academic-research-skills report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The five submitted members support publication of a bounded source account, with the inventory limit below. I checked the declarations, evidence strength, annotations, containment relations and reconciliation against the supplied contracts and independently inspected consequential source paths. The supplied structural report says none; all four correction-answer files are empty. There are no structural findings or correction arguments requiring disposition.

The source checkout's HEAD equals the revision in the boundary, and its status was clean. The tracked top-level tree agrees with the declared workflow boundary, including separately shipped screening, adapters and evaluation tools. Added inspection uses SRC-1 only. It supplies implementation and doctrine/design evidence, not observed operation. No target code, tests, examples, models or services were executed. The reports consistently withhold execution, accuracy, benefit and causal attribution; source inspection does not strengthen those conclusions.

Runtime, workflow and authority

  • Entry and model components. RT-RTE-command-dispatch is supported by the Pi input transformation, activation writes and branch-local restore callbacks in SRC-1, pi/wrapper.js. Loading command/role text does not execute a research pipeline. RT-CMP-session-model and RT-CMP-cross-model distinguish external parametric models from shipped guidance, declared aliases from weight identities, and requested fallback controls from provider guarantees. SRC-1, shared/model_tiering.md and scripts/cross_model_codex_transport.py support those distinctions. Neither component claims observed parameter change or immutable serving weights.
  • Stage progression and delivery. RT-OBJ-pipeline-skill, RT-RTE-pipeline-stages and RT-CLM-complete-pipeline correctly identify model-interpreted policy rather than a deterministic stage engine. Independent searches of SRC-1, academic-pipeline/SKILL.md and academic-pipeline/agents/pipeline_orchestrator_agent.md support the stage order, mandatory decisions, bounded integrity/revision loops, Stage-5 FULL checkpoint and optional Stage-6 terminal acknowledgement in the workflow account. The three-round unresolved-integrity procedure is not a universal override of formatter refusal; the workflow's final-output section retains that distinction.
  • Tool mediation. RT-RTE-tool-scope, RT-BAP-tool-denial, RT-BAP-role-guidance and RT-CLM-pi-scope preserve different enforcement units. SRC-1, hooks/hooks.json, hooks/run_guard.sh, scripts/ars_write_scope_guard.py and pi/wrapper.js support conditional Claude denials, normal permission pass-through, launcher degradation and the absence of a wrapper-installed Pi hook. Main-session Bash and unsupported tools defeat a universal fence claim, which these records do not make.
  • Revision and final output. SRC-1, academic-pipeline/agents/pipeline_orchestrator_agent.md, independently supports the author-sidecar, exact-scope patch, one re-emission, structural escalation, three sequential re-review phases, deferral and fail-closed abort obligations in RT-RTE-revision-review. The workflow and memory annotations add consequences without changing that route's identity or promoting prescribed checker behavior into audited implementation. RT-RTE-final-output and RT-OBJ-ref-markers correctly separate prose joins, current policy stamps, terminal markers, clean artifacts and package-report freshness. SRC-1, scripts/verification_gate/__init__.py, scripts/verify_passport.py and academic-paper/agents/formatter_agent.md support the narrower helper checks and instructed formatter refusal. The workflow's post-formatter stdout-token policy is a separate channel, not a second interpretation of in-marker HIGH-BLOCK.
  • Secondary verification and method changes. Independent inspection of thread construction and returned-source binding in SRC-1, scripts/cross_model_codex_transport.py, supports the conditional containment and evidence-shape rejection in RT-RTE-cross-verification. It does not establish page support, model independence or deployed isolation. SRC-1, scripts/run_evals.py and scripts/check_ranking_lift.py, support RT-RTE-evaluation-admission: selected measured metrics, regression acknowledgement and open same-repository issue checks, not automatic successor generation or deployment. The separate learning, reflection and own-system-change qualifications remain justified.

Retained objects and later consumers

  • Recovery. RT-OBJ-run-ledger, RT-RTE-checkpoint-recovery and their memory annotations accurately separate executable locked append/chain validation from the prompt-layer reset ledger. Independent reads of SRC-1, scripts/run_ledger.py and academic-pipeline/references/passport_as_reset_boundary.md, support receipt validation, advisory cooperating-writer locks, pending-decision handling and fresh-session reload. Hashes do not authenticate user words or prevent valid-file rollback; an in-session authority declaration cannot evict tokens.
  • Corpus and cache. MEM-OBJ-literature-corpus and MEM-RTE-corpus-reuse distinguish source pointers, serialized screening reasons and later relevance selection. SRC-1, scripts/adapters/folder_scan.py, sr-screener/scripts/build_outputs.py and both literature-role files support metadata admission and read-only re-screening, not paper acquisition or forced inclusion. MEM-OBJ-verification-cache and MEM-RTE-cache-verification accurately distinguish the generic SQLite payload capability from its versioned resolver-decision producer. SRC-1, scripts/verification_cache.py and scripts/contamination_signals.py, support exact-key, TTL, typed/versioned gate reuse and optional stale bypass; stored matches are not retained source evidence.
  • Reading declarations. MEM-OBJ-human-read-log, MEM-RTE-read-attestation, EPI-OBJ-read-ledger and EPI-RTE-read-withdrawal preserve declaration, storage and transient resolution as different functions. SRC-1, commands/ars-mark-read.md, scripts/ars_mark_read.py and scripts/human_read_attestation_resolver.py, support peer-log rescission and latest-event/scope-based eligibility. Eligibility does not establish reading or claim support.
  • Other histories. MEM-OBJ-inquiry-ledger and MEM-RTE-inquiry-summary retain author adoption separately from automatic checkpoint projection. SRC-1, scripts/inquiry_branch_ledger.py, supports live-status selection and flag/moment/branch-count gating; the report leaves full publication safety unaudited. MEM-OBJ-review-criteria-binding correctly describes context identity, exact marker receipts and changed-target conflicts, not substantive criterion application; I checked SRC-1, scripts/review_criteria_binding.py. MEM-OBJ-process-record has a prescribed human reader, not an established automated lesson-adoption route, in SRC-1, academic-pipeline/references/process_summary_protocol.md. MEM-OBJ-activity-store and MEM-RTE-activity-history distinguish post-terminal append/render/delete from model input. SRC-1, scripts/adjudication_activity.py and the tracker contract support sequence-window selection, idempotence/conflict checks, deletion confirmations and the explicit model-input prohibition. Uninspected extraction completeness is not silently promoted.

Epistemic functions and reconciliation

EPI-OBJ-synthesis-report, EPI-RTE-source-grading, EPI-RTE-synthesis-content and EPI-CLM-synthesis-understanding retain substantive source grading and synthesis doctrine without inventing checked entailment or a generated discovery. SRC-1, deep-research/agents/source_verification_agent.md and deep-research/agents/synthesis_agent.md, support requested cross-source checking and scholar-pending interpretations. An indeterminate content relation is appropriate without an input/output witness.

EPI-OBJ-citation-summary and EPI-RTE-citation-lookup match the inspected reducer: a match wins; absent matches, an ID-keyed miss yields false; manual skips and title-only misses can yield unresolvable. Anchor-kind presence does not check its value. EPI-OBJ-claim-audit, EPI-RTE-claim-judgment and EPI-RTE-audit-disposition correctly separate injected retrieval/judgment from deterministic refusal classification. SRC-1, scripts/claim_audit_pipeline.py and scripts/claim_audit_finalizer.py, support the reserved corpus parameter, distinct uncited streams, run-bound unknown model identity and UNSUPPORTED-versus-AMBIGUOUS countercases. Computed refusal still needs its downstream consumer.

EPI-OBJ-screening-decisions, EPI-RTE-screening-merge, EPI-RTE-pilot-comparison and RT-RTE-screening-fanout preserve decision generation, merge/QC, reference-label comparison and builder admission. SRC-1, sr-screener/scripts/merge_decisions.py, sr-screener/scripts/build_workflow.py and sr-screener/templates/workflows/ta_screening.template.js, support pending work, override precedence, context/label freshness, generated-job scope and one retry. The local pilot oracle is not full-dataset truth. The criteria theory-builder account separately qualifies consumption, criticism, iteration and learning; protocol amendment instructions are not evidence of an exercised criticism cycle.

EPI-CLM-integrity-boundary and EPI-CLM-review-calibration agree with SRC-1, README.md: reported consistency is not empirical authenticity, and live review remains NOT_CALIBRATED. The epistemic ledger's function-level implementation statuses do not contradict broader runtime container statuses or establish lifecycle integration.

I checked every Part of: relation: the two research routes belong to stage progression; the four finalization subroutes belong to final output; merge, pilot comparison and corpus export/reuse belong to screening's declared producer-to-consumer route. The non-screening corpus imports are expressly outside that containment. The reconciliation correctly supersedes only the duplicate retained read-log object, not its distinct write/resolution routes. Other object distinctions, container/part relationships and annotations preserve identities. Its integration-issue dispositions neither erase missing inspections nor select stronger statuses. No unresolved cross-report contradiction needs correction.

Consequential omission searches and stopping decision

I used the tracked tree, stage handoff dependencies and alternate consumers to challenge three central conclusions: whether source checks were mistaken for scientific acceptance; whether retained material could shape later work beyond the listed histories; and whether another shipped path invalidated the host-dependent enforcement account. Inspection of the quantitative-synthesis role in SRC-1, deep-research/agents/meta_analysis_agent.md, found pooling feasibility, effect-size and GRADE instructions. This supports retaining the epistemic report's explicit unassessed quantitative branch, not converting its bounded findings into a system-wide denial of derivation or knowledge production. The helper/formatter, manual screening, Pi and inline alternate paths likewise support the reports' existing guarantee qualifications.

The retained-state search found two additional passport selection mechanisms in SRC-1, academic-pipeline/agents/pipeline_orchestrator_agent.md: confirmed standing constraints and integrity-excluded references. They are not fully represented by the memory inventory. They do not contradict corpus immutability: selection changes what reaches a writer, not the adapter-owned corpus. Their omission is local because the reports already establish selective retained authority, mixed admission and incomplete mechanism coverage, without a complete-inventory or absent-selection finding. The resulting comparison limit is below.

Further optional expansion stopped because the selected challenges have supported dispositions and no identified material contradiction remains undisposed. Full meta-analysis correctness, all integrity/audit branches, every adapter and all helper safety properties remain outside this verification's completed trace. Those limits do not license broader guarantees; nor does the existence of additional files require an exhaustive audit to publish the bounded account.

Blockers

none

Limits

  • The retained-state inventory is incomplete for two passport-to-dispatch selectors alongside MEM-RTE-corpus-reuse, RT-RTE-pipeline-stages and RT-RTE-checkpoint-recovery. SRC-1, academic-pipeline/agents/pipeline_orchestrator_agent.md, prescribes retaining user-confirmed standing constraints, selecting them by applicable stage and recording withdrawal; it separately records integrity NOT_FOUND references and removes active exclusions from later writer inputs until user restoration. These are doctrine/design, not observed operation or an audited persistence implementation. The supported corpus, recovery, cache and other named routes remain valid, but readers must withhold claims of complete passport selection coverage or exhaustive comparisons of user-constraint propagation and negative source-admission memory. The source states:

    A constraint the user sets for the whole run, such as a word ceiling, something not to do, or a narrower scope, is kept in the passport's standing_constraints[] (shared/handoff_schemas.md Schema 9; entry shape shared/contracts/passport/standing_constraint_entry.schema.json), so that compaction, a resume, or a subagent dispatch does not drop it. The run ledger keeps the conversation; this list is what later stages read. --- academic-pipeline/agents/pipeline_orchestrator_agent.md

Carrying. Every later writer dispatch, including revision rounds and integrity correction rounds, leaves the active entries out of the corpus context and the bibliography it carries, and names them as excluded by the integrity gate, so that a roadmap item or the draft text that mentions one does not bring it back. --- academic-pipeline/agents/pipeline_orchestrator_agent.md