About

1.1 Candidate retrieval is recall-first: keyword matching is intentionally broad (for example AI, artificial intelligence, machine learning, LLM, GPT, GenAI, Copilot). This catches more candidates, including false positives.

1.2 Context windows are merged: overlapping hits are deduplicated into a single chunk so nearby mentions are analyzed together.

1.3 Long/noisy blocks are cleaned: very long table rows and formatting noise are reduced so classifiers see readable text.

1.4 Traceability is preserved: each chunk keeps source report identifiers, section hints, offsets, and matched keywords.

2.1 Mention type labels: adoption, risk, harm, vendor, general_ambiguous, none. These are non-mutually-exclusive except that none stands alone and means no real AI mention / false positive.

This chart is shown here because mention type is the Phase 1 gate: it determines how chunks are routed to downstream classifiers, so changes here flow through the rest of the pipeline.

2.2 Routing logic: the risk classifier only runs on chunks tagged risk, adoption classifier only on adoption chunks, and vendor classifier only on vendor chunks. If Phase 1 misses a tag, Phase 2 for that branch will not run.

2.3 Taxonomies used: adoption type = non_llm, llm, agentic. Risk taxonomy = strategic_competitive, operational_technical, cybersecurity, workforce_impacts, regulatory_compliance, information_integrity, reputational_ethical, third_party_supply_chain, environmental_impact, national_security, none. Vendor tags include amazon, google, microsoft, openai, anthropic, meta, internal, undisclosed, other.

2.4 Signal and substantiveness: adoption uses 0-3 signals, while risk and vendor use 1-3 signals (weak implicit to explicit). Risk also gets a substantiveness label (boilerplate/moderate/substantive).

Exact Taxonomy Reference (Canonical Labels)

Labels are shown exactly as stored in classifier outputs and dataset fields for transparency.

3.1 Confidence handling: report-level adoption/risk trend counts use confidence thresholds (default 0.2) where confidence maps exist; explicit risk signal entries are retained. Signal heatmaps bin values into weak/strong/explicit.

3.2 Legacy compatibility: older risk labels (for example regulatory, workforce) are mapped to current canonical labels so longitudinal charts stay comparable.

3.3 Report denominator is explicit: we keep no-signal reports in the report-level dataset to show blind spots, not just positive cases.

Structured outputs: classifiers write to strict response schemas (Pydantic + JSON schema), reducing malformed labels.

Conservative prompting: prompts require explicit AI attribution and discourage category over-assignment.

Testing and reconciliation: repo scripts support QA checks, human-vs-LLM disagreement review, and merge-back of reconciled labels.

Known limitations: this release is still primarily LLM-labeled and keyword-seeded, so it can miss subtle non-keyword AI references and can still include some ambiguous cases.