tcx-evidence

name	tcx-evidence
description	Collect source-backed investment evidence at the start of analyst workflows. Use for research intake, source lists, fact versus assumption separation, and missing-evidence tracking before analysis.

Collect Evidence

Build the smallest evidence set that can answer the assigned question.

Choose one artifact by default

• Create an evidence_pack under trading/research/ only when source intake has independent reuse or cross-role handoff value, or preserves a material source conflict or gap separately from a role conclusion.
• Create a role_report under trading/reports/<role>/ when the output includes that role's analysis, conclusion, or downstream handoff. When collected evidence only supports that same report, create the role report directly; an evidence pack is not a prerequisite or default duplicate.
• When both are justified, create the evidence pack first. Put its authenticated Artifact ID in the role report's input_artifact_ids, retain every exact Snapshot/Dataset ID the report uses, and rely on service receipts/hashes for lineage. Do not copy the same body into both artifacts.
• When persistence has no reuse, provenance, decision, or audit value, hand off compact IDs without creating either artifact.

1. Identify the universe and workflow type. Use only relevant, callable source classes; mark missing universe support or unavailable routes as gaps.
2. Apply tcx-source-gate before external retrieval and retain its returned source IDs and gaps. If analysis exposes a material gap, conflict, stale anchor, or identifier mismatch within your assigned question and specialty, collect the additional evidence needed to resolve it without waiting for a new field-by-field instruction. Use evidence value, not a fixed source or call count, as the stopping rule.
3. Distinguish observations, source or management claims, analysis, and assumptions in natural prose where ambiguity matters. Use opened filings, releases, and exchange/regulator records when source-of-record status matters. Otherwise accept attributable OpenBB/provider data, credible institutional data, and reputable secondary reporting for the claims they competently and currently cover. A search snippet or unsupported assumption is not evidence.
4. State the identifier, source list, source-trust notes, market context, missing evidence, freshness, decision readiness, confidence, update triggers, invalidation conditions, and contrary evidence that matter.
5. If persisting, apply $tcx-artifact, the canonical shared quality floor, and use authenticated MCP.

Carry Snapshot, Dataset, and Artifact IDs plus a compact card into calculation or handoff context. Do not summarize away used Dataset rows or repeat an unchanged source call. Do not expand into another role's complete data family or broad just-in-case collection; hand off a genuinely cross-specialty gap.

For record_source_snapshot, omit caller-owned snapshot_id, retrieved_at, and recorded_at. Supply known_at only when an exact timezone-aware knowable time is genuinely supported; never repair validation with a guessed time. When an artifact cites returned source_snapshot_ids, set knowledge_cutoff to a timezone-qualified RFC 3339 time at or after the maximum service-returned snapshot known_at, preferably that exact maximum. Never use end-of-day or another future time, and never send a date-only value. If no exact bound is available, omit the optional cutoff; if no snapshot exists, use [].

Use ready-for-portfolio-risk when every conclusion-driving claim has current, attributable, fit-for-purpose support and material conflicts and gaps are explicit; mixed source classes are allowed. Use factual-baseline for descriptive work, and screen-grade or not-decision-ready only when a material gap limits downstream use. Include high/medium/low confidence with one reason independent of readiness. Use follow_up_requests=[] when none apply; otherwise provide structured objects with trigger, suggested_role, question, reason, and materiality.