Aurora Agent

AURORA Agent (bioprism)

Query-compiled inference for executable biology.

Context engineering, with receipts.

An MCP server and CLI built on the FIBER decision-context compiler: a typed decision query is compiled into the smallest decision-sufficient evidence region, delivered with a Context Certificate stating exactly what was omitted.

The preclinical glioma engine now compiles a complete bounded research workflow, adapts its next stage from typed outcomes, and admits only capability-matched institution-local workers before dispatch. Worker routing is deterministic and fail-closed; the engine never turns an unimplemented stage into a generic action and never makes clinical decisions. The route can also drive the autonomous engine end to end through the typed stage/action adapter; MCP exposes this as a simulation-only rehearsal, while production hosts supply their own local workers. Evidence is an execution gate, not a passive report: P01 cross-family triangulation must qualify before the P07 stage router can admit a worker. The evidence-gated stage engine returns an explicit evidence hold or capability hold with next actions, and only then runs bounded evidence, multimodal, mechanism, experiment, computation, replication, release, and federation stages.

Implementation of the AURORA BioPRISM / OncoWorld / FIBER blueprint (v0.6, 935 registered spec modules). A Rust workspace whose central idea is that context assembly is a compiler pass: instead of retrieval plus summarisation plus vibes-based compaction, a typed decision query is compiled into the smallest decision-sufficient evidence region, delivered as a Decision Section, and accompanied by a Context Certificate that states exactly what was omitted and whether the omission could have changed the decision.

Compile the smallest decision-sufficient evidence region. Never traverse the whole knowledge structure by default.

What the measurements actually say

The reference world ships 761 facts, 750 of them exploratory distractors that all consume the same protected cohort_id hub. FIBER compiles the query down to 11 facts (1.45% of the world) and the deterministic oracle still returns the correct verdict with all four leakage witnesses.

It is not alone in doing so. Under equal tuning, a 5-hop incidence walk and a BM25 retriever at k=11 select exactly the same eleven facts. The distribution's own compare_baselines.py measures the graph baseline only at depth 7 and unbounded — the two settings where it returns everything — and reports a 69× advantage that vanishes under equal tuning. That is a strawman comparison, and correcting it is what 43.38 and 43.41 require.

So the reference world cannot tell these methods apart. crates/worldgen makes the structure a parameter and builds one that can — distractors attached near the target instead of at a hub leaf, decisive facts behind a relay chain, and distractor tags camouflaged to tokenise into the protected vocabulary:

Three distinct failure modes appear. The graph walk has no usable depth: 5–10 pull in 98% of the world and still miss every decisive witness; 11 is the first sound setting and by then it has taken everything. BM25 reaches the right verdict from a 91% protected closure — right by luck, having dropped a protected fact that happened not to matter, and raising k to 50 never recovers it. FIBER is the only admissible strategy: right verdict and full closure, at 11 facts.

That last failure is why the harness ranks on admissibility rather than verdict alone — ranking on verdict would have crowned the strategy that violated the mandatory closure and got away with it.

This does not show FIBER wins generally: the discriminating world was built to expose these modes, just as the reference world was built to expose hub expansion. The full structural family sweep has now been run — 36 cells over attachment x relay depth x tag style x distractor count — and the two formerly missing baselines are in the panel. The sweep's headline is a negative result for FIBER: a plain backward walk over the directed factor edges, closure first, is admissible in all 36 cells at exactly FIBER's fact count, so on this family admissibility and cost cannot distinguish the compiler from that walk; the fixed-basis embedding retriever, by contrast, fails every camouflaged cell at the tight budget. Full analysis: docs/FINDINGS.md. How much of the blueprint the workspace actually covers, and which sections have nothing standing in for them: docs/COVERAGE.md. The crate layout and the blueprint path: docs/ARCHITECTURE.md.

The Rust glioma research engine tracks executable capabilities in docs/glioma/PROGRAM_PLAN.md. P10 F32 adds longitudinal replication transport through analyze_glioma_longitudinal_transport and the MCP tool glioma_longitudinal_transport_analyze: studies must bind the same estimand and effect unit and cover one shared time grid, with incomplete studies excluded rather than imputed. Different model systems incur the maximum transport gap and are included only when the request explicitly allows that maximum; signature similarity weights admitted sources. Effects are bounded to ±1,000,000,000 milli-units before fixed-point pooling. Its gate and threshold policy is explicitly repository-defined because this checkout has no configured detailed source blueprint for that slot.

P10 F03 plans prospective resolvers for conflicting preclinical claims through analyze_glioma_prospective_contradiction and MCP glioma_prospective_contradiction_plan. Rival statements, evidence reports, design reports, and local artifacts remain digest-bound. Evidence is counted by independent group, shared support across rivals cannot establish exclusive support, and low-quality, null, and unresolved findings remain visible. Resolver prediction intervals must separate a rival pair and pass feasibility, risk, and budget gates. A plan_ready result means the bounded portfolio covers all declared rival pairs; it does not mean an experiment ran or a claim was resolved. Its contract is repository-defined because this checkout has no detailed P10-F03 blueprint configured.

P10 F02 compares independent study effects through analyze_glioma_multistudy_concordance and MCP glioma_multistudy_concordance_analyze. It binds an exact estimand and unit, reports every eligible same-model pair by interval overlap and confident direction, and keeps cross-model pairs incomparable. It retains null, unresolved, and low-quality studies without imputation or pooling. The contract is repository-defined because this checkout has no configured detailed P10-F02 blueprint.

P10 F04 audits registered-outcome reporting through audit_glioma_registered_outcome_reporting and MCP glioma_registered_outcome_reporting_audit. It compares local registry protocols with reviewed result-report metadata, marks overdue missing primary outcomes, and separates incomplete, ambiguous, unregistered, and not-yet-due outcomes. It accepts no effect values and treats reporting differences as review signals rather than proof of publication bias or misconduct. Its contract is repository-defined because this checkout has no detailed P10-F04 blueprint.

P10 F05 adds build_glioma_registered_outcome_record and MCP glioma_registered_outcome_record_build for a single digest-bound study outcome. P10 F06 adds build_glioma_registered_outcome_evidence_panel and MCP glioma_registered_outcome_evidence_panel to combine compatible records across independent studies without pooling effects. The companion glioma_registered_outcome_sensitivity_analyze route evaluates caller-declared missing-result ranges as separate scenarios. These contracts are repository-defined because this checkout has no detailed P10-F05/F06 blueprints.

Autonomous agent process boundary

The Rust runtime also exposes an opt-in Docker-backed command boundary for callers that need an explicit Linux process boundary. It is not connected to agent tool dispatch or the trial provider; its limits and integration boundary are documented in docs/OCI_SANDBOX.md.

The Python SDK includes a secret-safe operator entry point for the autonomous brain:

cd python
python -m prism_sdk catalogue
python -m prism_sdk evidence-plan --domain science
python -m prism_sdk provider-status --provider openai

For keyless local development, the same boundary supports an explicit credentialless fixture: python -m prism_sdk provider-status --provider local and run --provider local --model local-model use the runtime's bounded in-memory transport; no key or network provider is contacted. For actual local-model inference without an OpenAI key, use the first-class Ollama preset: python -m prism_sdk provider-status --provider ollama (default endpoint http://127.0.0.1:11434/v1), then run with --provider ollama --model <installed-model> and --approve-provider-call. The Ollama path is explicit, credentialless, OpenAI-compatible, and loopback-by-default; it fails closed when the local server is unavailable rather than falling back to a synthetic response. For a grounded research pass, LocalNeurosurgicalAgent.grounded_real_data_research() / grounded_public_literature_research() (Python) or groundedRealDataResearch() / groundedPublicLiteratureResearch() (TypeScript) composes a bounded real-data or six-specialty PubMed context, an explicitly approved credentialless local-model call, and the matching authoritative draft-claim audit. It accepts Ollama or another caller-registered local provider, returns context/bundle digests and structured claims, and remains grounded_for_human_review; a provider outage fails closed and never produces synthetic evidence. Before the authoritative draft audit, the bridge requires every model citation to be present in the exact bounded context it received; unseen-but-valid snapshot records are rejected rather than treated as grounded. Set tool_loop=True (Python) or toolLoop: true (TypeScript) to expose bounded, read-only, credentialless snapshot tools to the local model: row search, a ClinicalTrials.gov trial-landscape view (neurosurgery_real_data_trial_landscape_view), and a cBioPortal/GDC molecular-coverage view (neurosurgery_real_data_molecular_coverage_view). Calls are capped by explicit turn and call budgets, recorded as a sanitized tool_trace, and their returned citation identities are added to the closure check; the final audit widens only to the same source facets, never to a network or patient-data tool. The trial and molecular views, plus the comparative neurosurgery_real_data_cohort_landscape_view, return descriptive aggregates plus exact rows for citation, not eligibility, efficacy, safety, treatment, or patient inference. A third review-queue view (neurosurgery_real_data_review_queue_view) exposes only explicit missing-link, abstract, date, or sample-count obligations for qualified human review; its bounded task rows are citation closed and never treated as clinical findings. The loop also exposes neurosurgery_real_data_reconciliation_view, which returns the canonical PMID/normalized-DOI crosswalk ledger (missing or shared identifiers, counts, and exact metadata rows) for human review. It never repairs or merges identifiers, fetches a source, or treats an identifier relationship as biological or clinical evidence; returned rows are citation-closed. It also exposes , a deterministic glioma topic-lane extractor covering integrated molecular identity, genomics, imaging, pathology, trials, outcomes, tumor microenvironment, and treatment-effect metadata. Topic membership is lexical and reviewer-facing—not relevance, evidence quality, biology, or clinical advice—and each returned record is citation-closed to the supplied real snapshot. The public-literature tool loop also exposes , which projects the real PubMed snapshot's missing DOI/abstract/MeSH/publication-type and duplicate- identifier obligations into citation-closed reviewer tasks. It is specialty-scoped, read-only, and never treats missing metadata as negative evidence. It also exposes , which returns bounded PubMed source-completeness and identifier-hygiene counts, review reasons, and exact metadata issues for the caller's fixed lane. Issues remain citation-closed reviewer work; they are never evidence rankings, negative findings, or clinical conclusions. The glioma loop also exposes , a bounded traversal of explicit study/profile/PMID crosswalks. Graph edges are identifier/provenance metadata—not causal or biological links—and every returned node is added to the citation closure set. It also exposes , which turns the validated snapshot and the fixed glioma request into a bounded next-evidence worklist. The worklist carries only local replay queries, source-linked metadata references, match counts, and explicit reviewer obligations; it never fetches a source, opens a patient asset, or authorizes a clinical action. The public-literature loop exposes the parallel for a fixed specialty lane. It compiles the checked-in PubMed snapshot into the same bounded, reviewer-owned local worklist while keeping PMID references and the boundary explicit; it never treats a planned query as proof that evidence exists. The glioma loop also exposes , a digest-bound inventory of source, record-kind, temporal, assay, and explicit linkage coverage plus bounded gaps. It preserves caller scope and omissions, but never converts coverage into a quality score or clinical claim. Both loops also expose , which projects the fixed lane's identity, spatial, functional, and temporal coverage states, missingness counters, and reviewer questions. It returns planning metadata only (never observation values or clinical inference), is provider-free/read-only, and rejects reports that drift from the caller's specialty lane. When the caller supplies an explicit UTC clock, the loops also expose a freshness view ( or ) that returns bounded source-age states and digest metadata. No host clock, fetch, quality inference, or synthetic fallback is used. The function accepts structured real-data facets (record kind, trial status/phase/study type and date bounds, molecular/genomic selectors, linked publication/MeSH selectors, and source IDs) or PubMed facets (publication type, MeSH term, and date bounds). A model may add a narrower facet or change lexical text, but cannot override a caller facet or increase its result limit; the specialty lane is never model-selectable. The lexical field is optional for facet-only searches; when omitted it uses the current bounded question (or caller text) as the selector. Tool results retain bounded source metadata such as trial status/phase/study type, molecular and genomic datatype labels, publication/MeSH labels, aggregate enrollment/sample counts, and abstract excerpts when present; recognized edges preserve the source crosswalk, while patient-level values are never projected. The glioma tool loop also offers a digest-bound identifier-reconciliation view for canonical PMID/normalized-DOI missing/shared rows; it is metadata-only human-review work and never repairs, merges, fetches, or clinically interprets identifiers. The operator commands expose the same mode with , , and ; persisted traces retain only search-text digests and structured facets. For these grounded helpers, an HTTP provider must resolve to loopback (, , or ); remote credentialless gateways are rejected before any evidence tool or network call. The bounded / helpers extend this into a finite autonomous fan-out: each pass re-renders the source context, audits its claims, and turns only model-reported unknowns into deduplicated metadata queries. The returned pass ledger, pending queries, termination reason, and loop digest are caller-owned and remain held for human review. The real-data loop also accepts the same structured registry, molecular, genomic, and PubMed facets as ; the normalized facet set is retained in the ledger and bound into its loop digest, so a restart cannot silently switch evidence slices. When a caller supplies an explicit lexical facet, it is used for the first pass; subsequent unknown-derived passes replace only that lexical selector while preserving every structured facet, so autonomous follow-up work changes the searched slice without widening its source boundary. Each persisted pass also carries a canonical digest of its claim payload; resume rejects missing or altered claims before another local-model call. The public-literature loop also accepts structured facets (specialty, publication type, MeSH term, inclusive date bounds, and limit); follow-up passes change only the lexical text selector, and the complete facet set is retained and resume-fenced. Pass a prior ledger as (Python) or (TypeScript) with a larger total pass budget to continue pending queries after a process restart; schema, provider/model, source bundle, and loop digest are revalidated before another local-model call. The provider-free capability router now maps the complete neurosurgical tool surface into both biomedical and neuroscience profiles: sanitized FHIR/DICOM imports, evidence programs and autonomous review waves, trial landscapes, molecular coverage, and the public-literature refresh/link/integrity queue, workbench, and portfolio tools. Natural-language routing remains lexical and abstaining; it only selects a reviewed capability and never authorizes a provider, patient-file access, clinical action, or external effect. / coordinates both planes in one source-separated digest: real glioma population evidence and specialty PubMed evidence retain independent loop/audit identities while their counts and pending work are aggregated for review. When both snapshots are supplied, the portfolio also runs the existing provider-free automatically. Its bounded exact PMID/normalized-DOI links, unmatched identifiers, and metadata mismatches remain a separate reviewer artifact and never imply cohort overlap, causality, or clinical applicability. The portfolio and CLI also accept an optional real, de-identified plus bounded query. The authoritative manifest projection contributes only asset-kind coverage, digests, and reviewer obligations; asset bytes, identifiers, and clinical values never enter the local-model context. The attachment is specialty-bound and remains a separate case-provenance plane from population and PMID evidence. The Python process boundary now exposes the same workflow as : it reads bounded, checked-in non-synthetic snapshots, runs Ollama on loopback (or an explicit in-memory fixture), and atomically persists a digest-bound answer/claim ledger. only continues a store whose question, provider/model, source selection, and child loop digests verify; no API-key argument or prompt exists on this command, and it remains research-only and human-review gated. When current public evidence is needed, add and/or together with . The refreshers use only the allow-listed credentialless public endpoints, validate each candidate snapshot, and atomically replace the selected files before the first model call. Refresh cannot be combined with because changing a source digest would invalidate the persisted loop; receipts are returned in and the run remains human-review gated. The same receipt is retained in the digest-bound output store and replayed on without re-fetching sources. Use (also supported by ) to apply a bounded JSON facet object to the glioma plane; the selected facet set is retained and resume-fenced. Use with either command to apply publication-type, MeSH, date-range, specialty, and limit facets to the PubMed plane; that slice is also retained and resume-fenced. For free-text routing, first runs the provider-free six-specialty intake, stops with when the routed snapshot is absent, and only then invokes the approved local model. Glioma requires the real glioma snapshot; the other specialties require the PubMed snapshot. The envelope preserves source-plane separation and an explicit human-review hold; it never falls back to synthetic evidence or emits clinical advice. Its is an atomic, digest-bound restart checkpoint: rechecks the question, route, source paths, provider/model, and bounded controls (only a larger pass budget is allowed), then hands verified child ledgers back to the worker. Checkpoints retain caller-owned research claims only—never keys, patient data, or hidden model state. To refresh that PubMed plane on any supported platform, use the credentialless NCBI boundary: . It retrieves six bounded specialty lanes, computes the Rust-compatible source and bundle digests, validates , and atomically replaces the snapshot only after every lane is linked and hash-checked. No API key, provider, patient data, or synthetic fallback is accepted. The grounded commands can perform that refresh inline with (and, for glioma, add ). Inline refresh is opt-in, refuses , and returns source digests and retrieval metadata in so the model never runs against an unreported corpus. For the complete glioma population plane, retrieves only aggregate metadata from ClinicalTrials.gov, NCI GDC, cBioPortal, NCI PDQ, and PubMed. It validates the Rust-compatible source hashes and required registry/genomic/portal/ guideline planes, then atomically installs a last-known-good snapshot; no patient rows, assay values, imaging, credentials, or synthetic fallback are fetched or retained. The real-data context also serializes each bounded reviewer obligation (task ID, source identity, and rationale), so an autonomous worker cannot mistake an unresolved metadata queue for a clean corpus.

Resumable evidence-backed provider calls in both SDKs now capture authoritative request, wire, credential, provider configuration, and transport identities before awaited caller callbacks. Observers and rehydrators receive detached projections; credential/provider graph checks repeat after callbacks and after the caller-owned dispatch transaction. That transaction must durably commit the private idempotency receipt with the provider_in_flight checkpoint before any transport call, and the graph is checked again before terminal settlement. This is a guarded same-process boundary, not an exactly-once claim: deployments still own authenticated durable receipt storage, provider-side idempotency, and uncertain-outcome reconciliation.

The Rust workspace also ships a dedicated, provider-neutral neurosurgical research agent in bioprism-neurosurgery. It routes de-identified glioma, cranial-base, craniofacial, encephalocele, spina-bifida and Chiari requests through deterministic read-only tools, emits explicit evidence gaps and a reproducible request digest, and always holds the result for human review. Each response carries a specialty-specific research profile covering identity, anatomy, time, evidence questions, confounders, and reviewer roles. It uses no OpenAI API or credential; see docs/NEUROSURGICAL_AGENT.md. A synthetic fixture exists only for offline contract tests; it is not used by the real-data path. The neurosurgery_evidence_audit tool adds per-specialty intake coverage for measured, unmeasured, uninterpretable, conflicting, and missing-provenance observation classes before the route executes. The neurosurgery_specialty_evidence_map tool expands that audit into four explicit dimensions for each lane—identity, spatial/anatomic, functional/intervention, and longitudinal context—so glioma, cranial-base, craniosynostosis, encephalocele, spina-bifida, and Chiari review cannot hide which domain inputs are absent, uninterpretable, conflicting, or provenance-incomplete. It is available through the Rust CLI (--specialty-evidence-map), MCP, Python specialty_evidence_map(), and TypeScript specialtyEvidenceMap(); it inventories supplied metadata only and never interprets imaging, pathology, genomics, or operative text. The map self-validates its digest and canonical source rows; mission audits rebuild typed glioma maps against the exact request and supplied snapshots before handoff. The neurosurgery_evidence_synthesis tool is the cross-plane handoff: it composes the redacted case audit with caller evidence, the validated real glioma population snapshot, and/or the validated six-specialty PubMed snapshot. Each plane stays separate, exact source identifiers and URIs remain inspectable, optional freshness reports are attached to the supplied bundle digests, and cross-bundle PMID correspondences are reported only as links (never as cohort or patient claims). Reference/query bounds, truncation, missing snapshots, and incomplete case coverage become explicit review items. The Rust CLI (--evidence-synthesis), Python evidence_synthesis(), TypeScript evidenceSynthesis(), and MCP expose the same no-key, network-free, read-only contract; raw case labels and values are never echoed. The same report is now included automatically by mission helpers: one-bundle missions expose the corresponding evidence plane, while dual glioma missions expose both planes and exact links in a single digest-bound handoff. Persisted synthesis reports self-validate their plane separation, lane counts, freshness bindings, asset/disposition projections, and provider boundary; mission audits also replay the report against the exact request and supplied snapshots so a structurally valid report cannot be rebound silently. The tool turns those explicit gaps into a bounded, source-linked caller handoff. It can query only a supplied local real-glioma or six-specialty PubMed snapshot, attaches stable source IDs/URIs for reviewer inspection, and keeps population/citation context separate from patient observations. It never fetches, invokes a model, writes state, or emits diagnosis, prognosis, treatment, triage, or procedural instructions; every plan remains held for human review. Rust, MCP, Python, TypeScript, and the offline CLI expose the same digest-bound contract with task/reference bounds. Every persisted plan now carries a and validates its task/source projections; mission audits replay the recorded bounds and local queries against the exact request and snapshot before handoff. The tool adds a deterministic, source-linked reconnaissance pass over the same validated snapshots: it groups exact lexical matches into specialty topic lanes, returns stable record IDs/URIs, preserves abstract availability and truncation, reports cross-topic overlap and explicit unknowns, and emits reviewer prompts. It does not rank evidence, summarize unsupported claims, call a model, or turn population literature into patient evidence; remains true. The Rust CLI (), Python , and TypeScript facades are parity surfaces for this report. Persisted briefs expose and ; mission audits replay the brief against the exact request and source snapshot, including topic counts, truncation, and source-link projections. The standalone projection is likewise digest- and topology-checked; its replay confirms every emitted node/edge came from the exact local glioma snapshot and persisted bounds. The shared now carries an and exact request replay guard; downstream evidence programs and research plans therefore inherit a tamper-evident intake coverage primitive for measured, unmeasured, uninterpretable, and conflicting states. The tool is the next autonomous worker wave: it turns the same explicit missing/uninterpretable/conflicting/provenance gaps into a bounded dual-plane worklist, querying only caller-supplied validated real-glioma and/or PubMed snapshots. Each step carries a source tag, trigger, deterministic digest, local match/truncation status, fallback-to-specialty-scan flag, and replayable references; missing sources remain explicit obligations. Rust, MCP, Python , TypeScript , and the offline CLI expose this provider-free surface. The lifecycle variants (, bounded , and , with matching Python and TypeScript methods) let a caller persist and resume a digest-bound checkpoint; changed request, query, or snapshot bytes are refused before replay. It never fetches, needs an API key, opens asset bytes, or promotes a population/citation match to a case finding, and remains true. drives the caller-owned checkpoint loop locally and writes no credentials or clinical state. It accepts plus the optional and , so the same worker can carry a real de-identified multimodal review projection and its persisted reviewer state through every wave. When a real de-identified is supplied, the acquisition report also carries the manifest report digest and bounded asset review items (missing source, digest, timestamp, uninterpretable, conflicting, or requested-class obligations). Start/advance/finish re-bind that digest on every replay, so a local worker cannot silently drop multimodal provenance while replaying population or citation queries. The offline CLI accepts the same projection with and alongside ; Python and TypeScript expose matching optional arguments on each lifecycle method. For real glioma research, use the provenance-bound public snapshot in and refresh it (without a provider key) with . The checked-in adds the real NCI GDC TCGA-LGG project (516 aggregate cases) alongside TCGA-GBM (617); it is generated by the same script with . The baseline remains unchanged for replay compatibility, while callers can opt into the broader glioma population bundle and its distinct source digest. The checked-in extended bundle also uses the broader real PubMed query under the stable source ID, so lower-grade and histomolecular terminology is not silently excluded from the citation plane. Each extended GDC project also carries aggregate file/data-type facets (for example somatic mutation, aligned-read, slide-image, transcript-fusion, and methylation availability) without exposing files, samples, or assay values. For an end-to-end candidate workflow, validates the baseline, refreshes a separate candidate from public endpoints, runs the core refresh audit, and writes a report without replacing the baseline; promotion remains an explicit reviewer action. It accepts the same , , and scope controls as the low-level refresh script, so the candidate audit can cover the wider real-glioma population without silently changing the baseline. The refresh script defaults to a bounded 20-record PubMed window and accepts for an explicit corpus size. and widen the real citation lane without losing query/source provenance; replacement of an existing snapshot is atomic and cleans its temporary backup after promotion. Source IDs are stable across retrieval dates; timestamps and content hashes carry freshness and change information without turning every refresh into a remove-and-add event. The tool is the restart-safe reconciliation layer for that workflow: give it two independently validated snapshots and it composes structural diff, coverage, freshness (when requested), review-queue obligations, and the research brief into one digest-bound report. It preserves stable source/record identity, emits explicit refresh-review reasons, and never accepts, merges, fetches, ranks, or writes a candidate snapshot. The Rust CLI (), Python , and TypeScript facades expose the same provider-free contract; human review remains required. For long-running work, the MCP tool provides digest-bound start/advance/ finish checkpoints so a caller can resume one read-only specialty tool at a time without hidden server state. Checkpoints also bind the canonical specialty route, session identity, event status, and terminal hold, so identity or route mutations fail closed before a resumed tool runs. Every terminal now carries a over its complete route, tool trace, evidence-gap projection, and nested provenance summaries. Rust callers can invoke for persisted-envelope checks and for exact request replay; session finish rejects a response that fails either structural envelope gate. Mission envelopes now also carry the same bounded plan, so a single provider-free mission exposes the specialty route, source-linked research plan, real-data/literature packet, and resumable acquisition worklist together without merging evidence planes. They also carry an : six protocol-defined review tracks per lane (for example glioma histomolecular identity, imaging phenotype, surgery/function, response endpoints, microenvironment, and trial design) are projected onto exact IDs in the attached real snapshots. and the Python / TypeScript facades expose the same agenda directly. Track matches are transparent lexical retrieval observations with bounded references, required observation classes, and specialist reviewer roles. Each track also carries metadata-only observation coverage copied from the typed intake audit (, , , or ), missing classes, and provenance gaps; this is a worklist signal, never a sufficiency score. Empty and truncated tracks remain unknown. The program is read-only, provider-free, network-free, synthetic-data-free, and human-review gated—it does not rank evidence, make a glioma classification, or emit treatment or operative guidance. When a persisted ledger is supplied with the manifest, its digest and pending/resolved counts are carried into both the evidence program and acquisition plan; stale or tampered reviewer state is refused. When a real de-identified is supplied, each track also joins its required observation classes to digest-only imaging, pathology, molecular, operative, functional, developmental, longitudinal, or anatomical coverage. , , and states make the next export/review obligation actionable without reading asset bytes; the optional flag is inventory metadata, not clinical sufficiency. Tracks also emit a deterministic for observation/provenance gaps and unresolved asset classes, giving a local worker explicit next metadata checks without inventing findings. Evidence-program reports self-validate their canonical tracks, coverage/count invariants, source references, freshness bindings, and digest; mission audits replay them against the exact request and supplied snapshots before handoff. Persisted case-asset projections expose and guards; synthesis, evidence-program, and acquisition joins refuse a tampered or request-mismatched report before it can enter a digest-bound handoff. This protects restart/review workflows without pretending that an upstream asset digest proves the asset's clinical truth. The offline CLI exposes the same pass with , and/or ; add to bound lane, track, reference, abstract, or freshness controls. is an all-six-lane query template for the checked-in PubMed snapshot. Mission envelopes also include a final receipt. It verifies specialty/status identity, request digests, real/public snapshot digests, required report-plane presence, the case-asset-to-synthesis and case-asset-to-evidence-program bindings, and the provider-free human-review boundary. is an assembly/provenance invariant only; it is not a clinical readiness or evidence-quality score. Persisted mission envelopes now have a single replay gate: Rust checks the terminal response/session chain and all nested receipts without inputs, while rebuilds the mission audit against the exact request and caller-owned snapshots. Changed request or snapshot bytes fail closed before a worker can reuse the packet. The MCP tool accepts ; Python , TypeScript , and the offline CLI expose the same no-key, read-only replay check. When a mission carries a DICOM or FHIR receipt, exact replay additionally requires the original sanitized metadata export ( or on the CLI, or the matching case-import object on the MCP validation call); otherwise validation fails closed rather than treating receipt-shape integrity as source replay. The MCP tool exposes all specialty profiles and read-only tool specs before execution, while searches the validated public bundle by stable record text, cBioPortal molecular-profile modality, trial status, exact registry phase or study-type facets, inclusive registry update-date bounds, exact GDC file facets, case-insensitive PubMed / indexing facets, inclusive PubMed / bounds, record-kind/source facet, or explicit relationship facet, including PMID/title/DOI/abstract/MeSH matches from the PubMed lane. These indexing facets narrow literature metadata but do not act as study-quality scores. Genomic-project hits additionally expose aggregate GDC file/data-type facets when present, keeping modality availability source-linked without returning files, samples, or molecular values. PubMed hits carry bounded source-text excerpts and indexing tags for reviewer inspection. Clinical trial hits also preserve optional ClinicalTrials.gov study type, aggregate enrollment target, intervention names, phases, and last-update date; portal-study hits preserve optional public sample counts; PubMed hits preserve publication dates. Partial PubMed chronology (year-only or month-only source dates) remains missing rather than being padded with an invented day. Missing upstream fields remain absent rather than guessed. Hits also carry explicit study↔profile/publication relationships so a caller can traverse the evidence graph without inferring links from prose. Molecular-profile rows describe available assay modalities only; they never expose mutation, expression, or patient-level values. Responses also include deterministic counts of profile modalities and explicit relationships so a reviewer can see assay coverage and cross-source connectivity before inspecting source-linked metadata. PubMed hits are metadata-only and require reviewer verification before substantive use; the Rust summary also exposes a PMID crosswalk to flag unmatched portal citations without inferring cohort identity. The Python SDK exposes the same provider-free lifecycle through , including bounded session iteration, catalogue discovery, and public record queries for UI or worker integrations. Persisted real-data and PubMed query results now expose plus exact replay; mission audits invoke those gates so a changed query, hit list, or count projection cannot be smuggled into a persisted mission. adds a digest-bound, provider-free ClinicalTrials.gov metadata reconnaissance over the same validated snapshot: bounded status, multi-label phase, study-design, intervention, update-date, source, missingness, and truncation projections. It never ranks trials or infers eligibility, efficacy, safety, outcomes, or patient-level meaning; multi-phase rows are counted explicitly rather than collapsed into a misleading trial total. adds the complementary cBioPortal availability ledger: exact alteration-type/datatype facets, per-study profile counts, analysis-visible and patient-level metadata flags, description coverage, explicit missing alteration/datatype counts, aggregate GDC project file/data-type facets, and explicit row/study/facet truncation or missing-facet reasons. It inventories only public assay metadata already in the snapshot—no mutation/expression values or sample identifiers—and is digest-bound, replayable, provider-free, network-free, and human-review gated. The canonical evidence packet includes this ledger automatically alongside the trial and comparative cohort landscapes. adds the comparative genomic-project view used by the autonomous loop and is included automatically in newly generated evidence packets and missions. It compares the source-linked TCGA/GDC projects already present in the validated bundle, reporting aggregate released-case inventory and per-project file/data-type availability with explicit truncation and missing-metadata reasons. The view is read-only, provider-free, and metadata-only: rows are citation surfaces, counts are descriptive planning context, and it never opens files, exposes samples or molecular values, merges cohorts, or makes a clinical claim. For natural-language entry, (and the Python / TypeScript facades) performs deterministic lexical routing into the six closed specialty routes. It returns bounded candidates, abstains on weak or ambiguous wording, and lists caller-supplied evidence snapshot classes, reviewer roles, and next research actions. The question is represented in the returned plan only by a SHA-256 digest; scores are routing units, not probabilities or clinical risk, and an explicit specialty is only a research-routing override. This makes free-text intake useful without adding a model provider, credential, network, patient-file, diagnosis, or procedure capability. The closed vocabulary includes specialist subtopics rather than only disease names: glioma histomolecular markers and treatment-effect terms; petroclival/cavernous-sinus and cranial-nerve topics; craniosynostosis suture and syndromic terms; encephalocele variants and CSF rhinorrhea; spinal dysraphism, tethering, and neurogenic-bladder terms; and Chiari measurements, cine-MRI, and CSF-flow terms. These are routing labels only and never become inferred findings. (and / ) composes that planner with a guarded research-only mission: ambiguous questions return a digest-only abstention, selected glioma routes require the validated real glioma snapshot (with PubMed as optional supplement), and the other specialties require the validated PubMed snapshot. Executed results contain no raw question or request payload and remain provider-free, network-free, read-only, and held for human review. Callers may optionally include a de-identified with observations, provenance, and evidence. It is validated before any bundle query and carried into the guarded route, so a real case can be reviewed without the old empty-case fallback; the case payload is never echoed in the intake envelope. If omitted, the mission still runs the route but exposes the resulting observation gaps for human follow-up. An optional plus carries real, de-identified multimodal asset metadata into the nested mission. The manifest is digest-bound, requires explicit asset states, and never opens bytes; use the same pair with the Rust CLI's . The intake mission also accepts and directly (Python , TypeScript , MCP fields, or CLI /). These imports take the same independently validated, digest-only route and may be combined with each other, but not with a second asset manifest. An already persisted ledger can be supplied in the same intake mission call (Python , TypeScript , or the MCP field). Its report digest and reviewer counts are validated before evidence handoff and rebound into synthesis, evidence programming, acquisition, and the final audit; it never changes the manifest or creates a clinical conclusion. If a real case is exported through FHIR, (Rust , Python , TypeScript , or CLI ) projects a caller-sanitized FHIR into that same digest-only asset boundary. The import requires , , bounded / metadata, and an explicit asset-kind/status/ provenance hint; it rejects identifiers, patient references, narratives, codes, measurements, and raw text. The Bundle is never echoed or interpreted, unclassified resources become reviewer tasks, and the report can be replayed against the exact request, Bundle, and hints without an API key or network. If the imaging archive exports standard DICOM JSON, (Rust , Python , TypeScript , or CLI ) projects only bounded series-level metadata such as modality, body region, study/series/SOP UID digests, dates, descriptions, and series number. It accepts one dataset or an array (up to 512 datasets and 4 MiB of metadata), refuses patient-identifying tags and , ignores unknown/private tags, never opens DICOM bytes, and never interprets an image. Missing SeriesInstanceUID, acquisition dates, modality, body region, and object-byte SHA-256 digests become explicit review obligations; the digest-bound report is replayable, non-synthetic, provider-free, network-free, and human-review gated. For a single end-to-end handoff, (Rust , Python , TypeScript , or CLI ) composes that real metadata projection with validated real glioma/PubMed records, evidence synthesis, the six-track review program, and a resumable local acquisition checkpoint. Every nested report is bound to the same request and DICOM manifest digest; the output remains provider-free, network-free, read-only, non-synthetic, and held for human review. The repeatable PowerShell wrapper validates inputs, runs the offline CLI, and writes a caller-selected report without promoting data or retaining credentials. For a mission-level glioma dossier, pass the same DICOM import as to (Python , TypeScript , or CLI together with ). The mission carries the DICOM receipt and verifies that its manifest digest is rebound through synthesis, evidence programming, and acquisition; this convenience lane is real-glioma-only and can be composed with a sanitized FHIR import for a multimodal digest-only manifest, but not with a second asset manifest or disposition. For a repeatable local run of that mission-level lane, use ; it validates the DICOM/manifest/synthesis bindings and refuses a nonzero mission audit before writing the report. The same mission envelope accepts a sanitized FHIR metadata import as (Python , TypeScript , or CLI ). It works with a real glioma bundle, a cross-specialty PubMed bundle, or both; the FHIR receipt's digest-only manifest is rebound through the same synthesis, evidence-program, acquisition, and mission-audit planes. FHIR resources and clinical values are never returned or interpreted. FHIR and DICOM imports may be supplied together; their independently validated digest-only projections are unioned into one multimodal manifest while both child receipts remain visible. A separate asset manifest or disposition ledger cannot be mixed into an import-backed mission. Intake missions and portfolios also accept an optional caller-clocked policy inline or via the CLI flag. Resulting real/PubMed freshness reports are digest-bound; omission means freshness is unclaimed and the server never consults its own clock. When it executes, only the planner's matched closed-vocabulary terms become bounded local real-data/PubMed filters; the original free text is never echoed into those reports. An explicit specialty-only hint uses that lane's canonical corpus term (for example for glioma) when no lexical terms were matched. The same intake orchestration is available without MCP: pipe a flat JSON intake query to or and pass the checked-in and/or snapshots. These CLI modes perform the same validation, provenance checks, and human-review hold with no provider, API key, or network. For a repeatable worker that refreshes both public bundles into non-promoted candidates, audits their drift, and then runs the portfolio against the validated candidates, use . It emits one machine-readable worker envelope and never promotes a candidate snapshot. Supply to bind a caller-owned source-age clock, or plus the optional to carry real, de-identified multimodal provenance into a selected-lane portfolio. A persisted ledger can accompany that manifest and is replayed into the nested mission's synthesis/acquisition audit; all-six-lane portfolios refuse both the manifest and its ledger. The PowerShell worker accepts the same ledger through . For cross-specialty reconnaissance, (and / ) fans those filters across one selected lane or an explicit all-six-lane portfolio. Each lane remains independent and source-linked; an all-lane portfolio requires both the PubMed snapshot and the real glioma snapshot because glioma is part of the requested scope. A selected-lane portfolio can carry the metadata-only case-asset manifest pair; an all-lane portfolio refuses a single-specialty asset attachment, including its reviewer ledger. A selected-lane call may carry through the nested mission. The worker verifies the selected lane's nested evidence-synthesis asset digest and coverage counts before emitting its envelope. Use (or / ) when a reviewer needs the explicit, bounded study/profile/PMID crosswalk: it returns source URIs, root traversal, component/isolate counts, omissions, and a digest without inferring biology, causality, or clinical action. A complementary report audits the same real snapshot by source, record kind, trial-update/publication-date axis, assay modality, abstract availability, and explicit study/profile/PMID linkage gaps. It preserves missing dates, exposes retrieval metadata, and binds a coverage digest; it does not score freshness or evidence quality, merge cohorts, or make clinical claims. Coverage reports expose and ; mission audits use the exact replay check before a local worker can consume the coverage plane. is the companion cross-source identifier ledger. It replays one validated snapshot and reports only exact PMID/normalized-DOI findings: portal PMIDs missing from the local literature window, PMIDs shared by multiple portal studies, and DOIs shared by multiple literature rows. Counts remain visible when findings are truncated, identifiers are never merged or repaired, and any finding sets ; this is metadata review work, not a biological, clinical, or evidence-quality conclusion. It is available as , , and with no provider, network, or API key. Real-data missions also attach a bounded inventory over the ClinicalTrials.gov rows and a inventory over cBioPortal assay/profile metadata. Both are digest-bound to the same validated snapshot, preserve truncation and missing metadata as review obligations, and never rank trials, infer eligibility, expose patient-level assay calls, or make efficacy, safety, diagnostic, prognostic, or treatment claims. is the explicit age posture companion: provide a caller-owned UTC timestamp and policy to classify each source as , , or . A future-dated source forces ; age is never treated as evidence quality, applicability, or clinical relevance. The report is digest-bound, read-only, provider-free, and available for the cross-specialty PubMed snapshot as . Freshness reports expose and exact replay methods for real-glioma and cross-specialty snapshots; mission audits refuse a stale or future-dated posture that has drifted from its caller-supplied clock or source bundle. Real-data missions include the ordered, source-linked , coverage audit, bounded and inventories, metadata review queue, bounded evidence packet, explicit evidence graph, digest-bound , and automatically alongside any optional bounded record query and the resumable human-review workflow. The plan and queue turn explicit intake gaps into caller-owned next-review tasks; the packet/context are source-addressable input for a caller-owned local model or reviewer. Neither is a model invocation or clinical conclusion. Public-literature missions carry the corresponding bounded PMID evidence packet and automatically run the lane-scoped before packet/brief/context handoff. Missing DOI, abstract, publication-type, and MeSH metadata plus duplicate identifiers remain explicit review obligations; they are never treated as negative evidence. The same mission envelope carries a bounded with stable source-linked reviewer tasks so real metadata gaps become actionable review work without a provider key or clinical interpretation. That queue exposes and , keeping persisted task rows tied to the exact integrity audit and public snapshot. The companion joins each selected lane's closed specialty profile (identity, spatial, temporal, evidence-question, confounder, and reviewer-role axes) to exact snapshot coverage, abstract availability, metadata gaps, and integrity-review counts. It is navigation metadata rather than a readiness or quality score: lanes are never ranked, missing fields are never imputed, and no diagnosis, prognosis, treatment, triage, or procedural action is emitted. Use with a JSON query on stdin, Python , or TypeScript ; public-literature missions attach the request-specialty workbench automatically. The integrity audit, workbench, matrix, and portfolio reports expose digest/exact-replay checks; persisted multi-lane review state must be replayed against the same public snapshot before use.

The neurosurgery_public_literature_portfolio pass composes that workbench into one bounded multi-lane handoff: every selected specialty receives an exact lexical query result, its profile and coverage lane, and a stable reviewer queue (all six lanes by default). It uses only the validated real PubMed snapshot and preserves explicit hit, review-item, omission, and truncation counts. The portfolio is provider-free (provider: none, network: false, synthetic_data: false), does not rank evidence or infer a clinical conclusion, and never fetches URLs, opens credentials, or writes durable state. Use --public-literature-portfolio <public> with JSON on stdin, Python public_literature_portfolio(), or TypeScript publicLiteraturePortfolio().

Observations may also carry caller-supplied UTC observed_at values and de-identified timepoint labels. The neurosurgery_evidence_audit response (and --temporal-audit CLI mode) now includes a digest-bound temporal_alignment report with ordered timestamps, same-time observations, undated records, required specialty classes without dates, and caller-order inversions. This is an explicit longitudinal metadata audit—not a progression, response, prognosis, diagnosis, or treatment model; dates are never inferred from free text. For refresh monitoring, neurosurgery_real_data_diff compares two validated snapshots and exposes added, removed, or changed public records plus source-metadata changes by stable identifier; it never copies abstracts, scores freshness, merges cohorts, or makes a clinical claim. Diff reports expose validate_integrity() and validate_for_inputs(...) so refresh decisions can be replayed against the exact before/after snapshots. The composed refresh audit applies the same nested integrity and exact-replay checks across the diff, coverage, freshness, review queue, and research brief planes. neurosurgery_real_data_review_queue then derives a bounded, digest-addressed human-review queue from explicit snapshot gaps (missing crosswalks, unlinked citations, absent/clipped abstracts, unknown registry dates, or unknown sample counts) without imputing values or assigning clinical urgency. neurosurgery_real_data_review_disposition applies caller-owned reviewed, unresolved, or not_applicable state to emitted queue tasks, verifies the queue digest, and preserves omitted or undecided obligations as pending; it never edits source facts or produces a clinical conclusion. neurosurgery_real_data_evidence_packet composes the validated summary, coverage, explicit crosswalk, bounded source-linked query hits, canonical ClinicalTrials.gov trial landscape, and review queue into one packet digest for a local model or human reviewer; nested omissions and unknowns remain visible. The packet also carries the canonical cBioPortal molecular-availability ledger (per-study profile/modalities, explicit description gaps, and boundedness) so a local worker can see what assay metadata is actually present before reasoning. Both real-glioma and cross-specialty literature packets accept an optional freshness query with an explicit UTC as_of and return the digest-bound current/stale/future-dated source posture when requested; omitting it never invents a clock or claims that the snapshot is fresh. The real-glioma packet also carries a canonical PMID/normalized-DOI reconciliation ledger; missing or shared identifiers remain explicit provenance-review obligations before a local model can rely on the crosswalk. neurosurgery_real_data_reasoning_context renders that packet into a deterministic, bounded local-model context with digest-bound headers, source-addressable record blocks, optional untrusted abstract excerpts, and explicit character/query omissions. It never invokes a model or turns source text into a clinical conclusion. The context envelope exposes and as well; a worker must verify the persisted context against the exact snapshot before handing it to a local model or reviewer. Mission construction binds its context query to the same record and freshness scope used by the evidence packet. The packet itself exposes and ; nested coverage, graph, query, trial-landscape, molecular-coverage, reconciliation, queue, and freshness projections must retain one bundle digest and exact persisted bounds before a local worker can consume the handoff. The packet schema is ; older , , and artifacts must be regenerated because the canonical trial, molecular, and identifier reconciliation ledgers are now part of the digest-bound handoff. The cross-specialty PubMed packet and reasoning context expose the same integrity and exact-replay methods, so non-glioma lanes receive identical stale/tamper protection before local-model or reviewer handoff. The six-lane literature matrix, workbench, and portfolio reports are likewise digest-checked and replayable against the supplied snapshot; a multi-lane handoff cannot silently drift from its per-lane query, review queue, or specialty profile. composes the same packet into a deterministic, restart-safe provenance → completeness → context review wave. It emits only source-addressable metadata tasks (including explicit freshness-policy checks, stale-source refresh actions, and bounded-projection expansion holds when a caller supplies an age clock or small result limits), accepts a persisted human disposition report to resume work, keeps truncation and unresolved items visible, and ends at a human-synthesis gate. If caps the workflow queue itself, the same hold remains active so omitted context actions or the human-signoff gate cannot be mistaken for a complete handoff; autonomous means orchestration, not clinical prioritization or approval. Persisted waves expose and , verifying packet binding, action dependency closure, bounded truncation, open-obligation counts, and exact snapshot replay before another worker resumes them. adds a lane-specific identity/spatial/functional/temporal coverage map for glioma, cranial base, craniosynostosis, encephalocele, spina bifida, and Chiari. It turns the generic route into an explicit specialist inventory while retaining source IDs, timestamp coverage, conflicts, and uncollected dimensions; it never interprets observation values. is the companion local-model boundary: it requires every caller-owned draft claim to cite a record emitted by that packet, blocks patient-case and clinical- action posture, and labels accepted rows without pretending to fact-check or clinically interpret claim text. A separate six-specialty PubMed snapshot at covers glioma, cranial base, craniosynostosis, encephalocele, spina bifida, and Chiari. Refresh it with , validate it locally, and pass it to / or the MCP tool. The route attaches only the requested specialty lane as unverified citation metadata and still ends at human review; it never converts abstracts into patient findings or clinical actions. The checked-in snapshot has 145 source-hashed records and 138 abstracts across the six lanes. Each PubMed lane uses a stable source ID across refreshes, keeping retrieval time and content change auditable without manufacturing a new source identity on every run. Python and TypeScript expose for a live, explicitly approved acquisition of 1--6 fixed lanes. Its pure preflight binds the lane/query set, parser surface, and transport configuration; execution permits exactly one PubMed ESearch → ESummary → EFetch sequence per lane and enforces the resulting request ceiling plus per-response, aggregate-byte, tree, record, and bundle bounds. EFetch accepts and strips only the allow-listed external NLM DTD declaration before parsing and projection; entity declarations, alternate doctypes, and malformed XML fail closed. The durable receipt retains source IDs, digests, and counts but excludes query strings and article content. Deployments can supply their separately registered NCBI and developer only as a pair; both are placed on every request while durable artifacts retain only a configured flag and integrity digest. The one-lane / helpers connect a reviewed single-lane plan to the generic evidence runtime's acquire/project callbacks, validate the transient bundle and receipt, and project only digest metadata without widening the approved source plan. This is the first reviewed live-retrieval adapter, not general web research or evidence validation; broader sources, shared coordination, uncertain-call reconciliation, evidence-quality enforcement, and independent claim-integrity review remain deployment work. Python and TypeScript also expose for two fixed public catalogue lanes (, ). After a network-free and literal dispatch approval, it issues exactly two bounded GETs per selected study: the study summary and one sorted molecular-profile page capped at 128 rows. It keeps only allow-listed study descriptors, aggregate sample counts, and profile metadata; sample, patient, clinical, and molecular-value routes are outside the plan. Plans bind the endpoint, fields, pagination, transport identity, and fixed study set. Receipts bind the normalized catalogue and counts; autonomous evidence receives only bundle/source digests. Full profile pages are refused as potentially truncated, missing counts remain unknown, and source metadata stays transient. The public API is documented as beta and can change; the adapter fails closed on identity, response, or profile-schema drift. See the official and . It supplies catalogue coverage, not evidence quality or exhaustive glioma discovery. For a safe before/after refresh, use : it validates the baseline, creates a separate candidate, runs the cross-specialty refresh audit, and leaves promotion to an explicit human reviewer. bridges the real glioma literature index to a selected public-literature lane using exact PMID and normalized DOI identifiers only. It makes the 12 known glioma overlaps, unmatched bounded windows, metadata field drift, and identifier conflicts explicit; it does not infer cohort identity, evidence quality, biology, or clinical meaning. Use , Python , or TypeScript for this provider-free, read-only human-review handoff. is the pre-synthesis corpus gate: it audits selected lanes for missing DOI/abstract/publication-type/MeSH metadata and duplicate normalized DOIs, returning source-addressable issue rows and explicit truncation. Use , Python , or TypeScript ; it reports completeness obligations only and never silently repairs, merges, scores, or clinically interprets records. provides the lane-complete reviewer navigation view over the same validated snapshot, preserving per-lane profiles, source IDs, record/abstract counts, and bounded integrity obligations. It is deterministic and provider-free (, , ) and does not turn coverage into a clinical score. Each lane also reports non-exclusive metadata-derived design strata (human-indexed, animal/preclinical, in-vitro/cell-line, review/synthesis, imaging/diagnostic, surgical/procedural, developmental/genetic, outcome/follow-up, and interventional) with exact PMIDs. Overlap and unclassified rows are review obligations, never evidence-quality grades or clinical conclusions. For restart-safe cross-specialty refresh review, compares two independently validated snapshots and composes a bounded source/PMID diff, the six-lane coverage matrix, and optional caller-owned freshness posture. It reports changed field names and stable/added/removed identities without copying abstract text, and never fetches, merges, accepts, or promotes the candidate. Use the Rust CLI flag , Python , or TypeScript ; the report is provider-free (, ) and remains a human-review handoff. The and MCP tools (also available as / and / ) make that corpus a bounded local-model handoff: packet records are emitted with PMID/source links, and every accepted draft claim must cite one of those emitted PMIDs. The result is only a structural posture; it is not abstract fact-checking, study-quality assessment, or a clinical conclusion, and it requires no OpenAI key. (also / ) renders that packet into bounded, source-addressable context for a caller-owned local model. Abstract excerpts remain explicitly untrusted, citation/character omissions are reported, and no provider, network, API key, or clinical interpretation is involved. Public-literature missions include the same envelope automatically. adds a lane-complete reconnaissance pass: one bounded query can fan out across selected specialties (or all six), while each lane keeps its own packet, PMID identities, empty/truncation state, and digest. It reports corpus shape only; it does not merge cohorts or infer cross-specialty biology. Both SDKs also expose the typed glioma panel vocabulary, so assay provenance and explicit missingness can be submitted without inventing a diagnostic label. The dependency-free TypeScript SDK exports the same facade from . For marker-level grounding, (Rust , Python , TypeScript ) maps requested IDH1/IDH2, MGMT, EGFR, TERT, H3, 1p/19q, methylation, and related marker terms onto exact records in the validated real-glioma and PubMed snapshots. It preserves caller assay missingness, reports truncation and zero-hit review obligations, and never treats a literature or population match as a patient result. The map is read-only, provider-free, network-free, and human-review gated. For real case handoff, (Rust , Python , TypeScript ) accepts a caller-owned, de-identified manifest for imaging series, pathology, molecular assays, operative notes, functional/developmental assessments, longitudinal outcomes, and anatomical models. Every entry is bound to a SHA-256 content digest, source kind, explicit observation state, optional modality/body region, and an observation timepoint; the projection never opens asset bytes, extracts identifiers, calls a provider, or interprets a scan/report. It refuses synthetic manifests, direct identifiers, malformed digests, duplicate asset IDs, and specialty drift, then emits deterministic per-kind coverage, missingness, provenance gaps, and a bounded review queue. This is a real-data intake/provenance seam—not a pixel/content clinical parser—and it always returns a human-review hold. When an archive can export standard DICOM JSON, use (Rust , Python , TypeScript , or CLI ) to project bounded series metadata. It refuses patient-identifying tags and , ignores unknown/private tags, never opens DICOM bytes, and turns missing UIDs, dates, modality, body region, or object-byte digests into explicit reviewer obligations. Reports now include an ordered, typed research worklist that separates missing caller evidence from uninterpretable or conflicting evidence and names the observations and reviewer roles needed for the next review step; it never schedules a test or recommends care. Its / helper composes catalogue discovery, optional real-bundle querying, optional case-asset provenance, and the bounded session lifecycle; glioma missions require a validated public bundle and always return a human-review hold. Pass the de-identified manifest with / (or the CLI flag) to carry metadata-only multimodal provenance in the mission; the field binds that projection into the same ledger; also exposes asset/observation/provenance counts, requested kinds still missing, review-item counts, and truncation without exposing asset bytes or identifiers. When present, carries the bounded digest-only review obligations themselves. Pass a persisted with the mission to carry reviewer progress forward; the mission audit verifies its manifest/synthesis bindings and leaves unresolved or undecided items visible as workflow state. (Rust , Python , TypeScript ) applies caller-owned , , or state to returned review-item sequence numbers. The resulting ledger is bound to the manifest , canonicalizes decision order, rejects duplicate/unknown sequences, and keeps omitted or undecided obligations pending. It stores no local IDs, asset bytes, secrets, clinical meaning, or external workflow state. The direct evidence-synthesis MCP tool and SDK facades accept that persisted ledger alongside the same manifest projection; digest/count mismatches fail closed, while the resulting synthesis exposes only the disposition digest and pending/resolved/unresolved counts for resumable human review. The offline CLI supports the same