Non-Human Experimental Inventory

Status: Research analysis artifact — proposal-only, not a canonical accepted record. Author: Claude (autonomous research-director pass), Cowork session, 2026-07-30. Repository baseline: commit 1012c0f on branch as checked out at analysis time. Scope: All located experiments, evidence runs, and evaluation packages whose subject is non-human (AI models, prompts, comparator software, repository artifacts). No human-subject experiments were located or are proposed anywhere in this document.

How to read this document

This is a reconstruction, not a re-statement of each experiment's own README. Fields marked MISSING were not found in the reviewed artifacts and are not assumed to be zero, absent, or negative — only unrecorded. This inventory does not cover 100% of repository files (see Coverage and Known Gaps at the end). It covers the canonical experiment registry (research/framework-engineering/experiments/registry.json), the six research/experiments/ packages, the three research/evidence-runs/ (ECR) packages, the research/framework-engineering/experiments/EX-FE-000{1,2,3} next-generation experiments, and the two major evaluation packages (FE-EVAL-2026-07-23, FE-BOUNDARY-2026-07-28) that themselves function as non-human meta-evaluations.

Two registries already exist in the repository and were treated as ground truth wherever they overlapped with this inventory rather than re-derived from scratch:

• research/framework-engineering/experiments/registry.json (machine-readable, generated 2026-07-24T17:38:48Z, EX-* IDs) — the canonical successor experiment registry.
• research/operating-system/hypothesis-registry.md (H001–H018) — the canonical hypothesis registry.
• research/evaluations/FE-EVAL-2026-07-23/framework-engineering-executive-brief.md — a prior, independent, adversarial repository-wide evaluation (authored by a different agent, "Codex"). This document is evidence, not settled fact, and is treated as such throughout this analysis.

E1 — FE-011A LLM Blind Pilot

• Experiment ID: EX-LEGACY-FE-011A (registry) / FE-011A (native)
• Title: LLM Blind Pilot — structured redesign vs. original framework
• Research question: Do Framework Engineering redesigns improve quality, consistency, evidence use, traceability, and actionability of blinded LLM outputs, relative to the original (un-redesigned) framework packet?
• Hypothesis: Structured FE redesign improves task output quality (informal; not in H001–H018 registry — tracked as native hypothesis FE-011A)
• Subject/system tested: SWOT, Five Whys, OODA Loop framework packets, original vs. FE-redesigned, worked by an LLM
• Experiment type: Same-model comparative pilot (LLM as both subject and — per the repo's own methodology-review — likely evaluator lineage)
• Date conducted: 2026-06-28 (internal pilot)
• Executing agent / model: MISSING exact model identity and version in reviewed files; registry logs runs as historical-manual: 8, reviewer: 3
• Provider: MISSING (not disambiguated by provider in available summary)
• Comparison condition / baseline: Original (un-redesigned) framework packet vs. FE-redesigned packet, same scenario
• Independent variables: Redesign condition (original vs. FE-redesigned)
• Dependent variables: Quality, consistency, evidence use, traceability, actionability (qualitative evaluator judgment, not a validated numeric scale)
• Controlled variables: MISSING — no documented control for evaluator identity/model relative to generator identity/model
• Sample size / trials: 3 frameworks × 2 outputs each = 6 output artifacts, each with an evaluator review
• Evaluation method: Evaluator review documents (evaluator-review.md per framework), qualitative
• Result (as reported): SWOT — Output 1 (redesigned) stronger; Five Whys — Output 2 stronger; OODA — Output 2 stronger; cross-framework pattern of higher traceability/evidence use at the cost of verbosity and process overhead
• Claimed conclusion: "Directional evidence that structured redesigns can improve traceability and evaluation readiness"
• Supported conclusion (this review): Directional and observational only. The pilot does not separate an FE-specific effect from a generic "more structure improves legibility" effect — the pilot's own methodology-review.md-referenced concern, restated in its own summary.
• Confidence claim (native): Not explicitly labeled; treated by research/theory/confidence-assessment.md as supporting "Structured redesign improves traceability" at Low confidence.
• Limitations (native, self-reported): Same-model evaluation bias risk; generic-scaffolding confound; pilot precedes any human validation
• Linked evidence/hypotheses: Feeds research/theory/confidence-assessment.md; referenced as input to FE-EVAL-2026-07-23's incremental-utility framing
• Supersession status: Current / not superseded; explicitly flagged by research/theory/research-program-status.md as needing "different-model replication"
• Raw data availability: Present (results/*/output-1.md, output-2.md, evaluator-review.md, run-log.md)
• Reproducibility status: Not independently reproduced; single internal run

E2 — FE-012A Blind Primitive Reasoning Extraction

• Experiment ID: EX-LEGACY-FE-012A
• Research question: Can a finite set of primitive reasoning operations reconstruct the reasoning flow of diverse knowledge artifacts without domain terminology?
• Subject: 20 non-EDF, non-Clarity knowledge artifacts (framework/methodology corpus)
• Type: Same-model, single-extractor, single-reconstructor internal run
• Date: 2026-06-28
• Executing agent: Single extractor pass EX-01, single reconstruction-analyst pass RA-01, same model family for both roles (per own summary)
• Baseline/comparison: None (no comparator condition; single-arm extraction-then-reconstruction design)
• Sample: 20 artifacts, 1 pass each
• Result: Reconstruction — Strong 9, Moderate 8, Weak 3, Failed 0; recognizable 17/20; 2 candidate primitives surfaced (Prioritize, Compare); vocabulary growth "slowed materially" after the artifact-set midpoint
• Claimed conclusion: "Provisional support for a finite primitive vocabulary with transfer across multiple domains"
• Supported conclusion (this review): Weak provisional support at best. Same extractor and reconstructor, same model family, prior awareness of the research goal, and subjective recognizability judgment together mean the "0 failures" and "17/20 recognizable" figures cannot distinguish a real transferable vocabulary from extractor-side pattern completion toward a vocabulary the extractor was primed to look for.
• Confidence claim (native, in confidence-assessment.md): Low — "partial stabilization and recognizable reconstruction, but independent extractor convergence has not yet been demonstrated"
• Limitations (self-reported): Same-model internal run, not independent/blinded; recognizability judged subjectively
• Raw data: results/extraction-results-2026-06-28.csv, reconstruction-results-2026-06-28.csv, stabilization-table-2026-06-28.csv, primitive-vocabulary-v0.1.csv
• Reproducibility: Not independently reproduced

E3 — FE-012B Primitive Grammar Synthesis Experiment

• Experiment ID: EX-LEGACY-FE-012B
• Research question: Can the FE-012A primitive vocabulary synthesize coherent procedural reasoning frameworks for novel (unseen) problems?
• Subject: 10 novelty problems, no existing framework may be copied
• Type: Same-model synthesis + same-model review (internal run)
• Date: 2026-06-28
• Result: 10/10 problems attempted; 7/10 completed without a new primitive; 3/10 flagged missing-primitive pressure; 0 outright failures; "Primitive Sufficiency Index" = 70% (native, explicitly flagged by its own authors as "experimental and should not be treated as validated"); missing-primitive requests: Synchronize, Schedule, Sequence, concentrated in coordination/phased-maintenance problems
• Claimed conclusion: Vocabulary "appears expressive enough" generically, weaker for temporal-coordination problems
• Supported conclusion (this review): The 70% figure is a same-model, same-designer, non-independently-reviewed proportion over a 10-item, researcher-authored problem set. It should not be read as an expressive-sufficiency rate in any general sense — at n=10 with no independent review, the confidence interval on "70%" is wide enough to be compatible with anywhere from roughly 35% to 93% true sufficiency under a naive binomial estimate, before even accounting for non-independence of design and evaluation.
• Confidence claim (native): Low
• Limitations (self-reported): same-model synthesis and review; novelty problems designed by the same research direction evaluating the result; no independent reviewer agreement measured
• Raw data: results/synthesis-run-2026-06-28.md, reviewer-run-2026-06-28.md
• Reproducibility: Not independently reproduced

E4 — FE-012C Manual Multi-Model Replication

• Experiment ID: EX-LEGACY-FE-012C-MANUAL
• Research question: Do GPT, Claude, and Gemini converge on primitive-grammar extraction when given the same packet instrument and a provided primitive vocabulary?
• Subject: 15 packets × 3 providers (GPT, Claude, Gemini) = 45 manually collected JSON responses
• Type: Cross-model (single point in time) convergence study, human-operated, no API
• Date: Prior to 2026-07 (exact date MISSING in reviewed files; precedes the repeatability run)
• Providers: GPT (unspecified version — MISSING), Claude (unspecified version — MISSING), Gemini (unspecified version — MISSING)
• Independent variable: Provider
• Dependent variables: Entry primitive, exit primitive, dominant primitive, reasoning shape, transition structure, artifact recognition, candidate missing primitives
• Sample: 45/45 valid JSON responses, 0 missing, 0 malformed
• Result: Entry-primitive full agreement 15/15; exit-primitive full agreement 11/15; dominant-primitive full agreement 12/15; reasoning-shape full agreement 9/15; strong transition agreement 14/15; artifact recognized in 15/15 packets (blinding did not hold); 0/15 packets produced a candidate missing primitive (vocabulary was provided in advance, so this is not independent evidence of sufficiency)
• Claimed conclusion: Strong cross-model agreement "strengthens confidence in the extraction instrument"
• Supported conclusion (this review): Convergence is real at the level measured, but confounded by two mechanisms the experiment's own interpretation.md already names: (1) universal artifact recognition (15/15) means agreement may reflect shared training-data familiarity with named frameworks rather than independent structural extraction from blinded text; (2) the provided vocabulary constrains the space of possible answers, mechanically inflating agreement on which of a small fixed set of primitives applies. See Contradiction C1 in the companion contradiction registry regarding this experiment's relationship to E5.
• Confidence claim (native): Not separately labeled; feeds H003, H013, H015 in the hypothesis registry at Moderate/Moderate/Moderate
• Limitations (self-reported, native): "Model agreement does not prove the theory," "disagreement is useful evidence," blinding weakened by recognition
• Raw data: responses/{gpt,claude,gemini}/*.json (45 files), reports/FE-012C-*.md
• Reproducibility status: Directly re-tested by E5 (repeatability run) — see below. This is the one experiment in the corpus with an explicit, dedicated repeatability follow-up.

E5 — FE-012C Repeatability Run 001

• Experiment ID: EX-LEGACY-FE-012C-REPEAT-001
• Research question: Does the same packet instrument, run again in fresh conversations with the same three providers, produce materially similar results over time?
• Type: Same-instrument, same-providers, different point in time (temporal repeatability test), explicitly not a new theory experiment per its own README
• Sample: 45 new responses (15 packets × 3 providers), paired against the original 45 from E4 → 45 paired comparisons, 90 total observations
• Result (literal/syntactic comparator): Stable 1/45 (2.2%); elaboration drift 8/45 (17.8%); structural drift 36/45 (80.0%); backbone-stable pairs 9/45 (20.0%); recognized-artifact changed in 41/45 pairs
• Result (semantic comparator, generous re-scoring layered on top, not a replacement): Stable 6/45 (13.3%); structural drift 38/45 (84.4%) — semantic normalization recovers some elaboration-drift cases into "stable" but does not reduce structural drift; if anything the semantic count of structural drift is nominally higher than the literal count
• Claimed conclusion (native, in the dataset-comparison report itself): "This is a measurement analysis task, not a validation of theory... drift does not by itself falsify the instrument"
• Supported conclusion (this review): This is the single most consequential quantitative result in the corpus and it is currently under-weighted in the hypothesis registry. Under both the literal and the semantic comparator, the large majority of same-provider, same-packet, fresh-conversation pairs show structural (not just wording) drift. This directly bears on H004 (Repeatability), whose kill condition reads: "Same-model outputs drift materially across repeated runs without explainable packet ambiguity." An 80% structural-drift rate is a plausible candidate for "drift materially" — see the confidence reassessment and contradiction registry for the full argument on whether this meets or falls short of the kill condition.
• Confidence claim (native, hypothesis registry, current): H004 = Low-Moderate, status Active (i.e., not yet downgraded to reflect this dataset)
• Limitations (self-reported): 15/45 Dataset-B files required quote normalization before parsing (data-quality, not reasoning, issue); repeatability conclusions remain sensitive to the constrained vocabulary and recognition effects
• Raw data: comparison/generated/FE-012C-dataset-comparison.md, FE-012C-drift-analysis.md (template, not populated at top-experiment level — populated version lives under comparison/, see note below), semantic-repeatability-report.md
• Note on internal inconsistency: experiments/FE-012C-repeatability-run-001/comparison/drift-analysis.md (top-level) still reads as an unpopulated template ("Do not infer repeatability before results exist"), while comparison/generated/FE-012C-dataset-comparison.md contains the fully populated results above. This is a within-experiment documentation-currency defect, recorded as Failure Mode F7.
• Reproducibility status: This experiment is the reproducibility test for E4. It has not itself been re-repeated a third time.

E6 — FE-013 Procedural Language Hypothesis

• Experiment ID: Not yet in the EX-* registry (native ID FE-013)
• Research question: Do procedural frameworks exhibit compiler-like structural properties (entry/exit/preconditions/postconditions/branches/loops/invariants/procedural ASTs) recoverable from blinded natural-language descriptions?
• Status: Designed and preregistered, not executed. responses/{gpt,claude,gemini}/ each contain only a .gitkeep. comparison/run-status.md is an empty table. comparison/observations.md, ast-comparison.md, constraint-comparison.md, invariants.md, unresolved-disagreements.md, preliminary-observations.md are all unpopulated templates carrying explicit "do not infer results before the experiment is run" guards.
• Why this matters for this inventory: FE-013 is listed in CURRENT_STATE.md under "Active Experiments" with a stated purpose, which could be misread as an in-progress or completed study. It is neither — it is a fully specified, zero-data experiment. This is recorded, not treated as a negative result (it is not a null result; it is an unexecuted design).
• Raw data: None
• Reproducibility status: N/A (not yet run once)

E7 — ECR-000001 Multi-Layer Procedural Representation

• Experiment ID: EX-LEGACY-ECR-000001
• Purpose (native): Instrument calibration, not hypothesis testing — explicitly converted to a calibration run per its own README ("determine whether the measurement instrument itself is reliable before using it for broader hypothesis testing")
• Primary hypotheses in scope: H003, H012, H013, H015
• Sample: 12 responses recorded in responses/ (registry: chatgpt: 3, claude: 3, gemini: 3 = 9 native-registry-counted + additional per directory listing = 12 present per file count)
• Result: MISSING a consolidated top-level findings document. instrument-pass-fail.md defines success/failure criteria but was not found populated with an outcome verdict in the files reviewed; execution-log.md was not fully reconstructed in this pass (see Coverage and Known Gaps).
• Claimed conclusion: MISSING (not synthesized into a findings statement discoverable in this pass)
• Supported conclusion: Cannot be assessed from the artifacts reviewed. This experiment functions in the corpus primarily as the calibration precursor to ECR-000002/ECR-000003, and its packet/schema design was reused (with revision) by both.
• Reproducibility status: MISSING
• Flag: This is a Phase-14 evidence gap (missing synthesized findings for a "complete"-status registry entry) rather than a null result — the registry marks status: "complete" with real run data present, but no reconstructed top-level conclusion was located.

E8 — ECR-000002 Recognition Sensitivity

• Experiment ID: EX-LEGACY-ECR-000002
• Purpose: Test whether recovered procedural structure is stable across recognition levels (canonical / paraphrased / structural / graph-only) of the same underlying procedure, across artifact families P001 (reasoning-heavy), P002 (execution-heavy), P003 (static/hierarchical)
• Primary hypotheses: H013, H003, H015; secondary H001, H002, H005–H008, H011, H012, H014, H016
• Sample: Responses present (registry: chatgpt: 2, claude: 2, gemini: 2); comparison/ directory contains only 2 substantive files (p001d-graph-recognition-plan.md, p001c-observation-note.md) plus a .gitkeep — materially thinner than ECR-000001 or ECR-000003
• Result: No consolidated findings document located. This run appears to have been superseded in practice by ECR-000003, which re-scopes and re-executes an expanded version of the same recognition/representation question with a frozen comparator.
• Claimed conclusion: MISSING
• Supported conclusion: This evidence run appears functionally abandoned or absorbed into ECR-000003 rather than formally closed or superseded on paper. No supersession record was found declaring ECR-000002 superseded by ECR-000003, even though ECR-000003's own README describes itself as following on from ECR-000002's recognition-sensitivity question with an expanded design. Recorded as Failure Mode F6 (silent absorption without a supersession record).
• Reproducibility status: MISSING

E9 — ECR-000003 Representation Sensitivity (EXP-001, EXP-002, EXP-003)

• Experiment ID: EX-LEGACY-ECR-000003
• Purpose: Which features carry procedural recognition and structural extraction — topology (EXP-001), surface representation (EXP-002), or domain semantics under matched control-flow (EXP-003)
• Hypotheses in scope: H016, H017, H018, H013, H003, H015, H002
• Instrument: Comparator 3.1.0 (frozen, official measurement) with Comparator 3.2.0 explainability layer (non-scoring, post hoc)
• Sample: EXP-001 12/12 responses, EXP-002 12/12 responses, EXP-003 9/9 responses — 33 total, 0 missing, 0 malformed; tolerant-parsing events 5 / 4 / 3 respectively; leakage findings 3 / 5 / 5 respectively
• Result:
  • EXP-001 (topology perturbation): structural backbone mixed; primitive and constraint comparison disagreement; representation compliance full agreement
  • EXP-002 (cross-representation): structural backbone mixed; primitive/constraint disagreement; compression/elaboration patterns observed but did not erase disagreement
  • EXP-003 (cross-domain isomorphism): structural backbone, literal, conceptual, and dimensional profiles all mostly_stable; primitive sequence, transitions, dominant primitives, and constraints still in disagreement; recognition provider-dependent (GPT persistent, Claude gradual decay, Gemini persistent)
• Claimed conclusion (native, ECR-000003-FINDINGS.md): "Consistent with limited backbone stability under some transformations, persistent provider-specific variation, and continued recognition or domain-language pressure. The run supports cautious methodological claims more strongly than broad theoretical claims."
• Supported conclusion (this review): Concur with the native conclusion as written; it is already appropriately hedged. The one addition this review makes: EXP-003's "mostly_stable" backbone result is the strongest positive structural signal anywhere in the corpus, and it is exactly the result the repository's own next-research-decision process has flagged as needing an adversarial follow-up (negative non-isomorphic controls) before being trusted, precisely because a positive backbone-stability result obtained without a designed negative control cannot rule out "the comparator/rubric is lenient enough to call most things mostly_stable." This review concurs with that internal prioritization.
• Confidence claim: See Hypothesis Evidence Summary table inside ECR-000003-FINDINGS.md — all directions in {mixed, slightly_supported, waiting}, all strengths in {weak, moderate, insufficient}. No hypothesis reaches strong/high anywhere.
• Governance status: A Hypothesis Review Board packet and summary exist (review-board/HYPOTHESIS-REVIEW-BOARD-SUMMARY.md) proposing direction/strength updates to H002, H003, H013, H015, H016, H017, H018 — but HYPOTHESIS-REVIEW-BOARD-DECISION.md is blank (status: blank_human_decision_required). No proposed update has been ratified as of this analysis.
• Raw data: Present and certified (dataset-certification step documented); reports present (ECR-000003-FINDINGS.md, NEXT-RESEARCH-DECISION.md, explainability summaries per experiment)
• Reproducibility status: Not yet independently repeated; recommended next step per the repository's own NEXT-RESEARCH-DECISION.md is adversarial (negative controls), not repeatability

E10 — EX-FE-0001 Framework Engineering Boundary Discrimination Pilot

• Registry status: ready; run_count_by_provider: {manual: 6}
• Research question: Can an evaluator distinguish a non-redundant FE scope from adjacent method engineering / generic structured analysis, using only current canonical claims?
• Implemented vs. stated mismatch: The only located synthesis artifact (synthesis/synthesize--1784832869017.json) has "status": "template", "findings": [], "unresolved": [], "decision": null, and the experiment folder contains fixtures/offline-normalized-result.json and fixtures/offline-manual-response.md. The registry's run_count_by_provider: {manual: 6} most plausibly reflects infrastructure/fixture exercise runs rather than six substantive adjudicated evaluator trials. This inventory records the mismatch rather than resolving it; a synthesized finding was not located.
• Result: No usable finding located.
• Supersession/authority: synthesis_policy explicitly states "No confidence update from model agreement alone."

E11 — EX-FE-0002 Blinded Mechanism Boundary and Subsumption Test

• Registry status: ready; run_count_by_provider: {} (zero runs recorded)
• Research question: Is any FE mechanism non-redundant relative to adjacent disciplines, under blinded adjudication?
• Result: Not executed. Per FE-BOUNDARY-2026-07-28/README.md: "Stage A remains incomplete, Stage B remains blocked, and its experimental boundary classification remains inconclusive."
• Note: A large body of scaffolding exists for this question under research/evaluations/FE-BOUNDARY-2026-07-24/, FE-BOUNDARY-2026-07-24-v1.1/, and FE-BOUNDARY-2026-07-24-v1.2-machine-pilot/ (mechanism cards, reviewer packets, adjudication folders, role registries). These were inventoried by directory listing but not individually reconstructed in this pass — see Coverage and Known Gaps. They appear to be iterative attempts at the same blinded mechanism-boundary test, each superseding protocol details of the last, none reaching a completed Stage B per the 07-28 package's own account.

E12 — EX-FE-0003 Uncertainty and Evidence-Lineage Ablation

• Registry status: proposed, confidence 0.5 (self-declared design confidence, not evidentiary confidence)
• Research question: Does explicit uncertainty representation and traceable evidence lineage improve diagnostic reasoning/decisions vs. an otherwise-equivalent framework?
• Design quality (this review): Notably rigorous as a design — pre-specifies a 27-run minimum matrix (3 trials × providers × cases), blinded scoring, adversarial review, freeze-then-unblind sequencing, explicit falsification conditions, explicit stopping conditions (identity leakage, prompt drift, unequal case information, manual repair before evaluation), and a private identity key (identity-key.private.md) for later unblinding.
• Result: final-analysis.md explicitly states "Status: no execution data." Not run.
• This is recorded as a positive process finding (a well-specified null-state experiment with real blinding infrastructure already built) rather than a gap in this inventory, though it remains a research-backlog item.

E13 — FE-EVAL-2026-07-23 (Repository-Wide Evaluation Package)

• Type: Meta-evaluation / evidence-synthesis artifact, itself produced by a non-human agent ("Codex") reasoning over repository text — included in scope because it is a non-human analytical process operating on non-human artifacts (the repository itself).
• Result headline: Classifies Framework Engineering as "a provisional research and method-engineering program," maturity "research prototype / pre-validation," explicitly states no confidence-assessment statement reaches High, and flags the ECR-000003 board decision as blank/unratified — i.e., independently arrived at several of the same conclusions this analysis also reaches from the primary evidence.
• Use in this analysis: Treated as evidence (a data point about how the repository's claims read to an independent evaluator), not as ground truth. Its own findings are carried into the contradiction and confidence-reassessment sections below with attribution, not silently absorbed as consensus.

E14 — FE-BOUNDARY-2026-07-28 (Successor Boundary/Definition Package)

• Type: Successor meta-evaluation, produced 5 days after E13, incorporating externally sourced comparisons (systems engineering, method engineering, design science, requirements/knowledge/ontology engineering, decision analysis, quality engineering, organizational learning, cybernetics, multi-agent systems engineering)
• Result headline: "The evidence does not support presenting FE as a distinct engineering discipline... Framework Engineering is a repository-centered research program and a candidate integrated engineering profile," and explicitly designs (but does not execute) the matched-utility experiment (Arms A/B/C/D) described in next-experiment.md.
• Status note it makes about itself: "This is a proposed successor finding. It does not overwrite the controlling EX-FE-0002 state."
• Use in this analysis: Same treatment as E13 — evidence, cited with attribution, its proposed next experiment folded into the ranked backlog in non-human-next-experiments.md rather than duplicated.

Summary Table — Execution Status

Coverage and Known Gaps in This Inventory

This pass reconstructed 14 experiment/evaluation units from primary source documents (READMEs, protocol files, results/reports, response counts, registry.json, hypothesis-registry.md, confidence-assessment.md). It did not individually reconstruct, and marks as an open gap for the next research pass:

• The internal contents of research/evaluations/FE-BOUNDARY-2026-07-24/, -v1.1/, and -v1.2-machine-pilot/ beyond directory-level listing (mechanism cards, individual reviewer packets, adjudication records, source registries). These plausibly contain 50+ additional individually inventoriable sub-artifacts.
• Individual Evidence Decision Records (EDRs) under research/evidence-runs/*/edr/ and research/operating-system/edr/.
• The 100-artifact reference corpus (research/corpus/artifacts/, research/corpus/batches/) as individual characterization records — only aggregate statistics.md was reviewed.
• research/papers/, input-documents/, and research/frontier/ were not opened.
• ECR-000001/execution-log.md and ECR-000003's full pipeline/orchestration logs were listed but not read line-by-line.

None of the conclusions in the companion analysis documents depend on these unread materials; where a claim would require them, it is marked MISSING rather than inferred.

Repository Validation State (recorded, not repaired)

• npm run experiments:verify -- --all passed, verifying all 10 registry entries (EX-FE-0001, EX-FE-0002, EX-LEGACY-ECR-000001/2/3, EX-LEGACY-FE-011A/012A/012B/012C-MANUAL/012C-REPEAT-001).
• npm run research:validate failed with EMFILE: too many open files while parsing research/framework-engineering/experiments/EX-FE-0003/*, even after raising the shell's file-descriptor limit to 524288. This appears to be an unbounded-concurrency defect in the research-publisher build step (attempts to open >1000 files via Promise.all without a concurrency cap) rather than a content defect in any research record. Recorded as a research-system finding (see failure-mode taxonomy F8); not fixed as part of this analysis, per the instruction not to silently repair tooling encountered mid-investigation.
• Git state at analysis time: HEAD 1012c0f, working tree clean except an untracked .claude/ directory (local agent configuration, not a repository content change).