…ation proxy, transcript formats

challenge-protocol.md:
- Graduated activation (Level 3 = Assumption Audit only, Level 4 = + Blind
  Spot, Level 5 = all 5 lenses) instead of binary Level 5 only
- Challenge Debt: unresolved challenges tracked across sessions and surfaced
  at the start of the next relevant session instead of silently dropped
- Avoidance Confrontation now has distinct Level 3/4 behaviors, not just
  a single Level 5 mode

story-mapping-engine.md:
- Framing Variety Constraint replaces "no story twice" rule: a story may
  appear twice in a mapping if the angles are substantively different,
  tested via a 3-criteria Substantiation Test (different competency,
  different STAR section, different interviewer takeaway)
- Story Freshness Score: 3-factor decay function (use frequency risk,
  experience age risk, Proven Performer offset) replacing raw use-count
  overuse check. Verdict: Fresh / Fresh (Proven) / Moderate / Stale

calibration-engine.md:
- Practice Calibration Proxy (Section 8): plateau detection without real
  interview outcomes — when same answer type practiced 5+ times and scores
  vary ≤0.3 across 3 consecutive sessions, flag as practice ceiling and
  prescribe behavioral change. New status: practice-calibrating.
- Score Confidence Display (Section 9): scores carry calibration status
  context (Low/Medium/High confidence) so candidates don't treat estimates
  as ground truth
- De-escalation Check: root causes are explicitly resolved when affected
  dimensions improve, preventing over-diagnosis drift
- Explicit transition rule: practice-calibrating → calibrating when 3+
  real outcomes arrive

transcript-formats.md:
- Added Fireflies.ai as 9th supported format with disambiguation rules
  separating it from Grain
- Zoom Webinar distinction: Panelist: prefix handling
- Format Fingerprint for ambiguous cases: 5-signal fingerprint + medium-
  confidence detection decision, gives candidate basis to override
- Granola quality signal: AI summary cross-checked against raw transcript,
  flags possible truncation or hallucination

decode.md:
- Pathway-weighted batch triage: Referral / Warm intro / LinkedIn /
  Cold / Internal dimension added to ranking. A Long-Shot Stretch with
  a Referral ranks above an Investable Stretch cold application.
- JD Honest Limitations surfaced in every Standard/Deep output as a named
  section (not just reference knowledge). Quick Scans include a one-line
  reminder. Candidates never walk away treating a decode as a complete picture.

feedback.md:
- Memory reliability tagging for Type D (Post-Session Memory): Same-day /
  1-2 days / 3+ days tags applied to Question Bank entries
- Guided elicitation prompts for thin memories: surfaces what's actually
  there rather than accepting vague recall as complete capture
- Rejection Leverage defers to graduated Challenge Protocol levels instead
  of hardcoding Level 5 behavior

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>