REMORA: governed agentic AI.

REMORA should be positioned as a governance control plane for agentic AI: it decides when probabilistic model output is reliable enough to answer, when evidence must be consulted, when action must be blocked, and when a human must approve.[S1]

Its strongest validated value is selective reliability and safe tool execution — not universal full-coverage QA superiority. The right metric is governed action quality.

Claims in this whitepaper are scoped to the uploaded technical report and the public REMORA repository at head current main branch (v0.9.0).

Once a model proposal becomes a tool call, the cost of a wrong answer changes shape. The minimum control plane therefore needs consensus, uncertainty, evidence, policy and audit — together, not in isolation.

The decision function is expressed as D = g(C, U, P, E, R), producing one of ACCEPT, VERIFY, ABSTAIN or ESCALATE. Session stability is tracked by V(t) = H(t) + λD(t), a monitor for entropy and dissensus growth — not a proof of correctness.

The control-room route demonstrates the operational version of the framework. When REMORA cannot approve or reject safely, the reviewer can request site verification instead of forcing a binary decision.[S14]

The simulated lifecycle is PENDING_REVIEW, SITE_VERIFICATION_PENDING, EVIDENCE_RECEIVED, READY_FOR_REVIEW, then final approval, rejection or closure. Similar-case history can create a policy-learning candidate, but it cannot change policy without owner approval.

See Evidence for calibration and negative results. [S6]

• False consensus
  Multiple models repeat the same wrong claim.
  Diversity tracking and evidence route.
• Prompt injection
  Tool output contains adversarial instructions.
  PreToolUse hook re-evaluates policy on tool input.
• Compromised model
  An oracle model is silently swapped or fine-tuned.
  Audit graph records provenance and version hash.
• Policy drift
  Rules are loosened without review.
  Signed policy bundles and change audit.

REMORA is a policy and verification boundary in front of agent tools.[S12]

• [S1]paper/whitepaper.md — current technical framing.
• [S2]Repository quality gates — Ruff, claim consistency, result snapshot, pytest, frontend lint and frontend build.
• [S3]Cascade module — Oracle pool and canonicalization.
• [S6]OracleDiversityTracker — rolling pairwise agreement.
• [S12]Audit envelope — RDF graph + OTel spans.
• [S14]Control Room GUI — deterministic review-loop demo with follow-up request envelope.
• [S13]docs/thermodynamics/claim_ledger.yaml — claim status classes.
• [E1]NIST AI Risk Management Framework — Govern / Map / Measure / Manage.