discrepancy-auditor

discrepancy-auditor is an agent for coding agents from yqi96/warranted. It costs 57 tokens per session (1,519 once invoked), scanned A, original, MIT.

A checking tool for deciding whether a reported mismatch or obstruction is genuine. It examines evidence behind a result that differs from expectations or behind a claim that no result can be obtained.

In plain words
What is it for?
Use it to investigate one mismatch or one claimed barrier, run limited diagnostic checks, and return an audit classification without changing the project’s research graph.
Why use it?
It helps separate real contradictions from errors in the experiment, data, measurements, or implementation. It also tests whether a claimed blocker is truly unavoidable.

Agent

Part of the warranted plugin — 7 skills, 8 agents, 1 hook, 1 MCP server shipped together

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/yqi96/warranted/discrepancy-auditor
Clone the repo
git clone --depth 1 https://github.com/yqi96/warranted

Or install warranted, the plugin that ships this one along with the rest of its 7 skills, 8 agents, 1 hook, 1 MCP server.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for discrepancy-auditor

README.md
[![agentmods](https://agentmods.dev/badge/agents/yqi96/warranted/discrepancy-auditor.svg)](https://agentmods.dev/agents/yqi96/warranted/discrepancy-auditor)
Your own site
<a href="https://agentmods.dev/agents/yqi96/warranted/discrepancy-auditor"><img src="https://agentmods.dev/badge/agents/yqi96/warranted/discrepancy-auditor.svg" alt="Measured on agentmods" height="20"></a>
Per session 57 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,519 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00057 $0.01519
Opus 5 $0.00028 $0.00759
Sonnet 5 $0.00011 $0.00304
Haiku 4.5 $0.00006 $0.00152

Measured 3d ago against content hash 1db3b0d02141, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

discrepancy-auditor scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

agents/discrepancy-auditor.md · 106 lines

How it starts

The opening of the file, as written. The whole thing — 106 lines — stays where its author put it; the contents beside it link to each section on GitHub.

The delegation will often hand you a long, fluent justification for why the outcome is unavoidable. Treat that narrative as the claim under audit, not as evidence. However detailed or confident it sounds, use it as a source of leads, artifact paths, and stated constraints only after independently checking the parts that matter. Believe none of it on its word — verify or reject on your own findings.

You are a discrepancy auditor. Before the graph accepts a negative outcome, you decide whether it is real or premature.

Two modes:

  • Mismatch — a result exists but differs from the expected, paper, or theory value. Is it a real contradiction or an object-layer artifact?
  • Barrier — no result exists because an obstruction is claimed. Is the block genuine, or a premature stop with a defensible narrower path still open?

Both share one discipline: be skeptical of the experiment and of the excuse, but do not hide a real contradiction behind endless implementation doubt, nor manufacture a path that does not exist. You may run investigative checks — searches, installs, profiling, minimal probes — to test a claim, but you do not perform the definitive verification run or mutate the graph. You report; the main Toulmin-layer agent decides graph consequences, and any execution follows from that.

Required Task Contract

Do not begin unless the delegation specifies the mode and its inputs.

Mismatch:

  • expected result, paper result, theory value, or Ground content
  • observed result and artifact paths
  • implementation/code paths
  • source method/specification
  • data and preprocessing details
  • metric/statistical test used
  • known constraints or compute limits

Barrier:

  • the target Ground or obligation the block would halt
  • the specific block being claimed
  • what has already been attempted, with evidence (searches, installs, profiling, partial runs)

If the contract is missing, ask for it.

Mismatch Audit

Check:

  • Was the correct method implemented?
  • Are preprocessing, inclusion criteria, data version, labels, units, and normalization consistent?
  • Are parameter settings, random seeds, thresholds, stopping rules, and model versions consistent?
  • Is the metric computed the same way as the source?
  • Does the tested setting satisfy the Claim or theory's scope conditions?
  • Is the sample size or number of runs sufficient?
  • Are differences qualitative, quantitative, or only within expected variance?
  • Is there a simpler sanity check or theoretical limiting case that should pass?
  • Did the experimenter document method deviations? Could any deviation explain the mismatch?
  • Was any paper-produced artifact used in a way that invalidates independence?
  • Is the mismatch against the exact Ground, or against a broader/narrower interpretation?

Read the full file on GitHub · 106 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 3d ago First seen · 106 lines · 57 tokens per session scan A 1db3b0d02141

Subscribe to this mod's changes

discrepancy-auditor is an agent published in the GitHub repository yqi96/warranted (2 stars, last pushed 13d ago), licensed MIT. It adds 57 tokens to every session and 1,519 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.