Expand description
§hyprstream-decision-stub — System One P0.7 stub decision service + wire smoke test
A deterministic mock decision service returning jev-1-shaped, schema-valid answers behind the two contracts the real serving paths will implement:
- The jev-1 JSON/HTTPS facade (
facade):POST /v1/systemoneandGET /v1/modelswith the envelope, answer derivations, and error shapes of the jev-1 compatibility profile pinned byhyprstream-decision(P0.1a). This is the portability-prove surface: the stock TypeSafe SDKs (typesafe-sdk-python,@typesafe-ai/sdk, both MIT) run against it unmodified viaTYPESAFE_BASE_URL(S6a §4 — the adapter never speaks the/v1/systemonewire, so stock-SDK repointing is the proof). - The P3.5 Flight SQL/ADBC
decidecontract (flight): a prepared statement pins a question set plus the (schema, model, calib) version triple; bound parameter batches carry onestateper row; the result stream is ahyprstream-decisionarrowDecisionSchemabatch. Handles are content-derived and unknown handles fail loudly (not_found) — the drift semantics P3.5 requires of prepared statements.
The mock (mock) is deterministic: distributions are a hash of
(model id, canonical question serialization, state, row index), so every byte on the
wire is reproducible and the wire-overhead floor (tests/wire_floor.rs) can pin exact
sizes. It is not a model — it exists so contract conformance and overhead are
measurable before P1.x lands.
§Facade decision — bare bool/number entries (deferred from P0.1a, owned by P0.7)
The upstream jev-1 EntryType is string | object | array | null; bare booleans and
numbers in entry positions are outside it. hyprstream-decision’s IR admits
Entry::Bool / Entry::Number as a documented
superset-permissive extension (structured criteria routinely carry numeric
thresholds). Decision: the facade ACCEPTS bare booleans and finite numbers in every
entry position (state, instructions, criteria values) and treats them as the
IR’s native scalars — canonical serialization renders them as their JSON literals
(true, 42.5). It does not stringify them (that would change rubric text the
encoder trains on) and does not reject them (the IR, the Arrow contract, and the
scorer all handle them). Consequence, documented for P3.6 migration recipes: a
request using bare scalars is valid against this superset profile but is not
portable to the upstream service, whose schema admits only string/object/array/null.
§What the stub deliberately does not do
No auth (any Bearer token is accepted; a missing one is a 401), no rate limiting, no
persistence, no calibration. usage.input_tokens is a documented stub approximation
(ceil(serialized_bytes / 4) over the canonical serialization — S6a A11 leaves the
tokenizer unspecified; consumers must not compare counts cross-service), and
output_tokens is the sum of answered cardinalities.
Modules§
- facade
- The jev-1 JSON/HTTPS facade:
POST /v1/systemone+GET /v1/models. - flight
- The P3.5
decidecontract, stubbed: Flight SQL prepared statements that pin a question set plus the (schema, model, calib) version triple, bind batches of states, and returnDecisionSchema-conformant Arrow batches. - mock
- The deterministic mock decision model.
- render
- Rendering: IR answers → the jev-1 wire answer shapes.
- wire
- The jev-1 wire envelope:
POST /v1/systemonerequest → validated IR.