AutoSci epistemic report

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

Source-and-claim boundary

The question is how AutoSci turns imported papers, proposed hypotheses and experiment outputs into later reliance. Scope and exclusions: see boundary; evidence identity/layers: see Source register, SRC-1. The only frozen source is implementation and doctrine/design, not an observed research run. This report distinguishes scientific evidence assessment from schema checks, retention and score-directed execution.

The runtime is provisional. Its containers remain valid; the records below split consequential checking, disposition and integration parts rather than superseding them. One additional implementation qualification matters: Hiera loop re-evaluates selected existing candidates, whereas tune creates config variants; reporting an experience revision does not establish that revision drives scheduling. See EPI-RTE-hiera-select and EPI-RTE-hiera-summary. No supplied runtime fact requires supersession.

Material coverage

“Direct” means source bodies read in this attempt; “supplied” means the runtime finding, not a new source inspection. All paths below are SRC-1 at the frozen commit. Available but uninspected operations prevent a system-complete negative.

Entry point or operation	Source path	Coverage and records / limit
Paper import, matching and writing	.claude/skills/ingest/SKILL.md	Direct source-resolution through rebuild body; RT-RTE-knowledge-cycle
Ask, crystallization, init fan-in, CI intake	.claude/skills/ask/SKILL.md, .claude/skills/init/references/parallel-ingest.md, .github/workflows/daily-arxiv.yml	Supplied: RT-RTE-knowledge-cycle, RT-RTE-parallel-ingest, RT-RTE-daily-intake
Wiki compiler, lifecycle transition, cleanup	tools/research_wiki.py	Direct compiler, transition, edge/citation dedup bodies and complete CLI operation-name search; RT-OBJ-context-brief, EPI-RTE-edge-dedup
Browser metadata/edge mutation	tools/serve.py	Supplied: RT-RTE-browser-write; counterpath to enforced scientific acceptance
Automated lint and repair; semantic maintenance	tools/lint.py, .claude/skills/check/SKILL.md	Direct check collector, quality heuristics, fix dispatch and fix bodies; semantic checklist/constraints; EPI-RTE-wiki-check, EPI-RTE-wiki-repair
Ideation filtering and refinement	.claude/skills/ideate/SKILL.md, .claude/skills/refine/SKILL.md	Direct filter/deep-validation and refinement-loop bodies; EPI-CLM-validated-ideas
Experiment hypothesis evaluation/update	.claude/skills/exp-eval/SKILL.md	Direct load/check/synthesis/disposition/integration bodies; EPI-RTE-verdict-check, EPI-RTE-verdict-disposition, EPI-RTE-verdict-integrate
Experiment execution/research gates	.claude/skills/exp-run/SKILL.md, .claude/skills/research/SKILL.md	Supplied: RT-RTE-research-verdict
MCP chat, reply and native search	mcp-servers/llm-review/server.py, .mcp.json	Direct registration-name search confirms three exposed tools; supplied dispatch/retry/thread evidence: RT-RTE-review-call; no extra evaluator inferred
Hiera evaluation, status, finalize, tune, loop	tools/hiera/evaluator.py, tools/hiera/cli.py, tools/hiera/experience.py, tools/hiera/scheduler.py, tools/hiera/tuner.py, tools/hiera/acquisition.py	Direct named bodies; RT-RTE-hiera-search, EPI-RTE-hiera-summary, EPI-RTE-hiera-select
Hiera remote worker, recovery, admission, contract	tools/hiera/remote_worker.py, tools/hiera/remote_adapter.py, tools/hiera/contract.py	Supplied: RT-RTE-hiera-search; no claim of independent remote measurement authentication
Setup activation, writer policy	setup.sh, runtime/policy/writers.yaml	Supplied: RT-RTE-setup-skills, RT-CLM-writer-contract
Other CLI functions / prompt paths	tools/research_wiki.py, .claude/skills/novelty/SKILL.md, .claude/skills/exp-pilot-eval/SKILL.md, .claude/skills/paper-plan/SKILL.md, .claude/skills/review/SKILL.md	CLI names inventoried; query/projection/maturity/backfill bodies and standalone novelty/pilot/review/writing procedures not fully inspected. Their exact evaluator criteria and all downstream publication checks remain unassessed; no system-wide absence of acceptance, derivation or withdrawal is claimed.
Translation and activation alternatives	i18n/	Setup supplied, language equivalence unassessed; conclusions about active English bodies do not prove all language surfaces equivalent.

Missing evidence and prevented conclusions: no populated, provenance-linked research instance establishes produced/checked/accepted claims; no run trace establishes activation or model obedience; no comparison establishes causal benefit, reviewer component effect or transfer. External provider weights, parameter updates, endpoint resolution and search quality are uninspected; see RT-CMP-claude-code and RT-CMP-review-llm. Experiment raw data, reference outcomes and instantiated Hiera contracts are not supplied, preventing scientific-validity assessment. These limits do not erase inspectable compiler, integrity and continuation capacity.

Consequential warrant claims: EPI-CLM-verdict-certainty, EPI-CLM-validated-ideas, RT-CLM-sidecar-confirmation and RT-CLM-writer-contract. README branch claims outside this pin remain excluded as in boundary; no single knowledge-production claim is substituted for these heterogeneous routes.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-paper-pages	Paper identity, reported methods/results, summary and commentary	see record	SRC-1 .claude/skills/ingest/SKILL.md	Imported warrant unknown; summary fidelity unobserved
EPI-OBJ-idea-hypothesis	Proposed hypothesis, novelty/risk/failure statements	see record	SRC-1 .claude/skills/exp-eval/SKILL.md	No candidate instance; proposed is not accepted
EPI-OBJ-experiment-results	Measured effects/setup and result interpretation	see record	SRC-1 .claude/skills/exp-eval/SKILL.md	Raw measurements and provenance unprovided
EPI-OBJ-evaluation-receipt	Execution/result/status claims and numeric score	see record	SRC-1 tools/hiera/evaluator.py	Status success is not scientific success
RT-OBJ-context-brief	Selected retained summaries/relation labels; no new warrant	see record	SRC-1 tools/research_wiki.py	Truncation can omit reasons and qualifications
RT-OBJ-task-contract	Mostly non-truth-apt execution policy; snapshot/digest assertions	see record	SRC-1 tools/hiera/contract.py	Objective and evaluator validity remain external
RT-OBJ-writers	None as authorization commands; declarative policy	see record	SRC-1 runtime/policy/writers.yaml	Not a scientific evaluator or runtime permission gate

Authority-route ledger

Each row is a consequential function, not an inferred successful episode. Repeated mechanism fields refer to the owning record. No observed operation is asserted. Textual scientific transformations remain indeterminate unless their inputs and relation are established; an unobserved model proposal is not automatically classified as ampliative merely because entailment is unproved.

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-knowledge-cycle	content transformation	doctrine only	EPI-OBJ-paper-pages	truth-apt transformation: acquisition/import	External papers/metadata into wiki	Source interpretation and enrichment; see record	Requested ingest	Imported source claims plus interpretation	Prompt-mediated writing	Source warrant unknown, not validated by import	Allows later use of retained claims	see RT-BAP-retained-context	SRC-1 .claude/skills/ingest/SKILL.md	none	none	Import is distinct from new commentary; interpretation fidelity unobserved.
RT-RTE-knowledge-cycle	content transformation	doctrine only	EPI-OBJ-paper-pages	truth-apt transformation: indeterminate	Summary/My take and later answer	Foreground model; see record	Ingest or ask synthesis	Candidate summaries/answers	Advisory answer; optional authorized write	Neither source truth nor entailment checked by storage	Can answer before crystallization approval	see RT-BAP-retained-context	SRC-1 .claude/skills/ingest/SKILL.md	none	none	Need concrete inputs/outputs to separate faithful reshaping, inference and conjecture.
RT-RTE-knowledge-cycle	retention	implemented	EPI-OBJ-paper-pages	no content change	Persisted pages/edges	Filesystem/helper shape checks; see record	During writes	Available later candidates	Persistence, not acceptance	None added by retention	Retained candidates become readable	see RT-BAP-retained-context	SRC-1 tools/research_wiki.py	none	none	Direct writes and browser route can bypass skill gates.
RT-RTE-knowledge-cycle	content transformation	implemented	RT-OBJ-context-brief	truth-apt transformation: non-ampliative reshaping	Selected metadata/context to brief	Fixed budgets, connectivity, field extraction	Rebuild/compile	Selected/truncated pack	Selection/omission only	No added endorsement; lossy carrier	Changes available short context	see RT-BAP-retained-context	SRC-1 tools/research_wiki.py	none	none	Semantic completeness not guaranteed, original pages still reachable.
RT-RTE-knowledge-cycle	operational admission/selection/consumption	doctrine only	RT-OBJ-context-brief	no content change	Relevant full pages for ask	Model relevance; top K <=15	Later ask	Selected pages and cited answer	Advisory knowledge use	Source warrant remains unknown	Permits synthesis; crystallize separately gated	see RT-BAP-retained-context	SRC-1 .claude/skills/ask/SKILL.md	none	none	No observed operation.
RT-RTE-daily-intake	operational admission/selection/consumption	implemented	EPI-OBJ-paper-pages	no content change	Paper intake selection	Model high-confidence instruction; CI branch	Scheduled/manual auto-ingest	Possible committed wiki/raw changes	CI writeback wired; per-paper judgment prompted	Selection not truth acceptance	Can ingest without per-paper human approval	CI/action -> retained wiki -> future agent; permissive, later sessions	SRC-1 .github/workflows/daily-arxiv.yml	none	none	Provider-compatible alternative is inform-only; model compliance unobserved.
EPI-RTE-wiki-check	check/evidence production	implemented	EPI-OBJ-paper-pages	no content change	Representation consistency/quality heuristics	see record	Requested lint	Issue list	Advisory diagnostics	Only checked structural predicates	No universal rejection of wiki use	Caller via issues; advisory, maintenance call	SRC-1 tools/lint.py	none	none	No observed operation.
EPI-RTE-wiki-repair	content transformation	implemented	EPI-OBJ-paper-pages	truth-apt transformation: indeterminate	Defaults and reverse links	see record	Fixer call without dry-run	Changed fields/links	Filesystem mutation	Schema consistency, not scientific truth	Can change later metadata and relation use	Future readers via pages; permissive, until edited	SRC-1 tools/lint.py	none	none	Reverse completion may be entailed under schema assumptions; default values need not be true.
EPI-RTE-edge-dedup	lineage/freshness/recovery	implemented	EPI-OBJ-paper-pages	truth-apt transformation: non-ampliative reshaping	Duplicate relation rows	First canonical triple; see record	Explicit cleanup	Later duplicate evidence removed	Enforcing within cleanup	No adjudication of conflicting evidence	Changes surviving graph evidence	Graph readers via JSONL; enforcing survival, persistent	SRC-1 tools/research_wiki.py	none	none	Warrant may degrade when later evidence differs; see deduction below.
EPI-RTE-verdict-check	check/evidence production	doctrine only	EPI-OBJ-idea-hypothesis	truth-apt transformation: indeterminate	Hypothesis versus result portfolio	see record	After experiment collection	Verdict/concerns	Advisory until disposition	Conditional judgment, not an oracle	Proposes changed reliance/next tests	Foreground via reviewer text; advisory, current verdict	SRC-1 .claude/skills/exp-eval/SKILL.md	EPI-CLM-verdict-certainty	Agreement does not establish calibrated certainty	No actual evidence-consuming check instance supplied.
EPI-RTE-verdict-disposition	disposition/acceptance	doctrine only	EPI-OBJ-idea-hypothesis	no content change	Acceptance/rejection for linked hypothesis use	see record	After verdict; confirmation unless auto	Validated/failed/unchanged	Prompt-level permission; lifecycle shape code	Acceptance specified for supported main single-hypothesis case only	May unlock paper planning or avoid failed idea	Foreground via verdict/status; permissive, later research	SRC-1 .claude/skills/exp-eval/SKILL.md	EPI-CLM-verdict-certainty	none	Status transition cannot prove evidence was consumed; direct metadata path bypasses it.
EPI-RTE-verdict-integrate	lifecycle integration	doctrine only	EPI-OBJ-idea-hypothesis	no content change	Accepted supported claim into evidence graph/context	see record	After supported disposition	Supports edge and Idea updates	Requested connection and rebuild	No additional license beyond disposition	Can guide later paper/research	see RT-BAP-retained-context	SRC-1 .claude/skills/exp-eval/SKILL.md	none	none	Rejected/partial branches retain evidence separately; they are not acceptance.
EPI-RTE-verdict-integrate	retention	doctrine only	EPI-OBJ-experiment-results	no content change	Verdicts, rejection reasons and inconclusive evidence	see record	All verdict branches	Persisted outcome metadata	Requested write	Retention alone no acceptance	Allows failure anti-repetition and next tests	see RT-BAP-retained-context	SRC-1 .claude/skills/exp-eval/SKILL.md	none	none	No observed operation.
RT-RTE-browser-write	operational admission/selection/consumption	implemented	EPI-OBJ-idea-hypothesis	truth-apt transformation: indeterminate	Caller-selected metadata replacement	see record	HTTP PATCH/direct CLI	Changed metadata	Enforcing shape/existence, not review	No truth license	Can change status without experiment acceptance	Browser/CLI -> page -> compiler/model; permissive, persistent	SRC-1 tools/serve.py	RT-CLM-writer-contract	none	Caller identity does not authenticate human scientific review.
RT-RTE-hiera-search	check/evidence production	implemented	EPI-OBJ-evaluation-receipt	truth-apt transformation: acquisition/import	Candidate execution and result-file score	Command evaluator, contract score path and finite-number gate	Admitted/frozen run; local or remote	Failed/success receipt	Process execution and receipt recording	License only for recorded execution/result conditions; metric validity external	Can make candidate score-eligible	Ledger -> scheduler; ranking/permissive, sidecar horizon	SRC-1 tools/hiera/evaluator.py	RT-CLM-sidecar-confirmation	none	Success requires zero exit/no timeout/result file/finite score, not hypothesis confirmation.
EPI-RTE-hiera-summary	content transformation	implemented	EPI-OBJ-evaluation-receipt	truth-apt transformation: entailed derivation	Counts of ledger statuses per point	see record	Status/finalize call	Counts/revision and report	Reporting only	Entailed from ledger premises; no source-truth upgrade	No direct scheduler use of experience digest	CLI/report consumer; advisory, run/report horizon	SRC-1 tools/hiera/experience.py	RT-CLM-sidecar-confirmation	none	Revision includes current timestamp; not a stable outcome-only identity.
EPI-RTE-hiera-select	behavior/policy adaptation	implemented	EPI-OBJ-evaluation-receipt	non-truth-apt policy/content update: select next evaluation/config	Existing score/bout/config state	see record	Loop/tune call	Repeated evaluation or child variants	Code-enforced selection within these paths	Ranking has no broader scientific license	Changes next work based on results/tried configs	Scheduler -> evaluator via candidate ID; enforcing, run budget	SRC-1 tools/hiera/scheduler.py	none	none	No hypothesis criticism, benefit or transfer follows from numeric optimization.

Concrete deductions from inspected implementation (not target runs):

• Given two valid edge rows with identical endpoints/type but evidence E1 then E2, dedup_edges keeps the first canonical key and drops the second without comparing evidence. Semantic relation count is reshaped non-ampliatively, but the surviving evidence lineage loses E2. Malformed lines survive. See EPI-RTE-edge-dedup.
• Given ledger receipts for a point with statuses success, failed, success, the counter increments yield attempts=3, successes=2, failures=1. This is entailed only relative to those recorded statuses; it does not entail two scientifically successful trials. See EPI-RTE-hiera-summary.
• Given a failed idea with title X and failure_reason R, compiler extraction forms - X — R, applies the failed-section character budget and may omit/truncate it again at the pack budget. The brief's “avoid repeating” guidance therefore does not guarantee the model receives the whole failure rationale. See RT-OBJ-context-brief (SRC-1 tools/research_wiki.py).

Additional prompt-only functions: ideation screening/ranking and refinement specify operational selection and revision, not scientific acceptance; see EPI-CLM-validated-ideas. Their checks, dispositions and writes are distinct in that record. No uninspected standalone novelty or pilot criterion is inferred from these callers.

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
EPI-CLM-verdict-certainty	see record	SRC-1 .claude/skills/exp-eval/SKILL.md, doctrine/design	see record	RT-RTE-review-call, RT-RTE-research-verdict helper affordances; check/disposition remain doctrine	None registered	No independent reviewer comparison, calibration or experimental outcomes	Evidence-consuming acceptance procedure is specified with fatal-methodology priority and inconclusive withholding	Agreement is not sufficient evidence of high certainty; foreground reads reviewer verdict first
EPI-CLM-validated-ideas	see record	SRC-1 .claude/skills/ideate/SKILL.md, .claude/skills/refine/SKILL.md, doctrine/design	see record	RT-RTE-review-call transports model advice; wiki helpers afford writes	None registered	No novelty ground truth, exhaustive search, reviewer reliability or benefit comparison	Ranking/filtering/revision requested; no claim of accepted new science follows	Deep validation can be skipped, retained ideas are proposed, review score is not a truth oracle
RT-CLM-writer-contract	see record	SRC-1 runtime/policy/writers.yaml, doctrine/design and implementation counterpath	see record	RT-RTE-browser-write	None registered	No deployed grants supplied	Writer policy does not enforce scientific acceptance	Direct metadata writes make a universal approval invariant false
RT-CLM-sidecar-confirmation	see record	SRC-1 .claude/skills/hiera-experiment/SKILL.md, tools/hiera/cli.py, doctrine/design and implementation	see record	RT-RTE-hiera-search, EPI-RTE-hiera-summary, EPI-RTE-hiera-select	None registered	No evaluator validity, formal-confirmation instance or secure confinement evidence	Exploratory evaluation and report-only finalization are supported	Finite-score success is narrower than confirmation; arbitrary code is not sandboxed by cwd

Bounded conclusion

AutoSci specifies acquisition of literature claims and model-authored interpretation, then retains pages for later question answering, proposal generation and experiment planning. Source warrant is unknown rather than upgraded by ingestion. Compiler and cleanup routes demonstrably reshape retained carriers; they do not evaluate scientific truth and can remove qualifications or later evidence. See RT-RTE-knowledge-cycle, RT-OBJ-context-brief and EPI-RTE-edge-dedup.

Its strongest inspected scientific-acceptance design is hypothesis evaluation against experiment/sibling results, with methodological criticism, conservative disagreement handling, interactive veto (omitted in auto), and withholding on inconclusive evidence. Supported main single-hypothesis results may license validated status and paper planning; rejected results license failed status plus concrete anti-repetition memory. Accepted supported claims are then meant to acquire evidence edges and derived-context connections. Those phases are separately specified, not observed accepted ampliative knowledge. No supplied instance establishes what new proposition was conjectured, whether it followed from its premises, or whether a check accepted it. See EPI-RTE-verdict-check, EPI-RTE-verdict-disposition and EPI-RTE-verdict-integrate.

Automated integrity checks license representation-level conclusions only; deterministic repairs do not accept the truth of filled defaults or reverse relations. Direct/browser metadata writes also permit operational status changes without the experimental procedure. Consequently status alone cannot authenticate acceptance history. See EPI-RTE-wiki-check, EPI-RTE-wiki-repair and RT-RTE-browser-write.

Hiera implements contract-bound execution, finite-score receipts, reporting and result-directed continuation. Receipt counts are entailed within the recorded ledger domain, not from independently warranted scientific premises. Scores and tried-config history can change what runs next without producing a truth-apt explanatory theory. Report-only finalization explicitly withholds formal wiki promotion. The experience revision itself is not the inspected loop's selection signal. See RT-RTE-hiera-search, EPI-RTE-hiera-summary and EPI-RTE-hiera-select.

Localized scientific hypothesis criticism/iteration is specified separately from learning benefit or reflection on AutoSci itself; see RT-RTE-research-verdict and EPI-RTE-verdict-check. Implementation supports independent narrower capacities, but no observed activation, causal improvement, calibrated certainty or system-wide verdict follows. Uninspected novelty/pilot/publication paths prevent a complete negative about other acceptance routes. External distributed-parametric state remains independently unresolved, not proof of either weight learning or fixity; see RT-CMP-claude-code and RT-CMP-review-llm.

Shared records

Operative objects

EPI-OBJ-paper-pages

Label: Paper pages

Natural-language paper claims, method/result summaries and My take, with symbolic identity/importance/link metadata in Markdown under wiki/papers. The paper import interpretation is distinct from RT-OBJ-context-brief (derived selected carrier), not a duplicate. SRC-1 .claude/skills/ingest/SKILL.md, doctrine/design. No populated page instance or provenance-linked acceptance supplied.

Experiment & Results — setup, primary metrics, and results in one place. Do not stop at "they beat the baseline"; cite the numbers and their conditions. --- .claude/skills/ingest/SKILL.md

EPI-OBJ-idea-hypothesis

Label: Idea hypothesis

Natural-language Hypothesis, Approach sketch, Risks and Novelty argument in wiki/ideas Markdown, plus symbolically consumed lifecycle metadata. Scope is the scientific candidate and changed reliance, not the RT-OBJ-task-contract execution policy or RT-OBJ-context-brief copy. No counterpart among supplied object records. Source: SRC-1 .claude/skills/exp-eval/SKILL.md, doctrine/design.

• ## Hypothesis, ## Approach sketch, ## Risks, ## Novelty argument --- .claude/skills/exp-eval/SKILL.md

EPI-OBJ-experiment-results

Label: Experiment results

Natural-language setup/results/interpretation and numerical metric claims retained in experiment Markdown. Numerical values are symbolic measurements when consumed as metrics; prose interpretation remains natural language. They feed hypothesis judgment, distinct from idea hypotheses and sidecar receipts; no supplied object counterpart. SRC-1 .claude/skills/exp-eval/SKILL.md, doctrine/design; no measurement instance supplied.

• metrics and full ## Results section --- .claude/skills/exp-eval/SKILL.md

EPI-OBJ-evaluation-receipt

Label: Evaluation receipt

Symbolic JSON ledger receipt: execution status, score/direction, result digest and result-file data. Sidecar filesystem/ledger substrate; arbitrary result payload semantics are evaluator-defined, not automatically natural-language knowledge. Distinct from RT-OBJ-task-contract (instructions and frozen policy) and RT-OBJ-context-brief (wiki text). SRC-1 tools/hiera/evaluator.py, implementation. Available receipt schema is not an observed receipt instance.

self.ledger.set_status(candidate_id, "evaluated", last_attempt_id=attempt_id, last_depth=depth) --- tools/hiera/evaluator.py

Routes

EPI-RTE-wiki-check

Label: Wiki integrity and quality checks

No identical supplied route referent: the RT-RTE-knowledge-cycle container describes ingest/use, not this maintenance function.

The lint function collects missing-field, broken-link, orphan, field-value, conditional-required-field, target, reverse-reference, edge, citation and content-quality issues. This checks representation consistency and selected heuristics, not whether scientific assertions are true.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Issue objects/report to caller.
• Later read-back: inapplicable — checks alone do not retain or reload candidates; a separate fixer consumes issues.
• Delegated visibility: Local caller sees issue category and structured data; no provider is required for these checks.
• Selection predicate: All discovered pages and declared schema rules; content-quality checks use importance, maturity, references and empty sections.
• Invalidation or expiry: Issues are recomputed from current files; no semantic endorsement expiry exists.
• Activation or effect: Issues can guide the separate repair route or a person; no automatic rejection of all wiki use is established.
• Evidence limits: No lint run was supplied; regex/parser correctness and scientific truth are not established by collection of checks.

SRC-1 tools/lint.py; implementation.

issues.extend(check_content_quality(wiki_dir, pages)) --- tools/lint.py

EPI-RTE-wiki-repair

Label: Deterministic issue repair

No identical supplied route referent: the RT-RTE-knowledge-cycle container describes ingest/use, not this maintenance function.

fix_issues selects fixable issues in two categories. It adds reverse links or declared defaults, using structured issue data. This mutates representations without evaluating the underlying scientific claim. It is distinct from the preceding checker and from experiment acceptance.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: FixResult list and possibly changed Markdown.
• Later read-back: Changed fields/links can reach later page reads, compilation and subsequent lint.
• Delegated visibility: Local fixer receives structured issue data and schema-provided defaults/templates, not a review verdict.
• Selection predicate: Explicit fixer call; fixable xref-asymmetry or missing-field only; dry_run suppresses writes.
• Invalidation or expiry: Subsequent edits or repository restoration can replace changes; no automatic semantic rollback is established.
• Activation or effect: Writes are implemented when not dry-run; downstream scientific reliance is unobserved.
• Evidence limits: A reverse link/default can encode an unwarranted relation when its premise or schema is wrong; consistency is not acceptance.

Caller chooses repair/dry-run and can withhold it. Guidance is the schema and the checker issue, not an answer oracle. Trigger is requested maintenance, not a measured scientific-improvement objective. /check asks for user approval for non-deterministic suggestions; generic direct writes remain separate.

SRC-1 tools/lint.py; implementation.

elif issue.category == "missing-field": --- tools/lint.py

EPI-RTE-edge-dedup

Label: First-occurrence edge cleanup

No identical supplied route referent: the RT-RTE-knowledge-cycle container describes ingest/use, not this maintenance function.

dedup_edges canonicalizes symmetric endpoints and keeps the first row per canonical endpoint/type triple. Later duplicate evidence is discarded even if its text differs. Malformed JSON is preserved. This is a material lineage operation, not a semantic conflict adjudicator.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Counts plus overwritten edges.jsonl.
• Later read-back: Later load_edges/context rebuilds read the surviving rows.
• Delegated visibility: Local filesystem caller; no model evaluator.
• Selection predicate: First canonical triple wins; duplicate evidence text does not participate in the key.
• Invalidation or expiry: Overwrites current graph; recovery requires retained repository/file history.
• Activation or effect: Changes which evidence survives for later readers; no check chooses the better evidence.
• Evidence limits: Deduction from implementation, not a target execution; bibliography cleanup uses a separate endpoint-pair key.

Caller requests cleanup and can decline it. Code decides survival; no human per-row veto or reference answer is wired. Existing add-edge admission is in RT-RTE-knowledge-cycle; cleanup is a distinct post-admission operation.

SRC-1 tools/research_wiki.py; implementation.

kept.append(stripped) # preserve malformed lines --- tools/research_wiki.py

EPI-RTE-verdict-check

Label: Hypothesis evidence assessment

Part of: RT-RTE-research-verdict

The foreground agent loads the idea hypothesis, current experiment results/setup and sibling outcomes. It requests a reviewer verdict on statistical significance, effect size, validity, confounds and generalization, then synthesizes a final verdict. Fatal methodology findings take priority. This is an evidence-consuming check specification, not proof it was executed.

• doctrine conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: Supported/partially_supported/not_supported/inconclusive verdict, strength, concerns and proposed status.
• Later read-back: Recorded verdict is requested in the experiment Idea updates section; retained results can enter later rounds.
• Delegated visibility: Review provider gets hypothesis, setup, full Results, key_result and sibling summary, not all source data or the whole wiki.
• Selection predicate: Named linked idea required; aggregate experiments and method concerns guide judgment.
• Invalidation or expiry: New results can lead to another verdict; no formal statistical acceptance threshold is fixed.
• Activation or effect: Advisory verdict feeds separate disposition; the adapter itself cannot update the wiki.
• Evidence limits: No supplied results, review trace or reference-answer oracle. Foreground reads the review before its own supposedly independent verdict; actual independence is unestablished.

The model proposes and synthesizes; interactive user can veto subsequent writes, auto mode omits that veto. This is open research judgment, not a benchmark oracle. Theory-builder conditions 1–4: condition 1 conclusion status: afforded (localized hypothesis sections); condition 2 conclusion status: claimed (verdict requested against their content); condition 3 conclusion status: claimed (methodological criticism and changed reliance requested); condition 4 conclusion status: claimed (sibling results and further experiments requested). Addressability is page/section-level; persistence is requested across experiment rounds/sessions. Learning benefit, reflection on AutoSci method, autonomous theory building and exercised self-improvement are uninspected independently.

SRC-1 .claude/skills/exp-eval/SKILL.md; doctrine/design.

Fatal findings take priority: if either party finds a methodological issue (data leakage, unfair comparison), that finding takes precedence --- .claude/skills/exp-eval/SKILL.md

EPI-RTE-verdict-disposition

Label: Verdict-dependent reliance decision

Part of: RT-RTE-research-verdict

After the check, user confirmation is requested unless auto. Supported may advance a single-hypothesis idea whose main experiment this is to validated; partial leaves lifecycle unchanged; not_supported requests failed with concrete failure_reason; inconclusive leaves status unchanged. The intended use is subsequent research and possible paper planning, scoped to the linked hypothesis/evidence portfolio. No software proof of scientific validity follows from the status.

• doctrine conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: Planned/authorized status and relation changes, terminal report.
• Later read-back: Retained status and failure_reason can affect later ideation and paper planning.
• Delegated visibility: Foreground agent consumes final verdict; user sees planned changes unless auto.
• Selection predicate: Verdict branch and main-experiment/single-hypothesis qualification; user may withhold confirmation.
• Invalidation or expiry: Later edits/results can change reliance; transition helper restricts some transitions but direct set-meta bypasses it.
• Activation or effect: Prompt-level permissive decision to change reliance; transition code only checks lifecycle/preconditions.
• Evidence limits: No accepted claim instance is supplied. The word validated names a requested status, not demonstrated generalization.

Model proposes/decides branch; human veto exists only on interactive path. Proposal guidance is the verdict and methodological concerns. Rejection leaves status unchanged or marks failed; recovery is further experiments/manual edits, not a transactional multi-file rollback. No expected-answer oracle is supplied.

SRC-1 .claude/skills/exp-eval/SKILL.md; doctrine/design.

1. Do not modify idea status: insufficient evidence to make a judgment --- .claude/skills/exp-eval/SKILL.md

EPI-RTE-verdict-integrate

Label: Post-verdict wiki connection

Part of: RT-RTE-research-verdict

Following disposition, supported/partial branches request supports edges, rejected hypotheses request invalidates edges and concrete failure reasons; all paths record Idea updates and rebuild derived context. This is intended post-acceptance integration for supported cases, and retained rejection/evidence bookkeeping for other branches. It does not retroactively turn every retained idea into accepted knowledge.

• doctrine conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: Updated experiment/idea pages, evidence edges, rebuilt context and log.
• Later read-back: Later ask, ideate, research and paper-plan reads can consume status/edges/reasons.
• Delegated visibility: Foreground writer receives the verdict; downstream compiler exposes selected fields, often omitting the full reasoning.
• Selection predicate: Disposition branch determines supports/invalidates; all paths record verdict metadata.
• Invalidation or expiry: Derived brief is overwritten; source pages persist until edited; no automatic withdrawal propagation to every prose claim is established.
• Activation or effect: Connects disposition to retrievable pages and relationship summaries; actual behavioral dependence unobserved.
• Evidence limits: Integration is doctrine; helper availability supports only physical write capacity. Partial and rejected retention must not be classified as acceptance.

Admission owner and veto are those of disposition; graph helper rejects some malformed edges, not false evidence. Guidance and persisted content are branch-specific verdict, key_result and failure_reason. Recovery is subsequent explicit edits/rebuilds.

SRC-1 .claude/skills/exp-eval/SKILL.md; doctrine/design.

Fill in the ## Idea updates section of the experiment page (records changes to the linked idea, not a separate claim entity): --- .claude/skills/exp-eval/SKILL.md

EPI-RTE-hiera-summary

Label: Receipt counts and report derivation

Part of: RT-RTE-hiera-search

build_snapshot groups ledger evaluations by candidate point_id and counts attempts plus success versus other statuses. It derives timestamped/digested symbolic summaries. Success here is an evaluator status, not confirmation of the scientific hypothesis. CLI status/finalize consume the revision digest, not the full per-point counts; coverage_scores is separately printed as advisory coverage.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Point counts and revision; status/finalize expose revision and coverage, finalizer writes FINAL_REPORT.json.
• Later read-back: Later status/finalize recompute from persisted ledger; final report can be read by caller.
• Delegated visibility: Local CLI sees counts/digest; no model receives a scientific explanation automatically.
• Selection predicate: Evaluation receipts with a string point_id; status equal to success selects successes, everything else failures.
• Invalidation or expiry: Recomputed with updated_at, so revision can change without changed outcomes; ledger remains the source.
• Activation or effect: Implemented reporting, not an automatic wiki acceptance or policy gate.
• Evidence limits: No observed receipts. Counts are entailed only relative to ledger statuses, not evidence that commands or benchmarks were honest.

SRC-1 tools/hiera/experience.py; implementation.

row["successes" if receipt.get("status") == "success" else "failures"] += 1 --- tools/hiera/experience.py

EPI-RTE-hiera-select

Label: Score-directed continuation and config search

Part of: RT-RTE-hiera-search

choose_deep_candidate requires finite best_score and fewer than max bouts, orders by objective direction, bouts and ID, records a TUNE decision or returns STOP. cmd_loop consumes that decision and calls evaluate on the same candidate at deep depth. Unlike cmd_tune, loop does not materialize config variants. cmd_tune separately proposes bounded Cartesian configs excluding tried digests and evaluates immutable children.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: TUNE/STOP decision, subsequent evaluation receipts or tuned children.
• Later read-back: Ledger scores/bouts and tried configs are loaded by later loop/tune invocations.
• Delegated visibility: Executor receives selected candidate ID/config; no explanation-bearing theory criticism is required.
• Selection predicate: Finite best score, bout/budget limits and objective direction; tuning uses schema and deterministic enumeration.
• Invalidation or expiry: Changed receipts affect eligibility/order; exhausted budget or STOP ends continuation.
• Activation or effect: Enforcing code selection within loop; best score ranks operational priority, not scientific warrant.
• Evidence limits: No run or benchmark contract instance. Coverage priority and experience revision are reported but not the inspected loop selector.

Author/user supplies candidate/contract; admission and approval controls are in RT-RTE-hiera-search. Code chooses continuation, caller can stop or change permitted run arguments. No oracle is supplied by finite scores. Config changes persist as children; source/contract integrity failures reject evaluation. Theory content of candidate code is not determinable without an instance; scores alone establish neither content-directed criticism nor improved future capacity.

SRC-1 tools/hiera/scheduler.py; implementation.

if not finite_number(record.get("best_score")): --- tools/hiera/scheduler.py

The Hiera evaluator's success gate and consequence, supplementary implementation evidence for RT-RTE-hiera-search and EPI-RTE-hiera-select:

if execution.get("returncode") == 0 and not execution.get("timed_out") and result_file.is_file(): --- tools/hiera/evaluator.py

"status": "success", "score": float(score), "result": data, "result_digest": sha256_file(result_file), --- tools/hiera/evaluator.py

The scheduling consumer and variant generator are different operations:

receipt = evaluator.evaluate(decision["candidate_id"], "deep") --- tools/hiera/cli.py

configs = propose_configs(schema, parent.get("config", {}), tried, args.limit) --- tools/hiera/cli.py

if canonical_digest(config) in tried: continue --- tools/hiera/tuner.py

Reporting does not imply experience counts are the selection policy:

summary["experience_revision"] = build_snapshot(ledger)["revision"] --- tools/hiera/cli.py

Claims

EPI-CLM-verdict-certainty

Label: Agreement and hypothesis support

Part of: RT-RTE-research-verdict

This claim is a doctrine part of the research route (cross-kind containment): exp-eval specifies high certainty on reviewer agreement and supported main single-hypothesis transition to validated. Distinct from RT-CLM-sidecar-confirmation, which concerns exploratory versus confirmatory paths. SRC-1 .claude/skills/exp-eval/SKILL.md, doctrine/design. The claim's target is support for a named idea from a portfolio, not universal benchmark correctness. The procedure asks for evidence consumption but supplies no executed verdict or calibration evidence. Agreement alone does not warrant high certainty, especially when the foreground reads the other verdict before forming its own.

Both agree (same verdict): use that verdict, high certainty --- .claude/skills/exp-eval/SKILL.md

If the idea covers a single hypothesis and this experiment is the main experiment block, transition the idea to validated: --- .claude/skills/exp-eval/SKILL.md

EPI-CLM-validated-ideas

Label: Validated ideas and review-driven refinement

No counterpart among supplied claim records: this is candidate novelty/feasibility/ranking and artifact refinement, not writer ownership or formal sidecar confirmation. SRC-1 .claude/skills/ideate/SKILL.md and .claude/skills/refine/SKILL.md, doctrine/design. The caller claims filter/deep validation and writing of validated ideas, but outputs proposed status and permits skip-validation/default priority. It specifies first-pass feasibility/search/banlist checks; disposition eliminates low-feasibility similar work or known-failure-correlated ideas, then ranks survivors using novelty and review scores. Interactive user can adjust/veto; auto delegates. These are proposal-selection criteria, not acceptance of hypothesis truth. Expected-answer oracle: none supplied; web similarity and reviewer advice are not an exhaustive novelty oracle. Full standalone novelty procedure uninspected.

Write the validated ideas to the wiki (including eliminated ideas, with their elimination reasons recorded). --- .claude/skills/ideate/SKILL.md

(Skip if --skip-validation: proceed directly to Phase 4: Write to Wiki with default priority = 3 for all survivors.) --- .claude/skills/ideate/SKILL.md

Write top ideas (status: proposed): --- .claude/skills/ideate/SKILL.md

Refinement separately specifies reviewer checking, score/verdict termination, direct artifact fixes and wiki relation changes, while experimental-data/domain-expert requirements remain unresolved. Author model interprets advice; artifact/wiki edits persist and rebuilds expose them to future readers. This is policy/content adaptation requested in prompts, not an implemented scientific verifier or an observed improvement. Guidance is reviewer weaknesses; no automatic rollback of incorrect edits is established. It is not evidence of revising AutoSci's own methodology: selected artifacts may be external work products.

Category D — Out of scope (skip): - Requires new experimental data → cannot resolve in refine loop - Requires domain expert judgment → mark as unresolved - → Record in unresolved_issues --- .claude/skills/refine/SKILL.md

Target score reached: score >= target-score → terminate, output final report --- .claude/skills/refine/SKILL.md