FOR COMPLIANCE OFFICERS · THE RUN PLANE

End-to-end governance for AI work — auditable, signed, built for regulatory review.

Every Skill chain, every agent run, every tool call lands in one cryptographically chained audit log aligned with NIST AI RMF, ISO/IEC 42001, EU AI Act, GDPR Article 22, and SOC 2 evidence expectations. Export as a cryptographically signed Evidence Pack your auditor can verify offline — without our server.

This is the Run plane — governing how your business operates with AI. Need to govern how your developers build with AI? See the Build plane →

See it in 5 minutes → Read the architecture

Control mapping out of the box

Aligned with NIST AI RMF, ISO/IEC 42001, EU AI Act, GDPR Article 22, and SOC 2 evidence expectations — pre-mapped per skill, with per-org overrides. Closed-set hallucination guard cross-checks every cited control ID against a 335-control truth set so fabricated references can't pass. Compliance team owns the policy; engineering team owns the workflow.

NIST · ISO 42001 · EU AI Act · GDPR · SOC 2

Cryptographically signed audit

Every audit row is part of a tenant-scoped chain so cross-tenant tampering is structurally detectable. Evidence Pack export signs the manifest; a regulator can verify the pack offline using a standalone verifier that is open source under the Apache License 2.0 — no Proofpane backend required.

Standalone offline verifier (open source, Apache-2.0)

Workflow + agent unified

Skill chains (intake → triage → review) and agent runs go through one policy gate, one audit chain, and one cost gate. Partial governance is no governance — auditors don't accept partial.

End-to-end governance thesis

"Why is this your production default?" — one URL, frozen JSON.

The agent's memory strategy becomes the default only by passing (a) a statistical significance gate over a content-hashed fixture; where the deployment turns it on, (b) an inter-rater reliability floor (Krippendorff α with bootstrap CI — the same measure clinical-trial reviewers use to prove humans agree above chance) gates automatic promotion on top, and below the floor the decision stays human-reviewed. A skill's model becomes its default when a named person accepts a comparison recorded on your own cases, or when an auto policy applies one the comparison classed improved. The verdict, the confidence interval, the fixture hash, the DLP rule-set fingerprint, the approving operator — all frozen on the audit row. Your auditor reconstructs why this is the current default from the bundle alone. No meeting required. No engineer dragged in. Six years from now, same answer, same hash.

# /promotions row — frozen at promote time, replayable forever
{
  "dimension": "memory_strategy",
  "winning_variant": "simple",
  "verdict": {
    "winner": "simple",
    "rank": [1, 2],
    "per_variant_retention": [0.91, 0.29]
  },
  "significance": {
    "effect_size": 0.620,
    "ci_low": 0.215,
    "ci_high": 0.890,
    "n": 22
  },
  "evidence": {
    "fixture_hash": "8a3f72…d401",
    "dlp_rules_version": "b1c43e…ff02",
    "audit_run_id": "exprun_a1b2c3…"
  },
  "measurability": {
    "alpha_HH": 0.86,
    "ci": [0.81, 0.91],
    "band": "auto-eligible"
  }
}

Why teams pick us

The product surface isn't load-bearing — the substrate is. These three properties are what hold up under a regulator's third question, not the first.

Multi-tenant by construction

Every record carries its tenant identifier and the audit chain is scoped per tenant, so cross-tenant tampering is structurally detectable — not just policy-detected. The isolation contract is pinned by a comprehensive test suite that runs on every change.

Multi-tenant isolation, pinned by tests

Self-improving policies, approval-gated

When quality drifts, the system proposes a prompt edit, sandboxes it against past failing samples, and asks an admin to approve before going anywhere near production. Every approval carries approver, rationale, and linked diff in the audit chain. Reversible. Never auto-applied.

Approval-gated prompt evolution

Open verifier — checking evidence needs no vendor

The Evidence Pack verifier is a standalone script, open source under the Apache License 2.0, so your auditor can read what it checks. Your auditor downloads it once and verifies any pack offline — forever, even if Proofpane disappears. Other vendors' compliance reports need their server to validate. Ours don't.

Standalone offline verifier (open source, Apache-2.0)

What the auditor receives

One signed zip per time range. The manifest enumerates every event in the slice, the chain hash at start and end, and the compliance coverage matrix. The verifier is standalone and open source under the Apache License 2.0 — download a real sample pack and check it yourself.

# Verify offline — no Proofpane server required.
$ proofpane-verify q3-2026-evidence.zip
✓ Pack version: evidence-pack/1.8
✓ Manifest signature valid
✓ Audit chain intact: no gaps
✓ Control mapping observed:
    NIST AI RMF  → all configured controls
    ISO 42001    → all configured controls
    EU AI Act    → all configured obligations
    GDPR         → Article 22 obligations
    SOC 2        → Trust Services Criteria
✓ Tenant isolation: confirmed
✓ Runtime intervention chain: 4 pauses, all resolved
✓ Pack is verified.

Are you an engineer, not a compliance officer? See the developer-facing pitch →