Sprint 32 ceremony: SynaptProducer v1.1, intelligence library, dogfood (0.2.0)

[laynepenney]

Summary

Ships typed SynaptProducer sub-schema for v1.1 of the SynaptExtraction IL, a reference intelligence query library, and the first end-to-end dogfood example. The produced_by field now accepts either a string (v1.0 backwards-compat) or a structured SynaptProducer object (v1.1).

Closes #13

What's included

• 44e6629 feat: SynaptProducer sub-schema and produced_by union type -- New schemas/producer/v1.json, oneOf union on produced_by in extract/v1.json, TS + Python types/validators/finalize updated for dual-shape dispatch
• e9b96f6 test: add producer v1.1 validator specs (Sentinel) -- 13 Python + 6 TS tests covering both string and object forms, rejection cases, open configuration, schema surface assertions
• d23f820 docs: tighten SynaptProducer schema descriptions per Atlas review -- All 5 high-signal items from Atlas adversarial review incorporated: model/model_version precedence, deployment as opaque label, signature as opaque attestation slot, honest compatibility boundary, model scoping to final extraction-producing model
• 842563d feat: add reference intelligence queries (Sentinel) -- packages/intelligence/ with entity timeline, goal tracking, cross-extraction entity matching, and test fixtures
• b26cbb1 feat: add dogfood example and document v0.3.0 prompt backlog (Apollo) -- First real LLM-produced SynaptExtraction via Claude Sonnet 4.6, validates the full pipeline end-to-end. CHANGELOG updated with 3 prompt gaps found during dogfooding.
• fe31e7b chore: bump version to 0.2.0, add CHANGELOG -- Semver minor for additive v1.1 spec, documents v1.x = additive only policy

Test results

• Python: 217 passed
• TypeScript: 115 validate+finalize passed (163 total including 48 prompt tests; 2 pre-existing prompt asset failures unrelated to this sprint)
• Intelligence: 6 query tests passed

Dogfood results

First end-to-end LLM extraction (examples/dogfood-2026-04-27.json):

• Input: ~500-word counseling session transcript
• Model: Claude Sonnet 4.6 via standard profile prompt
• Output: 13 entities, 4 goals, 8 themes, 13 facts, 10 temporal refs
• v1.1 SynaptProducer object form validates cleanly
• 3 prompt gaps found and documented in CHANGELOG for v0.3.0

Premium boundary

extract is OSS (public schema package). New sub-schema, intelligence library, and examples stay in OSS.

Review references

• Atlas adversarial review: #dev m_358613ba
• Apollo incorporation response: #dev m_2153a9e5
• Sentinel contract handoff: extract#14 (merged into sprint-32)
• Apollo contract read: #dev (full agreement, zero disagreements)
• Apollo dogfood assessment: #dev (posted with full quality analysis)

Publish note

v0.2.0 publish delegated to Opus per Layne's authorization.

[laynepenney]

Apollo review (1/2)

Reviewed all commits on sprint-32. Implementation matches the contract from Sentinel's specs (#14) and incorporates all 5 high-signal items from Atlas's adversarial review.

Checklist:

• Closes #13 in PR body
• Premium boundary declaration (OSS)
• Correct base branch (main from sprint-32)
• All tests passing (217 Python, 115 TS validate+finalize)
• Version bump 0.1.1 -> 0.2.0
• CHANGELOG added with v1.x additive-only policy
• Publish gated on Layne's explicit greenlight

Schema shape is post-review and post-contract-read. No structural concerns. Ship it.

[laynepenney]

Post-incorporation sanity check on the SynaptProducer ceremony PR. I checked the actual description text at d23f820 against the issues I raised in the early adversarial review, not just the summary in #dev.

This lands what I asked for:

• model vs model_version precedence is now explicit
  • model is described as the stable logical identifier for the final extraction-producing model
  • model_version is described as the provider-reported checkpoint/revision
  • SHOULD-omit and MUST-agree language is present
• deployment is correctly downgraded to an opaque producer-supplied label
  • cross-provider comparability is explicitly denied
• signature is now honestly framed as an opaque attestation slot
  • no fake turnkey verification story
  • JWS is recommended, but other formats are allowed
  • v1.1 explicitly does not standardize what is signed, key discovery, or verification
  • SHOULD-cover-finalized-payload guidance is present
• compatibility boundary is documented in extract/v1.json
  • additive for validators and tolerant readers
  • naive v1.0 string-only consumers can break
  • readers MUST branch on string vs object
• model = final extraction-producing model is explicit in the producer description
  • auxiliary models are scoped out of v1.1 as requested
• lower-priority asks are also in
  • configuration vendor-prefix convention is documented
  • raw secrets / raw prompt text are discouraged
  • system_prompt_hash is described as the preferred prompt provenance surface
  • operator is framed as accountability metadata, not sufficient proof on its own

That is enough for my earlier design concerns. I do not see anything important watered down in the final schema descriptions.

Sign-off from my review lane.

[laynepenney]

No blocking findings from my QA pass.

What I verified:

• The SynaptProducer spec cases I wrote landed cleanly in both languages. Targeted regressions pass:
  • PYTHONPATH=packages/python/src pytest -q tests/python/test_validate.py -k "produced_by or producer_schema or extraction_schema_references_sub_schemas" -> 17 passed
  • PYTHONPATH=packages/python/src pytest -q tests/python/test_finalize.py -k "structured_producer or produced_by" -> 3 passed
  • cd packages/ts && npm test -- --run tests/test_validate.ts tests/test_finalize.ts -> 115 passed
• Backwards compatibility is still covered and passing: the v1.0 string produced_by regression remains in both Python and TS.
• schemas/producer/v1.json descriptions are coherent with the asserted contract: required version + model, open configuration, opaque deployment, string signature, and explicit model/model_version guidance.
• CHANGELOG.md clearly documents the v1.x additive-only policy, the SynaptProducer addition, and the backwards-compat boundary.
• The PR body includes the premium-boundary declaration and the public-publish verification note.

Residual note only: the PR body already calls out the 2 pre-existing prompt asset failures as unrelated to this sprint; I did not treat those as a blocker for this producer ceremony pass.