feat: compose evidence-bound Event Intelligence (v2.18.3)

CHANGELOG.d/2.18.3-event-intelligence-dossier.md

@@ -0,0 +1,22 @@
+# 2.18.3 Event Intelligence Dossier
+
+LineageWeave can now compose a strict, digest-bound Event Intelligence Dossier
+from source evidence, multi-clock temporal context, its knowledge graph and
+ontology, TEPP temporal/topic artifacts, fast-mlsirm psychometric artifacts,
+and a structured contextual-orchestrator verdict. The channels keep separate
+methods, versions, uncertainty, and authority; unavailable channels are
+explicit and no blended event score is invented. A JSON Schema, versioned
+OWL-Time/PROV-O profile, SHACL shapes, canonical example, and validator CLI are
+included.
+
+The semantic profile separates the generated dossier entity from the
+`prov:Activity` that used evidence and generated it. Source posts now support a
+first-class EventAssertion instead of being classified as influences on the
+real-world event, while the compact Buyer `evidencesEvent` edge remains a
+non-causal read-model projection. Event intervals and assertion, document,
+availability, and cutoff instants are explicitly grounded in OWL-Time.
+
+The legacy contextual-orchestrator lineage adjudication path now requests one
+strict JSON verdict, treats record labels as untrusted JSON evidence, requests
+the orchestration trace, and fails closed on malformed, duplicated, non-finite,
+or free-form output instead of regex-extracting an arbitrary number.

docs/adr/0015-adaptive-contextual-orchestrator-default.md

@@ -13,7 +13,10 @@ Ordinary LineageWeave LLM consumers request `mode="auto"`.
 
 The orchestration plane owns provider/model selection, test-time compute, workflow depth, verification, fallback, and known-price optimization. Quality sufficiency is the first constraint; cost is minimized among execution paths that satisfy it. Unpriced models are not treated as free.
 
-Explicit `verify` remains for the citation-bearing post-chat and lineage adjudication paths because those are deliberate checked-judgment contracts, not product defaults. Explicit route or conduct modes may be used only for documented ablation, incident response, or a bounded domain requirement.
+Checked post-chat and lineage-adjudication paths retain strict evidence and
+output contracts while using `mode="auto"` under ADR 0079 and their
+consumer-specific ADRs. Explicit route or conduct modes may be used only for
+documented ablation, incident response, or a bounded domain requirement.
 
 LineageWeave continues to own strict output parsing, evidence identifiers, IRT projection, and fail-closed domain validation.
 

docs/adr/0064-lineage-evidence-and-tree-assembly.md

@@ -18,7 +18,9 @@ or promoting an inferred relation to fact.
 - Fuse independent temporal, secondary-key, text/embedding, and optional LLM
   channels through the RankWeave weighted convex fusion contract. A missing
   channel is dropped and weights are renormalized; it is never replaced with a
-  fabricated negative or score.
+  fabricated negative or score. A structured `refuted` verdict remains a real
+  negative score; `insufficient_evidence` drops the LLM channel for that whole
+  candidate comparison so every candidate is ranked with the same weights.
 - Keep the channel-score breakdown and provenance on every candidate decision.
   Candidates below the minimum fused-score floor remain roots rather than being
   force-attached.

docs/adr/0120-event-intelligence-dossier.md

@@ -0,0 +1,263 @@
+# ADR 0120: Compose Event Intelligence without collapsing scientific or provenance authorities
+
+- **Status:** Accepted; amended 2026-08-21
+- **Date:** 2026-08-20
+- **Decision owners:** LineageWeave product and scientific integration maintainers
+- **Related:** ADR 0003, ADR 0004, ADR 0016, ADR 0034, ADR 0065, ADR 0074, ADR 0079, TEPP ADR 0011
+
+## Context
+
+The Buyer-surface stack ending at PR #264 makes LineageWeave evidence easier to
+reach, but the product still exposes its event-intelligence inputs as separate
+features:
+
+- the LineageWeave knowledge graph computes an evidence-backed neighborhood;
+- the LineageWeave ontology provides semantic identifiers and labels;
+- TEPP owns temporal-event and topic-model scientific artifacts;
+- fast-mlsirm owns calibrated psychometric estimates and their uncertainty;
+- contextual-orchestrator supplies bounded model routing and LLM judgment; and
+- source posts and model artifacts carry independent provenance.
+
+No versioned object required all of these channels to share the same immutable
+snapshot, knowledge cutoff, evidence identities, ontology references, method
+versions, uncertainty, and content digests. A UI or downstream integrator could
+therefore join unrelated clocks, show an LLM verdict as though it were a
+psychometric score, or omit an unavailable scientific channel without saying
+that it was unavailable.
+
+The first ontology profile correctly treated the dossier as a `prov:Entity`,
+but also made `usesEvidenceBundle` a subproperty of `prov:used` with the dossier
+as its domain. Because PROV-O defines `prov:used` for an activity using an
+entity, ordinary RDFS reasoning would infer that the dossier entity was also a
+`prov:Activity`. The same profile made the direct Post-to-EventEpisode
+`evidencesEvent` edge a subproperty of `prov:influenced`, which could be read as
+the later source document influencing the real-world event rather than
+supporting an assertion about it.
+
+Those are semantic-model defects, not cosmetic vocabulary choices. They would
+make standards-aware consumers infer roles the product does not intend.
+
+## Decision
+
+LineageWeave publishes **Event Intelligence Dossier v1** as an evidence-bound,
+deterministic composition contract. The wire contract remains version 1 because
+this correction occurs before the profile is released on protected `main`.
+
+### Dossier entity and generation activity
+
+The dossier is a buyer-facing read artifact and a `prov:Entity`; it is not an
+estimator or an activity. A separate `DossierGenerationActivity`, subclassed
+from `prov:Activity`, represents deterministic composition.
+
+```text
+DossierGenerationActivity
+  -- usesEvidenceBundle / prov:used --> EvidenceBundle
+  -- usesEventAssertion / prov:used --> EventAssertion
+  -- generatesDossier / prov:generated --> EventIntelligenceDossier
+```
+
+This preserves the PROV-O domain and range contract instead of relying on a
+single resource to be both the process and its output.
+
+### Event, assertion, and source separation
+
+A source post is not the event and is not asserted to have caused the event.
+The profile therefore introduces `EventAssertion` as a first-class
+`prov:Entity`:
+
+```text
+EventAssertion
+  -- supportedBySource / prov:wasDerivedFrom --> LineageWeave Post
+  -- assertsEvent --> EventEpisode
+```
+
+The existing `evidencesEvent` Post-to-EventEpisode relation remains available
+as a bounded Buyer read-model convenience edge. It is deliberately **not** a
+subproperty of `prov:influenced`, does not transfer source authority to an
+inference, and has no causal meaning. Producers that publish full RDF should
+retain the mediating assertion; bounded Buyer graph projections may publish the
+convenience edge together with the same evidence identifiers.
+
+### Multi-clock temporal semantics
+
+The dossier keeps six distinct clocks: event start/end, assertion, document,
+availability, and knowledge cutoff. Exact RFC 3339 values remain in the JSON
+contract and corresponding RDF datatype properties. The profile additionally
+uses OWL-Time resources:
+
+- an `EventEpisode` has exactly one temporal extent represented as a
+  `time:Interval` in the published SHACL profile;
+- assertion, document, availability, and cutoff clocks may be represented as
+  `time:Instant` resources; and
+- forward transitions remain distinct from retrospective reporting.
+
+A later document may describe an earlier event, but that reporting relation
+must never become a reverse state transition.
+
+### Versioning and imports
+
+`docs/ontology/event-intelligence-profile.ttl` declares a stable ontology IRI,
+`owl:versionIRI` for profile 1.0.0, and metadata imports for the LineageWeave
+base ontology, PROV-O, and OWL-Time. Runtime code and tests parse committed
+artifacts only and do not dereference imports over the network.
+
+### SHACL interchange constraints
+
+`docs/ontology/event-intelligence-profile.shacl.ttl` publishes closed-world
+constraints for the semantic boundaries that OWL/RDFS alone should not be
+expected to reject:
+
+- one evidence bundle and one generated dossier per generation activity;
+- one asserted event and at least one source per event assertion;
+- one OWL-Time interval per event episode;
+- required assertion, document, event-start, and knowledge-cutoff values; and
+- class constraints for PROV entities/activities and OWL-Time instants/intervals.
+
+The repository tests parse and inspect both the ontology and the SHACL graph.
+The production JSON validator remains authoritative for the current JSON wire
+artifact; the SHACL document is the standards-based RDF validation contract for
+external graph consumers.
+
+## Dossier contents
+
+The dossier contains:
+
+1. a source snapshot identity and the six distinct clocks;
+2. versioned ontology references for event and graph assertions;
+3. immutable evidence references with source authority, URI, digest,
+   availability time, and recorded time;
+4. an evidence-backed LineageWeave graph neighborhood and method-labelled
+   relevance with uncertainty;
+5. an optional TEPP artifact that uses the same snapshot and cutoff and retains
+   TEPP model, engine, and digest identity;
+6. an optional fast-mlsirm artifact that retains its construct scale, estimate,
+   standard error, model, engine, and digest identity;
+7. an optional contextual-orchestrator verdict that cites evidence and records
+   trace, operation, policy, prompt digest, verdict, confidence, and rationale;
+8. buyer-facing claims whose complete supporting evidence IDs are explicit;
+   and
+9. a SHA-256 over the RFC 8785 JCS canonical dossier payload after removing
+   `dossier_sha256`.
+
+The JSON Schema is `schemas/event_intelligence_dossier_v1.schema.json`. The
+runtime implementation is `lineageweave.event_intelligence`; the validator CLI
+is `lineageweave-validate-event-intelligence`.
+
+## Authority rules
+
+| Channel | What it may assert | What it may not replace |
+|---|---|---|
+| LineageWeave knowledge graph | Graph neighborhood and graph relevance | TEPP topic inference or psychometric calibration |
+| LineageWeave ontology | Semantic identifiers and relation meaning | Observed source evidence |
+| TEPP | Temporal/topic artifact under its own model contract | LineageWeave authorization or source-of-record data |
+| fast-mlsirm | Calibrated estimate and uncertainty on a named scale | TEPP temporal/topic truth |
+| contextual-orchestrator | Evidence-bounded supported/refuted/insufficient verdict | Numerical relevance or psychometric measurement |
+| source evidence | What was available and recorded | Model-derived inference or real-world causation |
+
+The composer never averages these outputs into one number. A missing TEPP,
+fast-mlsirm, or orchestrator channel is serialized as exactly
+`{"status_code":"unavailable"}` rather than a zero, null score, or fabricated
+fallback.
+
+## Temporal and validation rules
+
+Every evidence item must satisfy:
+
+```text
+available_time <= knowledge_cutoff
+```
+
+The TEPP artifact must match both `source_snapshot_id` and
+`knowledge_cutoff`. Event time may precede assertion, document, or availability
+time; those clocks remain separate so retrospective reports do not leak into a
+historical model.
+
+Runtime reconstruction rejects:
+
+- unknown or missing fields;
+- non-UTF-8 or invalid JSON in the CLI;
+- future evidence relative to the cutoff;
+- unknown evidence IDs;
+- graph edges whose endpoints are absent;
+- mismatched TEPP snapshot or cutoff;
+- orchestrator attempts to add a psychometric score;
+- free-form, duplicated-field, non-finite, out-of-range, or malformed lineage
+  adjudication output;
+- unsupported channel states; and
+- altered payloads whose dossier digest no longer matches.
+
+The ontology tests additionally reject a return to a dossier-domain
+`prov:used`, a Post-to-Event PROV influence, a generic temporal-entity range,
+or an unversioned/unimported profile.
+
+## Consequences
+
+### Positive
+
+- Standards-aware consumers no longer infer that a dossier entity is also its
+  generation activity.
+- A source document supports an assertion about an event instead of being
+  represented as an influence on that real-world event.
+- Buyers receive one inspectable event-intelligence artifact rather than a set
+  of unrelated widgets.
+- Disagreement between graph, topic, psychometric, and judge channels remains
+  visible and auditable.
+- TEPP and fast-mlsirm can evolve behind their own versioned contracts without
+  LineageWeave reimplementing their mathematics.
+- JSON consumers retain the strict existing contract while RDF consumers gain
+  versioned imports and SHACL constraints.
+
+### Costs and limitations
+
+- Full RDF exchange contains an additional assertion node and generation
+  activity that compact Buyer graph projections may omit.
+- Publishing SHACL shapes does not turn the current JSON runtime into a generic
+  RDF store or SPARQL service.
+- This ADR defines composition and validation, not a live TEPP HTTP service or
+  a new fast-mlsirm estimator.
+- A backend projection and Buyer UI still need to select authorized artifacts
+  and render the dossier.
+- Causal claims remain out of scope unless a separately validated model and
+  claim type support them.
+- An LLM verdict remains a fallible judgment channel and must be calibrated and
+  compared with human or statistical evidence for high-stakes use.
+
+## Rejected alternatives
+
+### Treat the dossier as both Entity and Activity
+
+Rejected because `prov:used` and `prov:generated` describe activity behavior.
+Using them directly from the dossier makes reasoners infer an unintended
+activity type and conflates a process with its output.
+
+### Make the source post influence the event episode
+
+Rejected because a report can be created after the event and may only support
+an assertion about it. A generic PROV influence edge is too broad and invites a
+causal reading the product cannot justify.
+
+### One blended event-relevance score
+
+Rejected because the component scales and authorities are not interchangeable,
+and a weighted average would obscure uncertainty and disagreement.
+
+### Copy TEPP topic or fast-mlsirm estimation into LineageWeave
+
+Rejected because production numerical authority belongs in those repositories,
+and duplicated implementations would drift.
+
+### Let contextual-orchestrator synthesize the complete artifact without a
+strict schema
+
+Rejected because source text is untrusted, free-form output is not a durable
+contract, and an LLM must not decide the scientific acceptance boundary.
+
+### Store only prose with citations
+
+Rejected because buyers and downstream systems need machine-checkable clocks,
+method versions, uncertainty, ontology identifiers, and content digests.
+
+## References
+
+See `docs/doctoring/EVENT_INTELLIGENCE_REFERENCES.md` for APA 7th references
+and requirement traceability, including PROV-O, OWL-Time, and SHACL.