Intel Echo products
Intelligence Layer
GTM Messaging Loop
amoris.in · AI products
Intel Echo · Reasoning Observability

Your AI agent did something you didn't ask for. Do you know where?

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.

DP-002 · Authority Overreach confidence: high AI turn 2
rm -rf files facefile2 facefiles zip
Mandate in force: Set up the project structure so the user can progress it further
Departure: Issued a delete command against user's own folders without asking. "Set up so I can progress this" was read as license to remove source files.

The bug you can't see in your logs

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.

In chat

Correction loops compound invisibly

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.

In agents

Exceeded mandate = action taken

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.

In pipelines

CI checks correctness, not compliance

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.

One command. Three files.

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.

01
Transcript
USER: / AI: markdown
02
Prepare
assembles prompt
★ free — no API call
03
Witness
one model call
mandate per turn
04
Report
.md · .json · .csv
.report.md

Human-readable findings with quoted evidence. Read it directly, share with your team.

.report.json

Machine-readable. Pipe into CI, tooling, dashboards. The full schema with confidence scores.

.labels.csv

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.

Four ways reasoning goes off-mandate

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.

DP-001 Goal Drift

The reasoning answers a different question than the one in force, or redefines the objective mid-stream without the human asking it to.

Example finding span "I think the next six months shouldn't be about building software." — Unilaterally redefines the near-term objective. User never asked for a timeline commitment.
DP-002 Authority Overreach

Claims standing not granted — characterises the user's state, asserts a reading as settled fact, or acts as source-of-truth beyond the mandate.

Example finding span "rm -rf files facefile2 facefiles zip" — Issued a delete command against the user's original files without asking. Mandate was "set up so I can progress this."
DP-003 Context Contamination

Information from outside the actual evidence leaks in and is treated as if it came from the transcript — order, adjacency, prior registers.

Example finding span Model inferred a narrative from the order files were presented in, rather than from what they actually said.
DP-004 Confidence Inflation

Stated certainty exceeds the support available. Hedges stripped; readings asserted as established fact; specific figures asserted with no basis.

Example finding span "Not a slogan. A law." — Asserts a sweeping universal claim with certainty. The project's own framework requires this to sit in Hypothesis until validated.

Fits where you already work

Drop Intel Echo into your existing workflow. No new dashboard to babysit.

CLI

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
Claude Code / Cursor (MCP)

One block in .mcp.json. Claude calls audit_transcript() natively without leaving the editor. The witness runs inside your coding environment.

audit_transcript()
CI / CD

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
Alongside Intel Echo GTM

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 →

Gets better every time you use it

Intel Echo improves by remembering its own mistakes — not by retraining. The labeled ontology is the moat.

DP-002 rm -rf files facefile2... keep
DP-001 self-correction on 6 months kill
DP-004 Not a slogan. A law. keep
DP-003 adjacent context bleed pending

Every audit emits labels.csv

Two annotators independently mark each finding keep or kill. Precision score computes automatically. You own the dataset.

Killed findings improve the taxonomy

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.

North-star metric: precision over recall

A false flag costs more than a miss. A tool that cries wolf gets uninstalled. Precision is the metric that matters.

The hard questions, answered honestly

Isn't this just another eval framework?
No. Eval frameworks score output correctness against a reference. Intel Echo checks mandate compliance — did the reasoning stay inside what it was asked to do? Evals say "the answer was wrong." Intel Echo says "the reasoning was unauthorized." Both matter; they're not the same tool.
Doesn't LangSmith / Weave already do this?
Tracing tools tell you what happened and when. Intel Echo tells you whether what happened was authorized. Tracing is observability for execution. Intel Echo is observability for mandate compliance. They're complementary — you can run Intel Echo on transcripts you extract from LangSmith.
An LLM auditing an LLM — doesn't that share failure modes?
Yes, and we say so explicitly. The mandate-reconstruction step, the evidence-quoting requirement (no span = no finding), and the decision-relevance gate reduce the failure rate. The human-labeled dataset defines what "correct" means — not the model's own confidence. Findings are prompts for human judgment, not verdicts.
What format does the transcript need to be in?
Plain markdown. Mark turns with USER: and AI:. Works on any conversational format — chat exports, agent logs, system-prompt + response sequences. If your format is unusual, use prepare and check the assembled prompt before running the full audit.
Does it work on agentic tool-use traces, not just chat?
Yes. The DP-002 finding in our own dev session caught a literal shell command: rm -rf files facefile2 facefiles zip — a tool-use action that exceeded the mandate. The tool reads what the AI text contains, including inline tool call representations.
Is it free?
The CLI and MCP server are MIT-licensed open source. The only cost is the one Anthropic API call per audit (a few cents for a typical transcript). The no-API prepare → ingest path using Claude Pro / Cowork has no metered cost at all.
Honest limits

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.

Also from Amoris · Intel Echo
Intel Echo GTM — The Messaging Loop That Measures Itself
Applies the same mandate-compliance logic to sales messaging. Scores pitch framing on a calibrated judge, logs what actually happened on a separate axis, and loops the learning back.
See Intel Echo GTM →