What CIRCUIT Is

Part 1 argued that existing AI governance is paperwork around a black box, and that the research community has spent the last eighteen months making the black-box excuse untenable. That is the gap. This is the framework.

CIRCUIT — Circuit Informed Risk & Control Understanding, Inventory & Transparency — is an open standard for AI interpretability risk management. It is deliberately small. The whole thing fits on a poster. That is the point. A framework a CISO cannot explain to a board in ten minutes does not get adopted.

The Score — Interpretability Maturity Score (IMS), 0 to 5

Existing risk tiers (Low, Moderate, High) classify intended consequence. They answer: if this system goes wrong, how bad is the blast radius? They do not answer: what do we know about why it might go wrong?

IMS answers the second question. It is a ratchet. Each level requires all the evidence of the levels below it. You cannot skip. You do not declare a level. You produce the artifacts that prove it.

IMS Floor Requirements by Tier

Minimum IMS levels required before a model may be deployed at each risk tier. Below-floor operation triggers Rule 8: remediation must be complete within 180 days.

The Circuit Risk Score — One formula, four bands

The Circuit Risk Score is what goes on the executive dashboard. It is the single number that tells you whether a model's interpretability posture is good enough for the job you are asking it to do.

The three multiplicands:

• Risk Tier: Low = 1, Moderate = 2, High = 3, Critical = 4.
• (6 − IMS), the interpretability deficit. IMS 5 yields a multiplier of 1 (no deficit); IMS 0 yields 6 (maximum deficit).
• Decision Consequence Weight: Advisory = 1, Recommended = 2, Automated = 3, Irreversible = 4, Catastrophic = 5.

Worked Examples

The Six KPIs

The score ratchets on evidence, and the evidence ratchets on measurements. CIRCUIT standardizes six KPIs. Each one maps to a question that matters in an incident.

KPI Floors at High Tier

The six KPIs require baselining for all deployments. For High-tier models, the following minimum thresholds apply:

The Three Model Categories — and Honest Ceilings

The Ten Hard Rules

The ten rules are the binding part of CIRCUIT. They are short on purpose.

• 01Every deployed model gets a registry entry with identity, owner, risk tier, IMS, and CRS, before it serves a single production request.
• 02IMS is evidence, not assertion. A level claim requires the named artifacts on file in the registry.
• 03CRS drives the approval ladder. Green auto-approves; Amber requires AIGC review; Red requires CISO and AIGC sign-off; Purple is not deployable. Consequence floor: any deployment at DCW ≥ 4 (Irreversible or Catastrophic) on a High- or Critical-tier model is governed at no less than Red, regardless of computed CRS.
• 04Six KPIs are baselined before production and re-measured on every material update.
• 05Every material model update triggers a new registry version and a stability KPI measurement.
• 06Circuit-aware (not behavioral-only) red team: semi-annually for High- and Critical-tier models, annually for Moderate-tier, not required for Low-tier.
• 07ACFR ≥ 1 in any model-quarter forces a Red band transition, a mandatory governance review within 30 days, and Purple-band escalation if ACFR is not declining at subsequent reviews within the 180-day remediation clock. ACFR is a count of P1 safety-circuit bypasses per model-quarter — not a ratio. High-tier expectation is 0.
• 08Vendor Transparency section is not optional. Partial responses receive IMS credit proportional to the sections answered. If a vendor has not responded within 60 days of two documented outreach attempts, the non-response is recorded in the registry and CRS uses IMS 0 until a response is received.
• 09Category C models cannot host High-tier autonomous or irreversible workflows. Period.
• 10Registry entries are preserved for the regulatory retention horizon — a minimum of three years.

The Registry — Eight Sections, YAML, Git-backed

The registry is not a new tool. It is a schema. The reference implementation is YAML because your platform team can diff YAML in pull requests. The eight sections, in order:

1. Identity. Model name, version, vendor, category (A, B, or C), owner, business process, upstream dependencies, hosting location, data classification of inputs.
2. Maturity. Current IMS, evidence inventory (artifact names, hashes, dates), history of level transitions.
3. Circuit Inventory. Named circuits relevant to the use case. For each: size, edge count, monosemanticity score, and link to the artifact.
4. KPI Baseline. The six KPIs with measurement date, methodology, thresholds, and current values. Time series.
5. Vendor Transparency. "Show Me Your Circuits" responses with Acceptable, Partial, or Unacceptable coding for each of the 29 questions.
6. Red Team. Latest circuit-aware red team report, MITRE ATLAS technique coverage, findings, remediation status, next scheduled engagement.
7. Compensating Controls. For any model in Amber or Red, the specific controls reducing residual risk.
8. Lifecycle. Deployment date, material update log, planned retirements, rollback procedures, incident history with ACFR contributions.

circuit_registry_entry:
  version: "1.1.0"

  identity:
    name: "fraud-classifier-v3"
    vendor: "internal"
    category: "A"           # open weights, self-hosted
    risk_tier: "High"
    owner: "security-team@company.com"
    consequence: "Automated"

  maturity:
    ims: 3
    evidence:
      - artifact: "fraud_sae_gemma3_9b_v3.pkl"
        type: "sparse_autoencoder"
        date: "2026-03-15"

  kpi_baseline:
    circuit_size: 142
    edge_count: 318
    monosemanticity: 0.87
    robustness: 0.82
    stability_across_versions: 0.91
    acfr_last_quarter: 0     # count of P1 safety-circuit bypasses this quarter

  crs:
    calculated: 27        # 3 × (6-3) × 3 = 27  (High=3, deficit=3, Automated=3)
    band: "Amber"

The Dashboard — Four Views, One Data Model

The dashboard is a reference implementation, not a product. Four views:

• Fleet view. Every model on one page, plotted on a CRS × IMS grid, colored by band. Dot size equals consequence weight.
• Per-model deep dive. One model, all eight registry sections rendered. KPI time series with thresholds overlaid.
• Executive summary. One-page PDF render per month. Five numbers, one chart, one list.
• Regulatory posture. Each row is a framework requirement; each column is a model; each cell shows the evidence link. Auditors love this view.