Qwen3.5 × LoGRA × IFBench: Gradient-Sketched Reinforcement Learning and Verifier Reliability

Research code, evidence archive and two manuscript drafts. This Hugging Face repository is public. Both manuscripts (Paper A, Paper B) are unsubmitted research drafts; nothing here is peer reviewed, and no label has been independently adjudicated by humans.

Reading guide. Start with Current status, then Verified findings and Limitations. Historical documents contain claims that were withdrawn; docs/technical/RESEARCH_HISTORY.md lists them. The pre-release README is preserved at docs/archive/README_pre_release_M5C3.md.

0. Latest release (M13): two papers and the M8–M12 archive

Status. Milestones M0–M12 are complete. The VM-side research workflow is closed; this release publishes all safe new M8–M12 artifacts. Both manuscripts are unsubmitted research drafts (not peer-reviewed, no DOI, no acceptance). Labels, oracles and specification interpretations are AI-assisted and not independently human-adjudicated; six oracle-specification disagreements remain matters of interpretation.

Paper A (summary). Compares conventional test-generation families at matched verifier-call budgets on injected-fault panels (RESERVE, Panel S, Panel F), 24 reconstructed upstream faults, alarm precision and benchmark-score sensitivity. Oracle-based methods lead oracle-free ones, the equal-budget combination is not better than its best constituent, the better oracle family differs by panel, B6 finds all 9 crash faults but 6 of 15 behaviour faults, and reference-free alarms have low precision at realistic prevalence. Limits: one AI agent wrote all oracles/labels; detection is conditional on a known-good parent; no human adjudication.

Paper B (summary). Tests prompt-only Qwen3.5-0.8B/2B/4B as input generators for verifier fault detection. Study 1 (frozen, 0.8B): no model stream matched RD-B9 at 500 calls (0.439 vs 0.658). Study 2 (frozen protocol, configuration chosen on development faults only, one confirmatory run on 24 held-out audit-derived faults): Qwen3.5-4B does not establish superiority over the strongest conventional baseline. Detection rises with size; post hoc, the 4B model alone reached two faults that no conventional stream reached, reproduced as version divergences against an AI-assisted reference — not confirmed semantic defects; an exploratory 250/250 hybrid is not statistically established. Actual GPU cost is reported (4B: ≈113 GPU s per 500-call stream).

Held-out detection probability (24 faults, confirmatory; from runs/m12/analysis_heldout.json):

Primary contrast (4B-G2 − RD-B2, N=500): +0.029, 95% CI [−0.104, +0.158]; the pre-declared margin of +0.10 was not met (H12-1 not supported). 4B − P1: +0.323 [+0.113, +0.531], Holm 0.042. Paper A representative result (Panel F, N=500, 190 mutants, 22 templates): B2 0.700, B5 0.846, B6 0.798, B6P 0.923 (B6P uses k× calls and is not matched).

Navigation. Paper index · Experiment index M8–M12 · Reproduction index · Checkpoint addendum (M8–M12 produced none) · A/B relationship · M12 reports · M12 final report · verified GPU ledger · security/privacy audit · license review · excluded artifacts · QA · HF verification · closeout.

Architecture diagrams. The five original diagrams (assets/architecture/) remain accurate for M0–M7; the new overview of Papers A/B is below (source assets/architecture/diagram6_paper_ab_overview.mmd).

Reproducing (CPU only).

python3 -m pytest tests -q                      # CPU tests (skips are reported, not counted as passes)
python3 scripts/m10_check_paper.py              # Paper A consistency checker
python3 scripts/m12/repro_check.py              # Paper B confirmatory/hybrid/complementarity reproduction (4/4)
bash scripts/release/build_papers.sh            # rebuild both PDFs
python3 scripts/release/verify_manifest.py docs/release/M13_RELEASE_MANIFEST.csv   # SHA-256 verification

GPU-dependent steps (new generation/training) are not needed to verify any reported number.

AI assistance. Implementation, oracle construction, analysis and drafting involved substantial Claude Code assistance under the project owner's direction; human authors remain responsible for reviewing and approving any submission. No external human validation of labels has occurred.

Not published (by design). Private blinded-review answer keys, credentials, agent-instruction files, model weights, and 18 third-party source snapshots whose licence could not be verified (see excluded artifacts). Older public revisions of this repository exposed earlier review keys; this cannot be undone by a new upload (see the security audit).

Contents

1. Overview · 2. Status · 3. Contributions · 4. Architecture · 5. Methodology · 6. Verified findings · 7. Repository structure · 8. Installation · 9. Reproduction · 10. Datasets · 11. Results and limitations · 12. Paper · 13. Reproducibility and provenance · 14. Citation · 15. License

---

1. Project title and overview

Original motivation. NVIDIA's LoGRA (Zhang et al., arXiv:2610.06647) reduces the memory of reinforcement-learning post-training by keeping low-rank sketches of gradients instead of dense gradients or adapter parameters. This project asked whether a single-GPU port of that optimizer can train Qwen3.5-0.8B with reinforcement learning from verifiable rewards (RLVR) on instruction-following constraints, and how it compares with LoRA.

Background.

• LoGRA. For selected linear layers the base weight W stays frozen. A random projection A (rank r) turns each backward pass into a sketch S = G Aᵀ of the weight gradient G; a RowAdam-style optimizer turns the accumulated sketch into a low-rank update ΔW = U A that is written directly into the base weights (not an adapter). NVIDIA's full method also has a predicted-KL step controller, which this project did not implement.
• Qwen3.5-0.8B is a hybrid model: 18 Gated-DeltaNet (linear-attention) layers and 6 full-attention layers with tied embeddings (752,393,024 unique parameters).
• RLVR for instruction following. A prompt carries verifiable constraints (e.g. "no commas", "exactly 3 bullet points"); a short program per constraint (the verifier) returns pass/fail and the reward is the share of constraints passed. IFEval (25 constraint types), IFBench (58 out-of-domain constraints) and the IF-RLVR training data/verifiers (allenai/IF_multi_constraints_upto5, open-instruct IFEvalG) are the benchmarks used here.
• Why verifier reliability matters. A policy optimised against a verifier can only be as good as the verifier's verdict. Narrow verifiers check the constraint, not the task, and implementation errors silently change both evaluation and reward.

How the project developed. The optimizer study (M0–M5B) produced a corrected, audited LoGRA port and inconclusive LoGRA-vs-LoRA results. While diagnosing why official IFBench accuracy rose without evidence of better answers, we found that a fixed one-word response satisfies about a third of IFBench prompts. That led to a systematic, CPU-only cross-benchmark verifier audit (M5C-0 to M5C-3), which became the project's principal publication contribution.

Two tracks, clearly separated.

2. Current research status

3. Research contributions

Only contributions supported by saved evidence are listed. The principal contribution is the verifier audit.

1. Single-GPU LoGRA gradient-sketch optimizer adaptation, mathematically validated against NVIDIA's reference (projection law, per-layer seeds, hook-based sketch accumulation, RowAdam, layer-scale preservation, trust region, stochastic rounding, exact-restore checkpoints) — docs/technical/LOGRA_IMPLEMENTATION.md, docs/audits/NVIDIA_LOGRA_SOURCE_AUDIT.md.
2. Qwen3.5-0.8B GRPO experiments with LoRA and LoGRA (inconclusive on accuracy; decode-matched efficiency analysis shows the earlier wall-clock advantage is a LoRA unmerged-adapter overhead, not an optimizer advantage).
3. Cross-benchmark audit of the IFEval, IFBench and IF-RLVR verifiers with one protocol, including the training-time implementation.
4. Reproducible fixed-response, wrong-prompt and counterfactual probing (prompt-independent acceptance at template, instance and prompt level).
5. Independent verifier-oracle implementations (56 of 112 templates) with metamorphic and property-based tests.
6. Validated candidate defects and specification ambiguities (four IF-RLVR/IFBench candidates plus one already documented for IFEval), each with a minimal reproduction; upstream drafts prepared, not submitted.
7. Release-ready audit toolkit and tests (86 CPU tests pass; scripts regenerate every table and figure from stored results).

Auditing verifiers is not claimed to be new (IFEval++, IFBench PR #34 and others exist); see docs/research/M5C3_PRIOR_ART_UPDATE.md.

4. Research architecture

Figure 1. Two linked tracks. Track A (left) trains Qwen3.5 with GRPO under a LoRA or LoGRA optimizer and evaluates on IFBench. Track B (right) probes the official verifiers with fixed, wrong-prompt and transformed responses, compares them with independent oracles and classifies each template. The tracks meet in the shared evaluation analysis (saved IFBench-test responses are the natural-response pools of Track B). Editable sources: assets/architecture/.

Figure 2. LoGRA optimizer as implemented (not LoRA): projection A, hooks, sketch S = G Aᵀ, RowAdam, separate update scaling, ΔW = U A applied to the base weights. The predicted-KL controller is not implemented.

Figure 3. GRPO experiment pipeline shared by both optimizers, with logging, GPU-budget accounting and checkpoint validation.

Figure 4. Verifier audit pipeline and the five outcome categories.

Figure 5. Milestones M0–M5C-3 and release; open items are shown as not completed.

5. Research methodology

Details: docs/technical/.

6. Verified findings

Every number below is read from a saved result file (source column). Percentages with different denominators are never combined.

6.1 Benchmark collections audited

6.2 Fixed-response (content-free) acceptance — official verifiers, strict

This is expected narrow-constraint behaviour (the checks do not read the task), not by itself a software bug. Source: runs/m5c1/audit_raw.json, docs/reports/m5c1_cross_benchmark_tables.md.

6.3 Wrong-prompt acceptance

A response written for a different prompt (same template, identical arguments) keeps its passing verdict in ≥ 95% of cases for 19/19, 26/26 and 17/17 evaluable templates (IFEval / IFBench / IF-RLVR), i.e. all verifiers are prompt-independent given their arguments. Pooled retention of originally passing instances: 99.9% / 100.0% / 100.0%. Whitespace changes keep 99.7–100%; word shuffling keeps 76.4% / 84.5% / 91.8%; lexical replacement 62.8% / 70.0% / 64.6% (docs/reports/m5c1_cross_benchmark_tables.md §4).

6.4 Independent oracle coverage

65,910 verifier-vs-oracle comparisons (49,538 + 16,372); 2,218 disagreements (954 + 1,264), all adjudicated by hand (defect, interpretation, or oracle choice). The 17 newest oracles were written after reading the official code, so independence is limited. Source: docs/research/M5C2_ORACLE_COVERAGE.md, runs/m5c2/oracle_audit.json.

6.5 Defect candidates and specification ambiguities

Reproduced on the latest upstream heads (open-instruct 11826255, IFBench 1c40f0c1, 14 Oct 2026); none submitted upstream; none touched by IFBench PR #34.

Sources: docs/research/M5C2_VERIFIER_BUG_VALIDATION.md, runs/m5c2/defect_validation*.json, docs/upstream/.

6.6 RL evaluation results — exploratory (Qwen3.5-0.8B, IFBench-test, 300 prompts)

Interpretation limits: no significance or equivalence claim (3–5 seeds); the IFBench test set is not pristine; the adequacy rows are AI-judge-dependent and not independently validated; the optimizer difference is not interpreted. The direction (compliance up, task engagement down) agrees with published reports and is not claimed as new.

6.7 Resource consumption

7. Repository structure

Derived from the actual files (docs/inventory/FILE_INVENTORY.csv lists every file with SHA-256, purpose, milestone and publication status).

.
├── README.md                     this file (previous README: docs/archive/)
├── LICENSE                       Apache-2.0 (project code)
├── pyproject.toml, Makefile, requirements-audit.txt, .env.example
├── src/q35_logra_ifbench/
│   ├── methods/                  logra_core.py (sketch optimizer), logra.py, lora.py, grpo*.py
│   ├── rl/                       trainer, rollout, scoring, policy, grpo_core, checkpoint, evaluate, adapters, policy_kl, autotune_pin
│   ├── models/qwen35.py          model/tokenizer loading
│   ├── data/                     download, splits, ifbench_io (loaders), rewards, feasibility
│   ├── diagnostics/              verifier_audit (probes), oracles, oracles_m5c2, verifier_candidates, shortcuts
│   ├── budget/ledger.py          persistent GPU-hour ledger
│   └── config.py, seed.py, hardware.py, watchdog.py, logging/, evaluation/, publishing/
├── tests/                        data/, reference/ (21 NVIDIA-differential), rl/, unit/   (86 tests)
├── scripts/                      experiment drivers (m4/m5), analyses (m5a-m5c3), release tooling, publish_hf.py, gpu_ledger.py
├── configs/experiments/          m4_matched_pilot.yaml
├── manifests/                    dataset split/hash manifests (M4/M5A)
├── runs/                         evidence by run (indexed in docs/index/EXPERIMENT_INDEX.md)
│   ├── gpu_ledger/               ledger, hash-chained events
│   ├── paired_*, m5a_*, dev_*, rep_*, smoke_*, resume_*, repro_*, m5b_*   training / benchmark runs
│   └── m5c0..m5c3, _analysis, _logs, _pins                                CPU analyses, derived tables, logs
├── docs/
│   ├── technical/                LoGRA, Qwen3.5, GRPO, benchmark auditing, research history
│   ├── index/                    experiment index, checkpoint index, GPU accounting
│   ├── inventory/                file inventory, duplicates, path/dependency checks, secret scan
│   ├── audits/                   NVIDIA source audit, diff matrix, claim verification, reliability gate
│   ├── experiments/              frozen protocols (M4, M5A, M5B)
│   ├── reports/                  milestone reports, tables, publication records, REPRODUCTION_VALIDATION.md
│   ├── research/                 literature, novelty, verifier taxonomy, oracle coverage, judge reliability
│   ├── upstream/                 4 upstream issue drafts + reproductions (not submitted)
│   └── literature/, feasibility/, ops/, archive/
├── paper/                        main.tex, main.pdf, references.bib, figures/, tables/   (draft)
└── assets/architecture/          5 diagrams (PNG, SVG) + Mermaid sources

Not part of the archive: third_party/labs-molt (NVIDIA reference clone, pinned by commit), data/cache/ (downloaded datasets, upstream verifier clones, paper PDFs), CLAUDE.md (internal agent instructions). Empty placeholder directories are not listed.

8. Installation

Verified environment (docs/audits/environment.json, runs/m5c1/pins.json): Ubuntu, Python 3.10.12, torch 2.6.0+cu124 (CUDA 12.4), transformers 5.16.1, peft 0.21.2, flash-linear-attention / fla-core 0.5.2, triton 3.2.0; one NVIDIA L4 (23 GB) for GPU work. causal_conv1d is not installed.

9. Reproduction

All commands marked tested were executed on the release host during this release; exit codes and timings are in docs/reports/REPRODUCTION_VALIDATION.md. Commands that rewrite tracked evidence were run and the committed evidence restored afterwards. CPU-only commands set CUDA_VISIBLE_DEVICES=.

10. Benchmark datasets

Reserved untouched sets (manifests/m5a_reserved_seen_confirm.jsonl, allenai/IFBench_multi-turn) were never read. The IFBench-test set was used repeatedly and is not pristine.

11. Results and limitations

Terminology used throughout.

• Constraint compliance — the official verifier passes the response (deterministic).
• Substantive task fulfilment — the response engages with the actual task (needs a rater; here AI-rated, exploratory).
• Verifier implementation error — the verdict contradicts the instruction text / a literal oracle (reproduced defects, §6.5).
• Expected narrow-constraint behaviour — checks that ignore content by design (e.g. "no commas" accepts Yes.); not a bug.
• Reward shortcut — a policy raising the verifier reward through responses that satisfy the checks but not the task; documented in prior work (IFBench, IFDecorator, Wei et al.) and only illustrated here.
• Model-dependent observation — anything measured on Qwen3.5-0.8B (response length, repetition, accuracy trends).

Limitations. (1) No independent judge: adequacy labels come from four same-vendor AI raters (Fleiss κ 0.86 binary); cross-family and human agreement are unmeasured. (2) Oracle coverage: 56 of 112 templates; 17 oracles were written after reading the official code; one adjudicator. (3) Single-model RL evidence: one 0.8B model, one benchmark, 3–5 seeds, 10–18 passes per rated cell; no significance claim. (4) Statistical uncertainty: df = 2 intervals are descriptive; the test set was reused. (5) The LoGRA port lacks the predicted-KL controller and uses a different RL objective than NVIDIA's; no speed/stability/accuracy advantage is claimed. (6) The prior-art search was single-reviewer; no claim of priority.

11b. M6 verifier-audit extension (CPU-only, frozen protocol)

A pre-registered, CPU-only comparison of verifier-testing methods (fixed probes, random, boundary+oracle, property-based+metamorphic, metamorphic follow-up, broad oracle differential, and a combined framework B6) on injected and historical faults in IFEval/IFBench/IF-RLVR verifiers. Headline (held-out, N=500 calls, 189 mutants/36 templates): B2 0.877, B5 0.864, B6 (equal budget split) 0.863, oracle-free methods 0.38-0.42; B6 minus best of B2/B3/B4 = -0.022 [-0.057, 0.014], so the combined framework does not beat its constituents at matched budgets. Pooled B6P (0.944) uses k x more calls and is not a method gain. Oracle availability is the decisive factor.

• Full details: docs/reports/M6_WORK_REPORT.md; comparison docs/reports/M6_BASELINE_COMPARISON.md; failures docs/reports/M6_FAILURE_CASES.md; historical faults docs/reports/M6_HISTORICAL_FAULTS.md; real-world impact docs/reports/M6_REAL_WORLD_IMPACT.md; threats docs/reports/M6_GENERALIZATION_AND_THREATS.md; protocol/amendments docs/research/m6/.
• Constraints honoured: zero GPU, no paid APIs, upstream sources not modified, no upstream issues/PRs posted, manuscript not submitted. The private blinded-review key is not uploaded.

11c. M7 verifier-reliability extension (CPU-only evidence; GPU pilot designed but not run)

Branch research/m7-verifier-reliability-and-rl-impact. Reproduced M6 (0 disagreements), added a 13-fault real-fault registry (6 root-cause commits, including upstream PR #1718), a 57-alarm single-annotator adjudication, denominator sensitivity, score-impact re-scoring (sub-bullets, IFBench version drift, PR #1718 rollout impact) and a 7-page working manuscript (paper/m7/main.tex, not submitted). Headline: oracle-aware testing dominates; the equal-budget combination (B6 0.863) does not beat its best constituent (B2 0.877); real faults are harder (B6 finds 5/13). Not done: the reward-correction RL pilot (GPU occupied by unrelated services; see docs/reports/M7D_RL_REWARD_CORRECTION.md); no independent human review exists. Index: docs/research/m7/M7_MASTER_STATUS.md, docs/reports/M7A_*, M7B_*, M7C_*, M7D_*, M7E_*, docs/reports/M7_WORK_REPORT.md.

12. Research paper

Manuscript draft — not submitted or peer reviewed. Working title: Auditing Verifiable Instruction-Following Benchmarks: Specification Gaps, False Verdicts, and Cross-Benchmark Reliability.

• PDF: paper/main.pdf · source: paper/main.tex · bibliography: paper/references.bib · figures: paper/figures/ · tables: paper/tables/ (appendices A verifier examples, B secondary RL context and C reproducibility are inside main.tex).
• Supporting reports: docs/research/ (taxonomy, oracle coverage, bug validation, literature, claim-evidence matrix), docs/reports/M5C2_FINAL_PUBLICATION_GATE.md, docs/reports/M5C3_PUBLICATION_PREPARATION.md (venue comparison and blockers).

The present recommended direction is a narrower benchmark-evaluation paper, not a confirmed Artificial Intelligence Review publication.

13. Reproducibility and provenance

• Source revisions pinned: NVIDIA reference d2be11c4… (docs/audits/PROVENANCE.md), IFBench 1c40f0c1…, open-instruct 11826255…, IFEval path commit e6890f85…, Qwen3.5-0.8B 2fc06364…, datasets (§10); machine-readable: runs/m5c1/pins.json, docs/audits/environment.json.
• Configurations: each run stores resolved_config.json, method_info.json, environment.json, git_revision.txt; indexed in docs/index/EXPERIMENT_INDEX.md.
• Checkpoints: docs/index/CHECKPOINT_INDEX.md — 92 checkpoint files; every locally present payload hash was re-verified against its manifest (0 mismatches); base weights are never redistributed; 46 payloads are in the repository, 34 tuning/smoke/resume payloads are local-only and 12 intermediate payloads were not preserved (manifests only).
• Seeds: M4 paired 3407, 42, 123, 2026, 777; M5A 3407, 42, 123; dataset split seed 20261014; audit seeds in the scripts (e.g. 20261015). GPU backward passes are not bitwise reproducible, so trajectories are not claimed to reproduce exactly.
• Logs: runs/*/train_metrics.jsonl, reward_metrics.jsonl (every rollout), gpu_metrics.csv, runs/_logs/.
• Tests: 86 CPU tests (make test), including 21 reference-differential tests against the pinned NVIDIA clone.
• GPU-hour accounting: docs/index/GPU_ACCOUNTING.md.
• Integrity: MANIFEST.sha256 (regenerated at each upload and verified by re-download), docs/inventory/FILE_INVENTORY.csv, docs/inventory/SECRET_SCAN.md.
• Historical corrections: docs/technical/RESEARCH_HISTORY.md; original M2/M3 reports carry correction banners.
• Known stale documents (kept, flagged): docs/reports/milestone_log.md (stops at M1), docs/feasibility/compatibility_matrix.md (M0 a-priori assessment), test_logra_manual.py (legacy check).

14. Citation and acknowledgment

Project citation template (placeholders only; no venue, DOI or acceptance exists):

@misc{qwen35_logra_ifbench_audit,
  title  = {Qwen3.5 x LoGRA x IFBench: Gradient-Sketched Reinforcement Learning and Verifier Reliability},
  author = {TODO},
  year   = {TODO},
  note   = {Research archive and unsubmitted manuscript draft; https://huggingface.co/Dhruv1000/qwen35-logra-ifbench (see repository page for current visibility)}
}

Please also cite the directly reused methods and benchmarks:

• LoGRA: S. Zhang, Y. Zhang, J. Hu, Y. Li, H. Zhang, B. Xu, J. Kautz, Y. Dong, LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches, arXiv:2610.06647 (NVIDIA); code port from skzhang1/labs-molt (Apache-2.0).
• IFBench / IF-RLVR: V. Pyatkin et al., Generalizing Verifiable Instruction Following, arXiv:2507.02833.
• IFEval: J. Zhou et al., Instruction-Following Evaluation for Large Language Models, arXiv:2311.07911.
• Qwen3.5 (Qwen team), open-instruct and IFBench code (Allen Institute for AI), PEFT/LoRA, GRPO. Further related work is cited in paper/references.bib and docs/research/M5C2_SYSTEMATIC_LITERATURE_REVIEW.md.

Authorship and affiliations are intentionally left unspecified.

15. License and responsible use

• Project code and documentation: Apache-2.0 (LICENSE). pyproject.toml authorship is a placeholder.
• Upstream method code: the LoGRA port is adapted from NVIDIA labs-molt (Apache-2.0, attribution in logra_core.py); the clone itself is not redistributed. IFBench, open-instruct and instruction_following_eval code are Apache-2.0 and are not copied into this repository (pinned by commit, kept in data/cache); docs/audits/logra_core_PREFIX_d6962d6.py.txt is a snapshot of our own earlier code.
• Model: Qwen3.5-0.8B is Apache-2.0 (Hugging Face model card, checked 14 Oct 2026); weights and tokenizer files are not redistributed. Checkpoints in runs/ contain only method state (adapters / LoGRA master and optimizer state), a derivative of the base model - use them under the base model's licence.
• Datasets: IFEval Apache-2.0; IFBench_test and IF_multi_constraints_upto5 are ODC-BY-1.0 with a note that they include third-party model outputs under separate terms; use for research and education per Ai2's responsible-use guidelines. Run directories contain prompts and model-generated responses derived from these datasets (including generated text that may look like e-mail addresses or passwords; the secret scan found no real credentials).
• Responsible use: results are exploratory; do not cite withdrawn claims; AI-rater labels are not human ground truth; verifier findings are drafts until maintainers respond. No private data, tokens or credentials are included; earlier revisions of this repository were publicly visible (see docs/release/SECURITY_PRIVACY_RELEASE_AUDIT.md).

11d. M8 (evidence-first verifier-testing sprint; branch research/m8-24h-verifier-evidence)

CPU-only follow-up to M6/M7: confirmatory RESERVE panel (240 unseen mutants) and template-disjoint Panel S (spec-first oracles) fixed before running; 24 reconstructed upstream verifier faults; oracle provenance/availability/error ablations; deployment-mode alarm calibration; validated score-impact; an RL qualification gate (NO_GO_GPU, 0 GPU hours). Start at docs/research/m8/M8_MASTER_STATUS.md and docs/reports/M8_FINAL_REPORT.md; manuscript draft paper/m8/main.pdf (not submitted; all labels single-author). Reproduce the primary tables on CPU: make m8-reproduce-core; regenerate the paper numbers/tables and check them: make m8-paper.

11e. M9 (oracle independence, alarm reliability, journal readiness; branch research/m9-oracle-independence-and-validation)

CPU-only. New confirmatory Panel F (24 fresh templates, new specification-first oracles, frozen before any run), budget-matched alarm calibration with verified math, real-fault reanalysis by crash/behaviour and commit cluster, clean-environment reproduction, privacy audit of the public HF repo, blinded review/adjudication packets (unlabelled), manuscript paper/m9/main.pdf (not submitted). Start at docs/research/m9/M9_MASTER_STATUS.md and docs/reports/M9_FINAL_REPORT.md. Clean rebuild: bash scripts/m9_clean_repro.sh (uses requirements-m9-clean.txt).