Intel Echo reads your AI conversation and shows you exactly where the reasoning went off-mandate — with the verbatim sentence as evidence. One CLI command. Zero guessing.
rm -rf files facefile2 facefiles zip
AI agents fail two ways. You know the first: hallucination, wrong output. The second is harder — the agent confidently did something adjacent to what you asked. It redefined the goal, claimed authority it wasn't given. Your traces show it ran. Your evals say the output looks fine. Nothing flags the mandate breach.
5 turns of uncaught drift → 3–8 correction cycles → 10,000–40,000 tokens of waste per session. And the output may still be wrong because the mandate was never restored.
An agent that exceeded its mandate didn't just produce bad text. It issued a refund, drafted a clause, committed code, deleted files — actions that must now be unwound. The cost is not the tokens. It's the consequence.
Your CI passes because it checks whether the output was right. It does not check whether the agent stayed inside what it was asked to do. Those are different questions — and only one of them is being asked.
Intel Echo is a CLI. You give it a transcript. It gives you a report. Under the hood, a single model call runs the witness — it reconstructs the mandate at each turn before judging drift.
Human-readable findings with quoted evidence. Read it directly, share with your team.
Machine-readable. Pipe into CI, tooling, dashboards. The full schema with confidence scores.
One row per finding. Two empty verdict columns for human annotators. The dataset grows here.
No server. No database. No dashboard. The transcript is the source of truth.
Every finding includes the exact quoted span, the active mandate at that turn, and the auditor's confidence. Findings that wouldn't change your next decision are hidden — noise is the cardinal failure.
The reasoning answers a different question than the one in force, or redefines the objective mid-stream without the human asking it to.
Claims standing not granted — characterises the user's state, asserts a reading as settled fact, or acts as source-of-truth beyond the mandate.
Information from outside the actual evidence leaks in and is treated as if it came from the transcript — order, adjacency, prior registers.
Stated certainty exceeds the support available. Hedges stripped; readings asserted as established fact; specific figures asserted with no basis.
Drop Intel Echo into your existing workflow. No new dashboard to babysit.
Run audit on any markdown transcript. Or use the free prepare → ingest flow with Claude in Cowork — no API key, no metered cost.
intel-echo audit session.md
One block in .mcp.json. Claude calls audit_transcript() natively without leaving the editor. The witness runs inside your coding environment.
audit_transcript()
Exits non-zero if mandate violations appear as actionable findings. Drop into any GitHub Actions workflow. Gate merges on clean reasoning.
--fail-on authority_overreach
The intelligence layer (this tool) audits reasoning compliance. GTM compresses sales correction loops. Run them together for the full AI cost stack.
See Intel Echo GTM →Intel Echo improves by remembering its own mistakes — not by retraining. The labeled ontology is the moat.
Two annotators independently mark each finding keep or kill. Precision score computes automatically. You own the dataset.
A killed finding (false positive) goes into taxonomy.js as a known_false_positive. The witness stops repeating that pattern permanently — across all future audits.
A false flag costs more than a miss. A tool that cries wolf gets uninstalled. Precision is the metric that matters.
The witness is one model call. It can miss subtle drift and occasionally over-flag. That is why correctness is defined by human-confirmed labels, not by the tool's own confidence.
This core is v0. The CLI and pilot faces come next. An LLM auditing an LLM shares failure modes. The mandate step, the evidence-quoting requirement, and the decision-relevance gate reduce that — but do not eliminate it. Treat findings as prompts for human judgment, not verdicts.