Calibration runs exercise each verifier in the Source Network against a seed corpus of claims whose ground truth has been established independently. Precision is the share of positive verdicts that were correct; recall is the share of real positives that were caught. F1 combines the two. A verifier with a missing API key returns unverifiable on every claim and therefore scores n/a in the table below; the row is committed to make the key-missing code path legible, not because it represents the verifier's calibrated reliability. See the calibration hub for the run index and the methodology protocol.

Detector empirics: per-FVS firing rates

Generated: 2026-05-01

Corpus size: 13 documents

Source: `data/adversarial_fixtures//document.md` + `data/worked_examples/.md` (top-level only)

What this measures

Firing rate per FVS detector and per coverage perspective. Genre + voice classification distribution. Absence-cluster dimension incidence. Computed by running `frame_check` over stdio MCP against each document and aggregating the structural fields. Recall and precision are not reported; those require gold-standard labels per document, which is operator-authoring work outside this harness.

Per-FVS firing rate (presence detection)

Read this table as: 'FVS-NNN fires as PRESENT in K of N corpus documents.' The complement (`N - K`) is the absence count for the same detector. The methodology paper's claim 'FVS-NNN absent in M of N documents' is `(N - K)` for this row, MINUS any rows where the same FVS-NNN fires as ABSENCE-pattern (next table) since those are not silently-absent but actively-detected-as-absent.

Per-FVS absence-pattern firing rate

V1-detector absence-pattern fires (`pattern_kind == "absence_detected"` in the structured emission). Today only FVS-007 Failure Framing has an absence-pattern detector in `frame_library.py:362-370` (fires when risks AND uncertainty are both missing AND unhedged-claim density exceeds 60%). A document where FVS-007 fires absence-pattern is actively-detected-as-absent on the failure-framing dimension, distinct from documents where FVS-007 is silently-absent (no fire, no evidence of absence-detection structural conditions).

Coverage perspective addressed rate

Genre classification distribution

`(abstain)` = classifier returned `null` (no feature-marker regex matched). Per the construct-honesty discipline, abstention is preferred over mislabeling. A high abstention rate on a corpus is not a defect; it is a measurement of how often the regex-based feature surface fires.

Voice classification distribution

Frame deepening per-detector firing rate

Three regex/feature detectors at `frame_deepening.py` (`detect_temporal_scope`, `detect_stakeholder_map`, `detect_falsification_conditions`). Each returns `Optional[dict]`: `None` when no signal is found, a dict of structural evidence otherwise. The MCP wire shape at `analysis.frame_deepening.` preserves this distinction (`null` vs object). 'Fires' here means the detector returned a non-`None` dict, regardless of how rich the dict's content is. Per `NEXT_STEPS.md` "Substrate-side composition: web exposure", web exposure of `frame_deepening` is gated on per-detector expert-grading; this aggregate is the empirical baseline the grading pass calibrates against. A 100% firing rate across the corpus is itself a finding: it tells the operator the corpus has CEILING saturation on this detector and grading needs documents specifically lacking the signal to measure the false-positive boundary.

Absence-cluster dimension incidence

Per-document detail

Reproducibility

Re-run via `python3 scripts/detector_empirics.py`. Determinism: the substrate is regex-only (no LLM); identical input produces identical output. Drift between this report and a future run either means (a) the corpus changed (added/removed/edited fixtures) or (b) a detector regex changed. The aggregate.json and per_document.json files in this directory are the machine-readable artifacts; this report is the human-readable summary derived from them.