Skip to content

MII CALIBRATION

Mobius Integrity Index (MII) — Calibration v0.1

MII is a composite score measuring how well a decision's provenance supports accountability, reversibility, and contestation.

MII is calibrated, not revealed.

MII Is a Vector (Implemented as a Scalar)

Treat MII as a weighted sum of components:

Component Weight Measurement
Provenance Completeness (PC) 30% % of required EPICON fields present
Test/Policy Pass Rate (TP) 20% Automated checks passed
Review Quorum (RQ) 20% Required human sign-offs met
Sentinel Challenge Score (SC) 20% Inverse of severity/flags
Drift Score (DS) 10% Semantic distance from prior similar decisions

Formula: MII = 0.3*PC + 0.2*TP + 0.2*RQ + 0.2*SC + 0.1*DS

Why 0.95 Is the Default Threshold

0.95 is a policy default, not a law.

  • 0.90 tends to allow sloppy attestations.
  • 0.95 balances throughput and legitimacy.
  • 0.99 may be too strict for high-velocity workflows.

The project will replace this intuition with data via pilot studies.

Action Semantics (What Happens When MII Drops)

MII Range System Response Override
≥ 0.95 Allow staging merge if Sentinel passes None
0.90–0.94 Warn + require justification Single sign-off + EPICON rationale
0.85–0.89 Block Dual sign-off + EPICON override
< 0.85 Circuit breaker + alert Emergency EPICON + post-mortem

Overrides are recorded as first-class attestations.

Contextual Thresholds (Per Workflow)

MII thresholds can be tuned per workflow based on risk tolerance.

Example: Low-Risk Configuration Updates

{
  "workflow_id": "config-update",
  "mii_threshold": 0.90,
  "required_sign_offs": 1,
  "sentinel_review_required": false,
  "rationale": "Speed > exhaustive provenance for non-critical configs"
}

Example: High-Risk Policy Changes

{
  "workflow_id": "content-moderation-policy",
  "mii_threshold": 0.98,
  "required_sign_offs": 3,
  "sentinel_review_required": true,
  "rationale": "Legitimacy > velocity for user-facing policy"
}

Calibration Methodology

Phase 1: Baseline (Current)

  1. Default threshold: 0.95
  2. No historical data for calibration
  3. Components weighted by intuition

Phase 2: Data Collection (Pilots)

  1. Run pilot workflows with MII tracking
  2. Record:
  3. MII score at decision time
  4. Post-hoc audit outcomes
  5. Rollback frequency
  6. Time-to-reconstruct intent
  7. Correlate MII with outcomes

Phase 3: Calibration (Post-Pilot)

  1. Adjust weights based on predictive power
  2. Adjust thresholds based on outcome correlation
  3. Publish calibration results

Component Details

Provenance Completeness (PC)

Required EPICON fields: - epicon_id - title - author_name - cycle - scope - justification (with VALUES, REASONING, ANCHORS, BOUNDARIES, COUNTERFACTUAL) - counterfactuals

PC = (fields present / fields required) * 100%

Test/Policy Pass Rate (TP)

  • CI tests passed
  • Lint checks passed
  • Type checks passed
  • Policy compliance verified

TP = (checks passed / checks required) * 100%

Review Quorum (RQ)

  • Required approvals obtained
  • CODEOWNERS satisfied
  • Sentinel review completed (if required)

RQ = (approvals obtained / approvals required) * 100%

Sentinel Challenge Score (SC)

Inverse of flag severity: - No flags: SC = 1.0 - Low severity: SC = 0.9 - Medium severity: SC = 0.7 - High severity: SC = 0.4 - Critical severity: SC = 0.1

Drift Score (DS)

Semantic distance from prior similar decisions: - No drift detected: DS = 1.0 - Minor drift: DS = 0.8 - Significant drift: DS = 0.5 - Major drift: DS = 0.2

Falsifiability (How to Prove MII Wrong)

Run N decisions through Mobius and compare: - MII score - Post-hoc audit outcomes - Rollback frequency - Time-to-reconstruct intent

If low-MII decisions do not correlate with worse audit outcomes, MII is broken.

If high-MII decisions still produce legitimacy collapse, MII is incomplete.

MII is a hypothesis. Help falsify and improve it.

Override Recording

All overrides are recorded with:

{
  "override_id": "OVERRIDE-2026-01-24-001",
  "original_mii": 0.87,
  "threshold_required": 0.95,
  "override_type": "dual-sign-off",
  "justification": "Emergency fix for production incident",
  "approvers": ["human:approver1", "human:approver2"],
  "epicon_ref": "EPICON_C-202_EMERGENCY_incident-fix_v1",
  "timestamp": "2026-01-24T15:30:00Z"
}

Governance of MII Parameters

Changes to MII formula, weights, or default thresholds require:

  • RFC filed in docs/11-SUPPLEMENTARY/rfcs/
  • 72-hour minimum review window
  • ≥2 maintainer approvals
  • EPICON documenting intent and rationale

See docs/RFC_PROCESS.md for details.

"We heal as we walk." — Mobius Substrate