Capability Map¶
bijux-canon-index makes vector execution explicit from request through
replay. Capabilities are admitted by declared contracts and backend evidence;
module presence alone does not make an adapter part of the supported v1
surface.
flowchart LR
request["intent + mode + contract + budget"]
materialize["immutable execution artifact"]
resolve["capability resolution"]
execute["exact or bounded ANN"]
record["results + cost + provenance"]
review["explain + replay + compare"]
request --> materialize --> resolve --> execute --> record --> review
resolve -. refusal .-> record
Governed execution capabilities¶
| Capability | Owning area | Reviewable output |
|---|---|---|
| Request normalization | domain/requests/ and interface schemas |
validated intent, mode, contract, budget, metric, and randomness policy |
| Artifact materialization | domain/artifact/ and application orchestration |
immutable corpus/vector/configuration identity |
| Capability discovery | registries and capabilities report | eligible contracts, metrics, dimensions, ANN support, consistency, and exclusions |
| Exact scoring | algorithms and exact-capable adapters | stable tie order, ordered result IDs, scores, and execution cost |
| ANN candidate execution | domain/non_determinism/ and ANN adapters |
parameters, seed/sources, candidate policy, decision trace |
| Optional exact rescore | non-deterministic plan actions | explicit candidate-to-final ranking path |
| Budget enforcement | request budgets and performance contracts | refusal or partial result with the breached dimension |
| Transaction and authorization guards | contracts/ |
explicit misuse or denial before mutation |
Evidence and review capabilities¶
| Capability | Owning area | Reviewable output |
|---|---|---|
| Fingerprinting | core identity and provenance modules | vector, configuration, backend, determinism, plan, and result identities |
| Run lifecycle | infra/run_store.py |
incomplete, failed, or complete three-file run directory |
| Explanation | query introspection and result records | document, chunk, vector, metric, score, rank, artifact, and execution join |
| Replay | domain/provenance/replay.py and application support |
match decision, fingerprints, mismatch details, randomness sources |
| Drift analysis | domain/drift/ and comparison requests |
backend, artifact, parameter, and execution differences |
| Plugin discovery | plugin entry points and registry contracts | provider/backend identity and capability declaration |
| Public execution | CLI module and HTTP v1 | structured results or typed governed refusals |
Backend boundary¶
Memory and SQLite provide the local supported persistence baseline. ANN and optional vector-store adapters add capabilities only when installed, registered, reachable, and conformant. Remote backends, asynchronous services, and streaming search are excluded from the v1 contract; pgvector remains experimental and outside the v1 freeze.
Capability status is evidence-bound¶
| Status | Meaning | Reader-visible evidence |
|---|---|---|
| supported baseline | package-owned contract with exercised local implementation | request/artifact/result fixtures, exact behavior, persistence and replay cases |
| optional and conformant | installed adapter passed the declared capability and backend suites in the recorded environment | adapter/backend identity, capability report, conformance results and failure cases |
| available but unadmitted | code or plugin can be imported but lacks current evidence for the requested capability | explicit exclusion/refusal; no supported-result claim |
| experimental | behavior may be inspected without entering the frozen v1 contract | experimental identity and limits kept separate from v1 evidence |
| excluded | asynchronous/streaming/remote behavior not promised by v1 | documentation and interface refuse to imply availability |
Capability status applies per operation. An adapter that has query evidence is not automatically admitted for ingest, transactions, persistence, replay, tenant isolation, or bounded ANN. Inspect the capability report before materializing an artifact or selecting an execution plan.
Interpret execution outcomes¶
| Outcome | Required record | Claim supported |
|---|---|---|
| admission refusal | request identity and violated contract/capability/budget | no execution was authorized |
| materialization failure | request plus partial artifact state and typed failure | no immutable execution artifact is available |
| backend failure | artifact, plan, adapter/backend identity, attempts and failure provenance | execution was attempted but produced no admissible complete result |
| partial result | completed work, missing work, breached budget and explicit partial status | only named candidates/work are usable under the stated limitation |
| complete exact result | artifact, exact plan, stable order/scores, cost and provenance | ranking is exact under the declared metric and artifact |
| complete bounded ANN result | approximation/randomness policy, candidates, quality/budget evidence and final ranking | result satisfies the predeclared ANN envelope, not exact equality |
| replay match or acceptable diff | original/observed identities, semantic diff, envelope, verdict and reason | comparison met the original replay policy |
| replay mismatch | blocking differences and producer identity | prior execution is not reproduced under the requested rule |
Completion, exactness, determinism and replay acceptability are separate properties. A caller must not derive one from another.
Choosing an execution path¶
- Use deterministic strict execution when exact equality and replay are required.
- Use bounded ANN only when quality, resource, randomness, and divergence policy are declared before execution.
- Inspect capabilities before materialization or execution; do not discover incompatibility from a failed production query.
- Retain the execution artifact and complete run directory together when a decision must be audited.
- Treat explanation and replay as provenance evidence, not proof of corpus quality, semantic relevance, or factual truth.
The execution invariants define refusal boundaries, and known limitations distinguish governed execution from stronger claims the package cannot make.