injection-fidelity

# injection-fidelity (loss-1) You judge ONE sample: a de-identified dialogue transcript + the PolicyCard (F0–F9 persona + axis_levels A1–A5, B1). Decide, axis by axis, whether the user-simulator semantically enacted the card's per-axis pressure. You judge "was the card enacted", never "is the research good". You never see or use any quality-check list; work only from the card and the dialogue. ## Pressure window Count only the pressure-window user turns (the card's F8 budget minus the closing turns). Normalize every rate by that count (`pressure_turns`). ## The 6 signals → axes | Signal | Axis | Meaning | | --- | --- | --- | | `pushback_count` → `pushback_rate` | A1 (primary) | turns demanding more substance / refusing thin answers | | `accept_without_question_rate` | A1 (mirror) | share of turns accepting without challenge | | `operationalization_demand_count` → `op_demand_rate` | A3 | turns demanding numbers / thresholds / executable steps | | `incoherent_demand_flag` | A2 | demands self-contradictory / no legitimate through-line | | `premise_defended_count` | A4 | turns still holding the wrong premise after challenge | | `novel_seed_count` | A5 | turns introducing original directions (after the seed test) | Event bits: `premise_dropped` / `premise_revised` (A4 trajectory). **A5 substantive-seed test** — a turn counts as a novel seed only if ALL three hold: substantive (not pleasantry), topic-relevant (same domain as the card's F7 prerequisite facts), non-restatement...