[Event Intelligence Gap] Complete calibrated TDT and CHRONOS event detection, schema, and temporal reasoning

[seonghobae]

Buyer problem

TEPP has numerous bounded event-intelligence gates for mention confidence, first-story decisions, link/track/segment metrics, schema slots, forecasts, and refusal of predictions as facts. Protected main still does not provide one calibrated event-intelligence workflow that converts documentary evidence into reviewed event mentions/instances, TDT tracking, CHRONOS-style schema proposals, temporally consistent hypotheses, and measurable downstream artifacts.

Product outcome

Deliver one coherent event_intelligence boundary that coordinates:

1. story/measurement-unit segmentation;
2. mention extraction with calibrated confidence;
3. pair link detection;
4. candidate event/topic detection;
5. first-story/onset detection;
6. longitudinal tracking with identity-switch evidence;
7. event-schema/argument proposal and instance admission;
8. next-event occurrence hypotheses;
9. interval-relation reasoning and contradiction detection;
10. explicit promotion from proposed/inferred evidence to reviewed event instance or transition.

Existing related work

Open PRs #50, #64–#66, #68, #70, #72, #74, #76, #85, #118, and related event/relation gates cover isolated metrics or refusal rules. They must be reviewed for consolidation, supersession, and composition into a product boundary. A crate per metric is not the final architecture.

Acceptance criteria

• Event mentions, event instances, schema hypotheses, forecasts, observed facts, and promoted transitions remain distinct typed objects with provenance.
• Every mention and argument cites exact source span(s), document identity, six-clock evidence, confidence/calibration, extractor/model version, and review status.
• TDT segmentation, link detection, topic/event detection, first-story detection, and tracking run as one versioned workflow with reproducible configuration.
• Metrics include WindowDiff/Pk, boundary precision/recall, link/cluster pair precision/recall, first-story false-alarm/miss, tracking switch rate, and calibrated probability scores as applicable.
• CHRONOS-style schema/argument proposals and occurrence forecasts are hypothetical until deterministic validation and authorized review; no LLM/reasoner output becomes observed evidence by identity.
• Allen/OWL-Time-style interval reasoning derives bounded implications and rejects contradictions without claiming unrestricted global satisfiability.
• Input→process→outcome and state-transition edges require strict forward event-time order. Citation, revision, translation, summary, support, contradiction, outcome_of, and retrospective reporting remain provenance.
• Unobserved pairs remain missing evidence, not negative relations. Association and precedence cannot be described as causal without an identified design.
• Known-truth simulations cover noisy mentions, duplicate stories, delayed reporting, revised/translated documents, overlapping events, partial schemas, track drift, and forecast calibration.
• Realistic anonymized documentary fixtures demonstrate exact-span extraction, temporal cutoff safety, event tracking, and schema provenance end to end.
• Results are persisted and exported as versioned JSON/JSON-LD/GraphML artifacts consumable by the end-to-end run and LineageWeave contracts.
• Production statement/branch coverage and public docs remain 100%; fuzz/property tests cover malformed, adversarial, oversized, duplicated, inconsistent, and prompt-injected evidence.

Scientific claim boundary

This feature measures detection, linking, tracking, schema, and forecast quality. It does not establish real-world truth, causal effect, or psychometric construct validity merely because a model/reasoner agrees.

Authority

Use ADR 0016, PRD §11, the temporal/event architecture, and the APA 7 research register.