Patent Pending · USPTO #63/984,669

Plan-Level
Verification
for Agent Decisions.

PLV by ThoughtProof: structured plan-level verification for agent decisions. Cascade pipeline, optional EAS attestation, two tiers.
Hosted at verify.thoughtproof.ai. Provider-neutral. Open-source CLI.
One of four separate ThoughtProof products — see Sentinel, PLV, RV and DQL.

Integrate via API → Try pot-cli locally
0
false allows observed in a 120-case validation set (Standard tier)
0 B→A
BLOCK→ALLOW regressions observed in the same 120-case set
$0.04
Standard tier · SERV Nano→Gemini cascade
🏛️ Patent Pending

0 false allows observed in a 120-case validation set (Standard tier). Measurement source: ThoughtProof canonical-120 suite (publication pending). These figures describe specific validation runs on a fixed 120-case set, not a continuous guarantee. A later summary on a different configuration — combined-nano-solo, 2026-05-25 (canonical-120-summary-2026-05-25T14-20-45-771Z) — recorded 1 false allow.

The Problem
Why single-model AI is a liability
in high-stakes environments
👻

Invisible Hallucinations

LLMs produce confident, well-structured nonsense. Without cross-model verification, hallucinated facts look identical to correct ones.

💉

Prompt Injection Blindspots

A single model can't distinguish adversarial inputs from legitimate ones — because it shares the same vulnerability profile with itself.

📋

Compliance Without Evidence

EU AI Act (Art. 9, 13, 14) requires documented human oversight and risk mitigation. A single model output produces no audit trail.

🪞

No Model Can Verify Itself

Self-monitoring produces correlated blindspots. On the public SHADE-Arena benchmark, same-family monitors score near random chance (AUC ~0.50); cross-family monitors score materially higher.

The Solution
How PLV works

A structured plan-level verification process for AI-agent traces. It maps reasoning to explicit plan steps, returns a public ALLOW/BLOCK/UNCERTAIN verdict, and can issue Base EAS attestations (opt-in) for on-chain audit trails.

How it Works
Plan-level verification pipeline
01
📥

Normalize

Standardize input across providers and formats

→
02
🧠

Segment

Convert trace + answer into explicit plan-step evidence

→
03
⚔️

Score

Evaluate each step for support, faithfulness, and hard-fail risk

→
04
⚖️

Cascade

Escalate ambiguous cases from fast models to Sonnet rescue

→
05
📦

Attest

Return ALLOW/BLOCK/UNCERTAIN plus optional EAS attestation

Why it Matters
Built for agents, teams, and compliance
🤖

AI Agents & Autonomous Systems

  • Verify before acting — surface unsupported plan steps pre-execution
  • Optional EAS attestation on Base for on-chain audit trails
  • Roadmap: verification output schema on ERC-8210 (draft, not shipped)
📑

Compliance & Risk Teams

  • Designed to support EU AI Act Art. 9/13/14 workflows
  • Structured Epistemic Blocks you can retain as evidence
  • Verdict, plan-step scores and escalation path recorded for human review
  • Provider-neutral, so the record does not depend on one vendor
🔒

Security & Enterprise

  • Provider-independent — no vendor lock-in
  • BYOK with the local CLI: your keys, your provider calls, your machine
  • The hosted API does receive the packages you send it, and forwards trace content to the model providers in the cascade; data-processing terms are contractual
  • Every verification returns a traceable record of its verdict and evidence
Benchmarks
Measured results, scoped to the runs they came from

Everything below is an observation from a specific run, not a service level. PLV publishes no general accuracy figure. Rows marked Research support the multi-model thesis (cross-family verification vs self-monitoring); they are not PLV Standard-tier product SLAs.

Metric Value Benchmark Layer
False allows 0 observed 120-case validation set (Standard tier) · canonical-120 suite, publication pending · exception: combined-nano-solo summary 2026-05-25 recorded 1 FA PLV product
BLOCK→ALLOW safety regressions 0 observed same 120-case set · Hard Rule clean PLV product
Standard tier SERV Nano→Gemini $0.04 · ~5s typical end-to-end, not an SLA PLV API v2
Thorough tier SERV Nano→Gemini→Sonnet ~15s typical end-to-end, not an SLA PLV API v2
Sabotage detection AUC 0.720 Research · SHADE-Arena hybrid figure · 208 internal runs · honest aggregate often cited as 0.713 @ 151 runs · publication pending Research (thesis)
Faithfulness detection 88.9% Research · RFEval (ICLR 2026) · 3-family OR-rule · 80/90 on 90-sample internal run · not a PLV tier score · publication pending Research (thesis)
Get Started
Two ways to verify.
# Option 1: API v2 (hosted)
curl -X POST https://verify.thoughtproof.ai/v2/verify \
  -H "X-API-Key: tp_v2_..." \
  -H "Content-Type: application/json" \
  -d '{"question":"...","answer":"...","trace":"...","plan_steps":[...]}'

# → { "verdict": "ALLOW", "schema_version": "plv-api-v2.0" }
# Option 2: Local development (CLI)
npm install -g pot-cli

pot-cli verify --input "your-trace.json"

# Evaluate locally before integrating the API
API Documentation → Try pot-cli →