Documentation map
The repository ships over ninety documentation files. These are the load-bearing ones — start here, follow references outward.
ARCHITECTURE.md
How the pieces fit and why the boundaries fall where they do. States the one idea — context
assembly is a compiler pass — then walks the layer diagram from foundation (ids,
scope) up through world/storage, compilation (fiber,
section), composition, evaluation, and the agent surface. Ends with a table of
cross-crate invariants and the exact test that pins each one, including “protected closure is
computed before any relevance step” and “a budget smaller than the closure fails rather than
truncating”.
FINDINGS.md
The measured results, including the ones that do not favour FIBER. Documents the three-way tie on the reference world, calls the distribution's own earlier baseline script a strawman, presents the discriminating world where FIBER is the cheapest admissible strategy (11 facts), and closes with an explicit section on what the results do not establish, plus defects found in the upstream v0.6 distribution. Every number is asserted by a test.
BASELINE_COMPARISON.md
The full equal-engineering comparison table for the reference world
(radiogenomic-integrity-demo-v1, 761 facts): ten strategies, each scored for
soundness, protected closure, and admissibility, with method definitions. The cheapest admissible
strategy on this world is the 5-hop graph walk at 11 facts — not FIBER.
DISCRIMINATING_COMPARISON.md
The same panel run on generated-discriminating-v1 (762 facts), where the methods
separate: the graph walk becomes unsound at every compact depth, BM25 reaches the right verdict
from an incomplete 91% closure — a contract violation that guessed right — and FIBER is the only
strategy that is sound, complete, and compact.
COVERAGE.md
Blueprint coverage, measured rather than asserted: which of the blueprint's 973 content modules the workspace actually cites, with the criterion stated honestly — a citation means “someone has taken a position on this”, never “this is done”. Also delimits what the MCP integration layer does and does not turn into implemented workspace code.
ADR-001-language-strategy.md
The language-strategy decision record, measurement first: scaling runs from 1,000 to 1,000,000 facts showing that parsing dominates compilation and that both scale with world size while the output stays at 11 facts. Concludes the optimisation that matters is architectural — a content-addressed indexed store — not a rewrite, and defines where Python and TypeScript sit relative to the Rust kernel.
HTTP_API.md
The bioprism-api gateway: a deliberately bounded local HTTP layer over the MCP
dispatcher, with bearer authentication, per-route documentation (capabilities, missions, events,
webhooks, operations snapshots and gates), explicit size limits, and root confinement inherited
from the MCP server.
PYTHON_SDK.md
The dependency-free Python SDK (python/prism_sdk): a standard-library client for
the MCP server and HTTP gateway that keeps Rust authoritative for canonical bytes, oracle
decisions, and release gates. Covers the typed request/report models, the artifact registry and
lineage surfaces, and the sync/async workspace facades.
TYPESCRIPT_SDK.md
The dependency-free TypeScript Fetch SDK for browser and Node consumers: typed API errors, request/response bounds, SSE cursor parsing, and the evidence and provider-connector contracts — including the rule that credential material is rejected at the boundary and readiness always stays a caller decision.
CAPABILITY_DASHBOARD.md
The capability_dashboard tool: an operator-facing projection answering, before an
agent builds a mission, which domains are actually callable through this transport versus only
declared. Documents the callable/partial/declared-only classification, the digest binding, and
what a “ready” dashboard explicitly does not grant.
AUTOPILOT.md
Autopilot: grant-authorised autonomous mission driving. bioprism autopilot
drives an instantiated workflow's mission under an explicit AutonomyGrant — the only source of
authority, with no default grant — classifying every failed step by its declared 40.36 retry
class and re-dispatching only the grant-authorised repair subset with rematerialised bindings;
terminal and cancelled steps are never re-dispatched. Success requires full step coverage, a
succeeded mission report, and by default a complete reconciliation with valid integrity; every
drive emits a digest-sealed report that autopilot verify recomputes. States its
limits up front: no scheduling or recurrence, no MCP tool exposure of the driver itself, no
cross-process resume, no wall-clock deadlines.
RESEARCH.md
The research runner: research template, research run, and
research verify execute a fixed protocol over synthetic decision
worlds — anchor the committed reference fixture against the cross-language parity
digest, then per distractor point generate, compile and certify, and run the full
equal-engineering panel, plus optional structural sweep, metamorphic mutation, and
minimization. Documents the request schema (deny_unknown_fields; the
question is recorded verbatim and never interpreted), the dossier contract —
findings derived by fixed public rules, locked to level observation, each citing
the sha256 of its supporting artifacts — and what verification does and does not prove. The
committed worked example in docs/research-example/ returns nine findings, seven
of them negative: FIBER is tied by directed-walk-full at every distractor point
and not separated in 36 of 36 sweep cells.
SECURITY_PRIVACY_AUDIT.md
The security_privacy_audit governance contract: a bounded, deterministic check
that a declared system's data assets, flows, identities, threats, reviews, and controls form a
coherent evidence chain — with the non-claims stated up front (not a scanner, not a
legal-compliance determination, not a deployment authorization).
EPISTEMIC_ADAPTIVE_ACQUISITION.md
The exact finite-horizon acquisition planner behind
epistemic_adaptive_acquisition: a policy tree (not a scalar value-of-information
summary) minimising expected terminal Bayes risk plus declared acquisition cost, with the
boundary made explicit — it plans only, and never executes an assay or establishes a causal,
biological, clinical, or predictive fact.