axis only s256 took the cross-border tunnel.

Open items

• [x] Teacher manifest rescan (2026-09-02, 32,830 rows, 0 errors): the 247 trajectories outside the three populations are byproduct manifests with status = error (neither passed nor failed; ver_passed False, q_passed None). Full breakdown: high_precision passed 12,257 (q_passed True 11,115 / None 1,142); byproduct passed 9,843 (True 8,609 / None 910 / False 324); byproduct failed 10,483 (True 8,404 / False 1,079 / None 1,000); byproduct error 247.
• [ ] Analysis fixes listed in CLAIMS_FROM_RESULTS.md §2 (a); human re-annotation (b); GPU experiments (c).
• [ ] Re-run result-to-claim after (a)–(c); then paper-plan (venue ACL, 8 pages) → paper-write.