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 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-cou