Agent Skills for Context Engineering epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is how acquired research, evaluation results and revised instructions acquire reliance and affect later behavior. Scope and exclusions are see boundary; source identity and layers are see its Source register. This is a complete artifact, partial loop, not an enclosing runtime. All source findings below use SRC-1 at the frozen revision. Direct inspection adds implementation and doctrine/design evidence, not observed operation or causal experiments. No target program was executed.

The consequential warrant claims are see RT-CLM-maintenance-guarantees, EPI-CLM-research-protocol and EPI-CLM-trace-analysis. These claims concern different consumers. Source suitability, lexical overlap, fixture agreement and model-assigned trace quality must not be treated as one evaluator of truth.

Entry point or operation	Source path, SRC-1	Coverage and record	Limit and conclusion prevented
Skill discovery and activation	.plugin/plugin.json, .claude-plugin/marketplace.json, skills/context-fundamentals/SKILL.md	Supplied RT-RTE-skill-delivery	External loader and compliance uninspected; no demonstrated behavioral effect.
Scheduled discovery, retrieval and parking	researcher/scripts/loop_step.py	Supplied RT-RTE-research-step	Actual fetched bytes and review actions not supplied.
Manual research CLI, including alternate dispatch	researcher/scripts/research_loop.py	Direct: retrieval through promotion and complete parser/dispatch; RT-RTE-mechanism-promotion, EPI-RTE-novelty-check	Initialization helpers only supplied through runtime; record-state labels are not semantic reviews.
Source rubric and publication-readiness checks	researcher/rubrics/content-curation.md, researcher/runbooks/autonomous-research-loop.md, researcher/scripts/validate_run.py, researcher/scripts/validate_repo.py	Direct: full rubric/runbook/run validator; repository source-evaluation validation functions; EPI-RTE-source-review, EPI-RTE-rubric-arithmetic, EPI-RTE-run-readiness	Remaining repository validator domains not independently traced; arithmetic consistency does not establish evidence truth.
Novelty comparison and report retention	researcher/scripts/novelty_check.py, researcher/scripts/research_loop.py	Direct full novelty implementation and report consumer; EPI-RTE-novelty-check	No semantic novelty oracle.
Router model calls, exact-match scoring, result reuse	researcher/benchmarks/sdk-runner/src/runRouter.ts, researcher/benchmarks/sdk-runner/src/common.ts	Supplied RT-RTE-router-benchmark	No supplied item-level run evidence or independent verification of reference labels.
Effectiveness execution	researcher/benchmarks/sdk-runner/src/runEffectiveness.ts	Supplied RT-ABS-effectiveness-executor	Absence is limited to this runner, not every evaluation in the artifact.
Prompt optimization, stop/selection, artifact and skill output	examples/interleaved-thinking/reasoning_trace_optimizer/loop.py, examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py	Direct full analyzer, loop capture-to-finalization and scoring/convergence; other internals supplied RT-RTE-prompt-optimization; new EPI-RTE-trace-analysis	No independently judged task outcome or qualifying produced theory inspected.
Retained idea/bookmark/post consumption	examples/digital-brain-skill/agents/scripts/idea_to_draft.py	Direct full script; RT-RTE-idea-draft	Its upstream write/review process and downstream publication untraced.
CI dispatch	.github/workflows/validate.yml	Direct full workflow; EPI-CLM-research-protocol	Declared check invocations, not supplied successful CI results or required-branch protection.
Other exposed evaluators, export and artifact operations	researcher/scripts/compare_skill_revisions.py, researcher/scripts/skill_health.py, researcher/scripts/check_activation_cases.py, researcher/scripts/run_benchmarks.py, researcher/scripts/validate_export.py, researcher/scripts/artifact_store.py	Paths identified, selected symbol searches and CI entry calls only; semantic implementations unassessed	Prevents a complete account of selection, export lineage, freezing and rejection across the researcher plane.
Other published guidance and helpers	skills/, template/	Included but not individually assessed beyond supplied skill-delivery witness	Prevents corpus-wide claims about evaluation, memory revision, belief checks or policy adaptation.
Other shipped examples	examples/book-sft-pipeline/, examples/llm-as-judge-skills/, examples/long-horizon-prompt-lab/, examples/x-to-book-system/	Available but uninspected; CI names the prompt-lab build/validation	Prevents system-complete negatives about learning, training, acceptance and judging.
Governance, claims, schemas and retained research instances	governance/, researcher/claims/, researcher/schemas/, researcher/runs/, researcher/exports/	Protocol references and selected consumers inspected, not complete data/lineage audits	No individual accepted research claim, authenticated reviewer, privacy invariant or observed lifecycle integration established.

External research originals, host internals, model parameters, private runtime state and provider execution are not supplied. These gaps prevent external-source verification, distributed-parametric learning conclusions and observed benefit attribution. Available but uninspected paths above are not access failures or evidenced absences. The coverage does not license a system-complete negative.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-skill-body	Advice includes empirical assertions about attention and loading; instructions themselves prescribe behavior	see record	SRC-1, skills/context-fundamentals/SKILL.md	Delivery is not verification of its empirical assertions.
EPI-OBJ-source-evaluation	Source suitability, gate judgments, mechanism summaries and evidential strength	see record	SRC-1, researcher/rubrics/content-curation.md	No particular completed evaluation inspected.
EPI-OBJ-mechanism-proposal	Activation conditions, predicted behavior changes and failure modes; status/reviewer fields are symbolic control data	see record	SRC-1, researcher/scripts/research_loop.py	Registry projection carries evidence IDs, not checked evidence semantics.
EPI-OBJ-novelty-result	Computed overlap values over selected tokens; semantic novelty remains a separate judgment	see record	SRC-1, researcher/scripts/novelty_check.py	Lexical statistics are truth-apt about the algorithm's inputs, not proof of novelty.
EPI-OBJ-trace-analysis	Claimed failures, quoted trace evidence, quality scores and recommendations	see record	SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py	Output can contain fallback scores rather than model judgments.
RT-RTE-router-benchmark	Ranking and exact agreement with fixture primary labels	see route for embedded result objects	SRC-1, researcher/benchmarks/sdk-runner/src/runRouter.ts	No new object declared for the same supplied result container.
RT-RTE-prompt-optimization	Revised prompts/skills mix policies with possible explanatory assertions; individual truth-apt units not determinable without outputs	see route for prompt/artifact objects	SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/loop.py	No inspected candidate output proves novel explanation or accepted knowledge.
RT-RTE-idea-draft	Imported idea, source link and first bookmark insight, where present	see route for retained input objects	SRC-1, examples/digital-brain-skill/agents/scripts/idea_to_draft.py	File content can be selected into “Supporting Evidence” without evidential assessment.
RT-CMP-cursor-model, RT-CMP-minimax-model	Parameters not individually accessible	see components	SRC-1, researcher/benchmarks/sdk-runner/src/runRouter.ts, examples/interleaved-thinking/reasoning_trace_optimizer/loop.py	No local weight update inspected; mutable provider identities do not pin exact weights.

Authority-route ledger

Fields owned by a declaring record are given by linked references rather than duplicated. Multiple rows may reference the same valid route container while separating its consequential functions. “Implemented” below concerns caller wiring, not reliable model compliance or observed execution.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-research-step	content transformation	implemented	EPI-OBJ-source-evaluation	truth-apt transformation: acquisition/import	External source to raw file/pointer	see route; byte cap, HTTP(S) and caller input, not source truth	see route	Retrieved/parked state; see route	Local file/state writes	Source warrant unknown; retrieval label grants no truth license	Later selection may advance or park work	see route	SRC-1, researcher/scripts/loop_step.py	none	Manual retrieve permits no file	Byte completeness and original source authenticity not established.
EPI-RTE-source-review	check/evidence production	doctrine only	EPI-OBJ-source-evaluation	truth-apt transformation: indeterminate	Source suitability and mechanism evidence	see record	see record	see record	Prescribed review	Suitability for research, not every claim's truth	Reviewer supplies proposal/rejection	see record	SRC-1, researcher/rubrics/content-curation.md	EPI-CLM-research-protocol	none	Human/judge execution unobserved.
EPI-RTE-source-review	disposition/acceptance	doctrine only	EPI-OBJ-source-evaluation	no content change	Admission to research pipeline	see record	After evidence review, as prescribed	APPROVE/HUMAN_REVIEW/REJECT; see record	Advisory protocol until external actor acts	Named criterion/intended use exist for source admission, not universal truth	Reject failed gates; prepare change only after review	see record	SRC-1, researcher/runbooks/autonomous-research-loop.md	EPI-CLM-research-protocol	none	No observed evidence-consuming acceptance decision inspected.
EPI-RTE-rubric-arithmetic	check/evidence production	implemented	EPI-OBJ-source-evaluation	truth-apt transformation: entailed derivation	Recorded gate/score/decision consistency	see record	see record	Expected verdict and mismatch errors	Error report/nonzero exit	Arithmetic and rubric-rule consistency only	Can fail repository validation	see record	SRC-1, researcher/scripts/validate_repo.py	EPI-CLM-research-protocol	none	Does not establish accurate scores or honest evidence.
EPI-RTE-novelty-check	check/evidence production	implemented	EPI-OBJ-novelty-result	truth-apt transformation: entailed derivation	Selected-token overlap with registry/corpus	see record	see record	pass/human_review/likely_duplicate	Return code and report	Lexical overlap only	Non-pass needs rationale at active readiness	see record	SRC-1, researcher/scripts/novelty_check.py	EPI-CLM-research-protocol	none	Semantic novelty needs separate judgment.
EPI-RTE-novelty-check	retention	implemented	EPI-OBJ-novelty-result	no content change	Report for subsequent readiness	see record	CLI novelty command	Stored report plus optional rationale	Persists caller-readable result	No promotion to semantic acceptance	Later validator can continue based on verdict/rationale	see record	SRC-1, researcher/scripts/research_loop.py	none	none	State advances even when checker returns non-pass.
EPI-RTE-run-readiness	check/evidence production	implemented	EPI-OBJ-mechanism-proposal	truth-apt transformation: entailed derivation	Active-run required artifacts or closed-run state	see record	see record	ok/errors	Promotion consumes predicate	Structural readiness in checked branch only	Default accepted/candidate branch can be blocked	see record	SRC-1, researcher/scripts/validate_run.py	RT-CLM-maintenance-guarantees	Closed-state/override exceptions	No claim-truth or authenticated-review check.
RT-RTE-mechanism-promotion	disposition/acceptance	implemented	EPI-OBJ-mechanism-proposal	no content change	Caller recommendation and mechanism ID	see route	see route	Registry admission or rejected ledger	Symbolic append/exception	Records a recommendation; semantic acceptance not established by CLI	Allows accepted and candidate entries; blocks some invalid/duplicate inputs	see route	SRC-1, researcher/scripts/research_loop.py	RT-CLM-maintenance-guarantees	Candidate and accepted share append branch	Review string/rationale need not demonstrate evidence-consuming decision.
RT-RTE-mechanism-promotion	retention	implemented	EPI-OBJ-mechanism-proposal	truth-apt transformation: non-ampliative reshaping	Projection to registry and ledgers	see route	see route	Registry fields plus separate reviewer/rationale event	File append	Evidence links retained, warrant not independently strengthened	Registry becomes input to later novelty checking	see route; deterministic novelty reader also see EPI-RTE-novelty-check	SRC-1, researcher/scripts/research_loop.py	EPI-CLM-research-protocol	none	Retention before demonstrated acceptance is not lifecycle integration.
RT-RTE-router-benchmark	check/evidence production	implemented	see route	truth-apt transformation: entailed derivation	Predicted top-1/top-3 against primary fixture label	see route	see route	Correct flags, aggregate accuracy	Scoring/report writes	Agreement for the fixture and supplied descriptions	No automatic publication or skill activation	see route	SRC-1, researcher/benchmarks/sdk-runner/src/runRouter.ts	none	none	Reference labels and resume identity bound inference; not task correctness.
EPI-RTE-trace-analysis	check/evidence production	implemented	EPI-OBJ-trace-analysis	truth-apt transformation: indeterminate	Trace failures and quality	see record	see record	Parsed model judgment or fallback	Output feeds optimizer and scores	Candidate diagnosis, no calibrated truth license	Can change next prompt and stopping	see record	SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py	EPI-CLM-trace-analysis	Truncated tool results/fallback	Requested diagnoses not evidence they occurred correctly.
RT-RTE-prompt-optimization	content transformation	implemented	see route	non-truth-apt policy/content update: rewritten system instructions; embedded truth-apt assertions remain indeterminate	Trace/analysis to prompt and optional skill	see route	see route	Revised instruction candidate	Caller/model proposal; length guard	No independent acceptance of factual assertions	Next round receives current_prompt	see route	SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/loop.py	none	none	Specific non-entailment not established; no blanket conjecture classification.
RT-RTE-prompt-optimization	operational admission/selection/consumption	implemented	see route	no content change	Continue/stop, best/final prompt	see route	see route	Score-based continuation or selected prompt	Enforcing orchestration, not epistemic endorsement	Task-loop proxy score only	Stops calls or supplies changed system text	see route	SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/loop.py	none	Score can be attached to successor not tested input	Final score is not independent final-prompt validation.
RT-RTE-prompt-optimization	behavior/policy adaptation	implemented	see route	non-truth-apt policy/content update: trace-fed prompt replacement	Prompt-controlled agent behavior	see route	On continued round	Updated current_prompt, optional durable artifacts/skill	System-prompt channel on next capture	No factual warrant added by prompt authority	Later capture policy can change	see route	SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/loop.py	none	none	Live execution and improvement unobserved.
RT-RTE-idea-draft	content transformation	implemented	see route	truth-apt transformation: non-ampliative reshaping	Retained content to scaffold	see route	see route	Source links and first insights printed	Formatting and selection	Original source warrant unknown; evidence heading is not evaluation	Caller gets scaffold, not published claim	see route	SRC-1, examples/digital-brain-skill/agents/scripts/idea_to_draft.py	none	none	Omitted insights and no freshness check limit completeness.
RT-RTE-skill-delivery	operational admission/selection/consumption	doctrine only	RT-OBJ-skill-body	no content change	Task-relevant instruction loading	see route	see route	Intended activated body	see RT-BAP-skill-instruction	Delivery grants no evidential warrant	Intended guidance during host session	see behavioral-authority record	SRC-1, README.md	none	none	Manifest packaging is implemented; host loading itself uninspected.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
RT-CLM-maintenance-guarantees	see record	SRC-1, README.md, doctrine/design	see record	see RT-RTE-mechanism-promotion, EPI-RTE-run-readiness	Uninspected	No execution or comparison; see record for supported forcing cases	Default protocol, not authenticated human review or universal readiness invariant	Override and closed-state branches; paid callers exist outside conservative step driver.
EPI-CLM-research-protocol	see record	SRC-1, researcher/runbooks/autonomous-research-loop.md, doctrine/design	Source evidence, failed-gate rejection, semantic novelty and explicit approval prescribed	see EPI-RTE-rubric-arithmetic, EPI-RTE-novelty-check, EPI-RTE-run-readiness, RT-RTE-mechanism-promotion	No reviewed research instance inspected	Rule arithmetic and lexical metrics are not independent research replication	Protocol distinguishes evidential review from deterministic gates	Source review remains external; CLI state progression is not mandatory semantic review.
EPI-CLM-trace-analysis	see record	SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py, doctrine/design and implementation	Explicit demand for concrete evidence and quality scores	see EPI-RTE-trace-analysis, RT-RTE-prompt-optimization	No analysis instance supplied	No independent oracle; sampled/model-visible trace is not proof of internal causal reasoning	A model is asked to diagnose exposed traces; output can guide prompt revision	“FULL” claim does not include full tool-result payloads; parse fallback can supply a neutral score.

Bounded conclusion

Acquisition and source review. The scheduler can capture source bytes, while the manual CLI can record retrieval without bytes. Source warrant remains unknown. The rubric supplies criteria and intended use for source admission into the research pipeline: mechanism specificity, implementable artifacts, advanced relevance and identifiable source, followed by weighted scoring and overrides. This is not a criterion for accepting every proposition as true. A human or judge must apply it; the scheduler parks that work. No particular evidence-consuming acceptance decision or post-acceptance lifecycle integration was inspected. See RT-RTE-research-step, EPI-RTE-source-review and EPI-RTE-run-readiness.

Deterministic derivations and their consumers. Rule arithmetic derives a rubric verdict from caller-entered gates and scores; novelty derives overlap and threshold labels from selected tokens; router scoring derives fixture agreement. These are checked implementation-domain deductions, not independent confirmation of source truth, semantic novelty or task correctness. The concrete deductions in EPI-RTE-rubric-arithmetic and EPI-RTE-novelty-check establish this without executing the target. Active readiness consumes novelty's verdict/rationale; default promotion consumes run readiness. The supported override and closed-state routes narrow their enforcing authority. See also RT-RTE-router-benchmark.

Retention is not acceptance. Promotion persists mechanism fields/evidence IDs and a separate review event. Both accepted and candidate statuses enter the registry. Subsequent novelty code reads both and uses their text in overlap scoring. That gives retained records a real selection role, not a new evidential license. Rejected recommendations are retained in a rejected ledger; the inspected novelty loader reads the registry, not that ledger, so preservation of rejection does not itself implement avoidance of rediscovery. Rollback or withdrawal from the registry is not supplied by the inspected promotion function. No broader absence is asserted. See RT-RTE-mechanism-promotion and EPI-RTE-novelty-check.

Trace-fed authority is operational. The analyzer asks for trace-grounded diagnoses, but its interpreted output and fallback numbers are not calibrated evidence of correctness. The optimization loop nevertheless consumes scores and recommendations, revises instructions and supplies them as the next round's system prompt. This implements behavior-shaping authority without a separate factual-acceptance decision. It can retain artifacts and emit a skill, but neither output label nor final score proves final-prompt improvement. Specific produced truth-apt theories, their entailment/non-entailment, claim-specific criticism and resulting acceptance are not determinable from inspected requests alone. See EPI-RTE-trace-analysis and RT-RTE-prompt-optimization.

For the prompt-controlled optimizer boundary and within-run horizon, evidence-responsive instruction replacement is wired and a next-capture authority path exists. Actual exercise and causal dependence of subsequent behavior on that replacement are unobserved, preventing exercised self-improvement attribution; successful improvement is separately unestablished. The runtime's independent theory-builder conditions remain see RT-RTE-prompt-optimization: localized produced theory and content-specific criticism uninspected, consumption afforded, qualifying iteration uninspected. Reflection is afforded by trace representation feeding prompt revision; fidelity and exercise remain unobserved. Learning is uninspected because no capacity gain is demonstrated. Autonomous orchestration is not evidence of autonomous theory building. Distributed-parametric parameter change remains uninspected at the provider; this does not negate the text-update route. See RT-CMP-minimax-model.

Other reuse remains bounded. The digital-brain consumer selects and prints retained insights under an evidence heading, preserving some text/URLs but checking neither truth nor freshness. Static skill delivery supplies behavioral guidance if an external host activates it, not proof that the guidance is warranted or followed. Available uninspected evaluators, training examples, export machinery and skill helpers prevent extending these route conclusions into a system-wide negative or a single epistemic verdict. See RT-RTE-idea-draft, RT-RTE-skill-delivery, and the coverage table.

Shared records

Operative objects

EPI-OBJ-source-evaluation

Label: Source evaluation

Source-native identity: completed sources/evaluations/*.json records. Natural-language gate evidence, mechanism descriptions and limitations are carried in symbolic JSON fields consumed by deterministic checks. Storage: run-local files. SRC-1, researcher/scripts/validate_run.py, researcher/rubrics/content-curation.md. No counterpart object is declared in supplied runtime; RT-RTE-research-step parks the work, rather than owning the completed evaluation's semantics.

Every approved or reviewed source must include: --- researcher/rubrics/content-curation.md

Specific quoted or paraphrased evidence for every passing gate. --- researcher/rubrics/content-curation.md

EPI-OBJ-mechanism-proposal

Label: Mechanism proposal

Source-native identity: run-local proposals/mechanism-proposal.jsonl and its projected fields. Form: natural-language activation scenario, behavior change, evidence references and failure modes within symbolic status/ID structure. Storage: JSONL files; accepted/candidate projection goes to the registry, review rationale to a ledger. This is the material candidate object traversing RT-RTE-mechanism-promotion, not a duplicate route or new source of approval. SRC-1, researcher/scripts/research_loop.py.

"activation_scenario": proposal.get("activation_scenario", ""), "behavior_change": proposal.get("behavior_change", ""), "evidence": proposal.get("evidence_claim_ids", []), "failure_modes": proposal.get("failure_modes", []), --- researcher/scripts/research_loop.py

EPI-OBJ-novelty-result

Label: Novelty result

Source-native identity: reports/novelty-result.json, with verdict, threshold, maximum scores, shared terms and optional human rationale. Symbolic numeric and label fields plus natural-language token/rationale strings on disk. No counterpart result object is declared in runtime; it is distinct from mechanism-admission outputs in RT-RTE-mechanism-promotion. SRC-1, researcher/scripts/novelty_check.py, researcher/scripts/research_loop.py.

"top_mechanism_overlaps": mechanism_overlaps[:10], "max_corpus_score": max_corpus_score, "top_overlaps": overlaps[:10], --- researcher/scripts/novelty_check.py

EPI-OBJ-trace-analysis

Label: Trace analysis

Source-native identity: AnalysisResult returned by TraceAnalyzer.analyze. Form: natural-language diagnoses, quoted evidence and recommendations within symbolic score/type/severity fields. Storage: in-memory result; formatted text can be persisted by the optimization container. Different from the provider parameters in RT-CMP-minimax-model; supplied RT-RTE-prompt-optimization names the container but declares no separate analysis object. SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py.

result = self._parse_analysis_response(analysis_text, trace.session_id) result.analyzer_thinking = analyzer_thinking result.analyzer_model = self.model --- examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py

Routes

EPI-RTE-source-review

Label: Source evidence review

The external human or judge agent is prescribed to read retrieved evidence, apply content-curation gates/scoring and record a source decision. This is distinct from the deterministic scheduler RT-RTE-research-step and promotion append RT-RTE-mechanism-promotion. Endpoints: source and rubric to completed evaluation, then proposal/rejection. The rubric is a source-suitability procedure, not a proof that every imported assertion is true. SRC-1, researcher/rubrics/content-curation.md, researcher/runbooks/autonomous-research-loop.md.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: uninspected — external evaluator execution is not supplied.
• Later read-back: completed evaluation can be consumed by evaluate/readiness; actual later review of rationale is unobserved.
• Delegated visibility: run-local evaluation/proposal files are available to subsequent agents/reviewers.
• Selection predicate: all gates, weighted score thresholds and O1–O4 overrides; failed gates reject, with human override described by rubric introduction.
• Invalidation or expiry: source refresh and conflicting-claim handling are instructed, not an automatic expiry implementation here.
• Activation or effect: prescribed rejection, human review or source approval for research/proposal work.
• Evidence limits: rubric and runbook do not establish executed evidence review, independent replication or individual claim acceptance.

Decision roles: human/judge supplies gate and dimensional judgments and decides disposition; failed gates prescribe rejection. Human override is stated. Mode: open research candidates, no reference-answer oracle. Trigger: retrieved research candidate. Proposed product change: mechanism extraction and skill proposal. Admission to published product remains a later instructed review/PR step. Preservation of rejected rationale and conflicting claims is prescribed; transactional product rollback is not supplied by this review instruction.

If any gate fails, record REJECT with the failed gate and stop. --- researcher/rubrics/content-curation.md

| Conflicting sources | Record both claims and prefer no published change until resolved | --- researcher/runbooks/autonomous-research-loop.md

EPI-RTE-rubric-arithmetic

Label: Rubric arithmetic validation

Endpoints: recorded evaluation booleans/numbers to recomputed total, expected rubric verdict/override and validation errors. Owner: Validator.validate_source_eval_shape and expected_content_decision; context is selected non-draft JSON with required evaluation keys in runs/fixtures. This checks the consistency of supplied judgments, not their evidential accuracy. Distinct from RT-RTE-mechanism-promotion: it runs in repository validation, whereas promotion's readiness subprocess calls the run validator. SRC-1, researcher/scripts/validate_repo.py, researcher/scripts/research_loop.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: accumulated validation findings; research CLI can store repository validation reports and forward exit status.
• Later read-back: uninspected — no subsequent semantic consumer of the stored repository report was traced.
• Delegated visibility: report/stdout and repository CI check are visible to caller/reviewer; deployment enforcement is external.
• Selection predicate: non-draft JSON with required source-evaluation keys under runs/fixtures; rule mismatches produce errors.
• Invalidation or expiry: recalculated from current recorded inputs; no assurance those inputs reflect current sources.
• Activation or effect: validation errors and nonzero return can block a compliant CLI/CI workflow.
• Evidence limits: no observed run; the arithmetic license assumes the encoding and recorded inputs, not truth of gate explanations.

Decision roles: evaluator/human authors scores; code calculates rule consequences; reviewer acts on report. Answer oracle: rubric rules, not empirical ground truth. Mode: deterministic consistency check. A concrete implementation deduction: with all gates true and D1=D2=D3=D4=2, the total is 2.0, no O1–O4 branch applies, and expected_content_decision returns APPROVE. The same recorded inputs can be supplied without sound supporting evidence; arithmetic alone cannot detect that semantic defect. This is a code deduction, not target execution.

if total >= 1.4: return {"verdict": "APPROVE", "override_triggered": None} --- researcher/scripts/validate_repo.py

if abs(float(recorded_total) - recomputed_total) > 0.011: --- researcher/scripts/validate_repo.py

EPI-RTE-novelty-check

Label: Novelty overlap check

Endpoints: selected proposal sections and registry/corpus text to token sets, Jaccard overlaps, threshold verdict and persisted report. Owner: Python checker, then research CLI recorder and run-readiness consumer. No counterpart check route in supplied runtime; RT-RTE-mechanism-promotion supplies registry writes but is not this comparator. SRC-1, researcher/scripts/novelty_check.py, researcher/scripts/research_loop.py, researcher/scripts/validate_run.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: JSON/text report; exit zero for pass and two otherwise. Wrapper forwards code but records novelty_checked regardless.
• Later read-back: active run validator reloads report and requires human_review_rationale for non-pass; future novelty invocations reload accepted/candidate registry entries.
• Delegated visibility: report and shared terms available to human/judge; top lists limited to ten entries, shared terms to thirty.
• Selection predicate: structured proposal sections or salient-text fallback; accepted/candidate registry entries; skill, fixture and run Markdown, with specified exclusions.
• Invalidation or expiry: current comparator inputs are read on invocation; saved report is not automatically recomputed by readiness.
• Activation or effect: verdict/rationale controls active readiness, not automatic semantic duplicate rejection or publication.
• Evidence limits: lexical similarity can miss paraphrase or flag shared vocabulary; no observed semantic review supplied.

Decision roles: caller authors proposal and optional rationale; code derives overlap; human/LLM semantic novelty review is prescribed separately. No answer oracle beyond comparator rules. Mode: open research proposals against a finite current corpus. The check does not admit changes itself or roll them back.

Concrete text transformation: proposal text ## Mechanism\nalpha beta yields tokens {alpha, beta}. A comparator text alpha gamma yields {alpha, gamma}; Jaccard is 1/3, rounded to 0.3333. If that is the maximum at default threshold 0.18, it exceeds both 0.18 and 0.315, so verdict is likely_duplicate. This follows inspected token extraction, rounding and branches; it says nothing about whether the two mechanisms are semantically duplicates. No target execution is needed.

return len(a & b) / len(a | b) --- researcher/scripts/novelty_check.py

if max_score >= args.threshold * 1.75: verdict = "likely_duplicate" --- researcher/scripts/novelty_check.py

if verdict != "pass" and not data.get("human_review_rationale"): self.error(path, "non-pass novelty verdict requires human_review_rationale") --- researcher/scripts/validate_run.py

EPI-RTE-run-readiness

Label: Run readiness check Part of: RT-RTE-mechanism-promotion

This predicate is the check portion of promotion, not its disposition or append. It is also exposed through validate-run. Endpoints: run files to ok/findings/exit status, consumed by run_readiness_ok. Owner: deterministic validator. Active mode checks state, queue, evaluation status, proposal placeholders, novelty rationale and PR headings. Closed mode skips those active checks. SRC-1, researcher/scripts/validate_run.py, researcher/scripts/research_loop.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: JSON/text ok report and process exit status; wrapper can save reports and set validated state on success.
• Later read-back: report storage is wired; promotion recomputes the predicate rather than trusting a saved report.
• Delegated visibility: findings available to caller/reviewer through stdout or stored reports.
• Selection predicate: current_state selects active versus closed checks; first sorted completed evaluation supplies source status.
• Invalidation or expiry: recalculated at promotion; no automatic freshness assessment of underlying sources.
• Activation or effect: failed readiness blocks default accepted/candidate promotion; allow-unready bypasses it.
• Evidence limits: declared retrieval status, headings and rationale presence do not authenticate source bytes, approval or claim truth.

Decision roles and override/recovery are see the parent route. No independent answer oracle. For active runs the evidence-consuming computation checks recorded fields, not the cited research. The closed-state branch plus supported close CLI confirms the runtime's default-path limitation. The closure validator returns when the file is absent; it does not require a closure file even though the runtime describes closed runs as checked through closure. This refines, rather than overturns, its less-than-active-readiness conclusion. No supersession is needed for the valid promotion container.

def validate_closure(self) -> None: path = self.run_dir / "reports" / "closure.json" if not path.exists(): return --- researcher/scripts/validate_run.py

if current_state == "closed": --- researcher/scripts/validate_run.py

EPI-RTE-trace-analysis

Label: Trace analysis check Part of: RT-RTE-prompt-optimization

The analysis call is a material part with different target and epistemic license from capture, rewriting and successor selection. Endpoints: exposed task/system prompt/thinking blocks, tool-call summaries and final outcome to model diagnosis, parsed patterns/scores/recommendations. Owner: TraceAnalyzer requests RT-CMP-minimax-model judgment; code parses/falls back. SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: AnalysisResult from model text or fallback parser; quick_score separately returns clamped numeric result or 50.
• Later read-back: inapplicable — this subroute returns current-iteration analysis; parent owns durable artifact storage/reload.
• Delegated visibility: separate analyzer call receives thinking text, task/prompt and tool-call summaries with at most 200 characters per tool result.
• Selection predicate: prompt asks for failure-pattern evidence and quality scores; invalid pattern enums/keys can be skipped, parse failure triggers fallback.
• Invalidation or expiry: each analysis is tied to trace session ID; no independent source verification or expiry check.
• Activation or effect: parent uses overall score in continuation/ranking and analysis in prompt rewriting.
• Evidence limits: no observed diagnosis; input/result truncation and synthetic fallback weaken any score-as-measurement interpretation.

Target/domain: exposed trace behavior, not full hidden model computation, final-answer correctness or transfer. Decision roles: model proposes diagnoses/scores; parser decides usable fields; loop decides continuation. No reference-answer oracle supplied. Mode: one trace or batch of traces; optimization parent repeatedly uses one caller task. Theory-specific criticism is requested only to the extent failure descriptions/evidence/suggestions might instantiate it; no particular claim/blame/revision unit was inspected. The parent's independent theory/learning statuses are unchanged.

Concrete fallback deduction: response text unparseable fails JSON parsing, matches none of the fallback numeric patterns, and produces overall_score 50 with parse-failure recommendations. This number can enter weighted iteration scoring, although it is not an evidence-based quality judgment. A different supported branch parses valid JSON scores directly; therefore not every score is synthetic. The distinction must remain local.

if result.overall_score == 0: result.overall_score = 50 # Neutral default instead of 0 --- examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py

result = self._fallback_parse_analysis(response_text, trace_id, str(e)) --- examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py

f" Result: {tc.result[:200] if tc.result else 'None'}..." --- examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py

Claims

EPI-CLM-research-protocol

Label: Research evidence protocol

Claimed operation: evidence review and failed-gate rejection precede mechanism extraction, novelty assessment, validated proposed corpus changes and human-approved merge. Distinct from the guarantee assertions in RT-CLM-maintenance-guarantees: this is the epistemic review procedure, not queue atomicity or reviewer authentication. SRC-1, researcher/runbooks/autonomous-research-loop.md, doctrine/design. Direct CI inspection adds only that named gates are dispatched on push/PR/manual workflow; implementations and branch-protection requirements are not implied.

Reject failed gates immediately and log why. --- researcher/runbooks/autonomous-research-loop.md

Use novelty_check.py as a fast mechanism-overlap gate, then apply human or LLM judgment for semantic novelty. --- researcher/runbooks/autonomous-research-loop.md

Every source cited in the change was retrieved. --- researcher/runbooks/autonomous-research-loop.md

EPI-CLM-trace-analysis

Label: Full trace analysis

Claimed operation: the analyzer has the full reasoning trace, including thinking blocks, and should identify failure patterns with evidence. Distinct from RT-RTE-prompt-optimization, which is a route container, not this claim. SRC-1, examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py, doctrine/design embedded in an implemented request. The formatter exposes all stored thinking blocks but truncates each tool result and formats selected fields, so “full” does not license complete observation of environment state or hidden reasoning.

You have access to the FULL reasoning trace including all thinking blocks between tool calls. This gives you unique insight into HOW the agent reasons, not just what it outputs. --- examples/interleaved-thinking/reasoning_trace_optimizer/analyzer.py

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member