50-verifier

A set of rules for writing verifiers, the tests that grade a coding-agent benchmark submission. It covers protected expected answers, requirement-based assertions, tolerances, and preventing submissions from cheating.

In plain words
What is it for?
Use it to place ground truth in protected test files, write behaviour-focused assertions, justify tolerances, design anti-cheat checks, and review verifier coverage before submission.
Why use it?
It helps ensure tests check only what the task asks for and that agents cannot simply read the expected answers. This makes benchmark results more reliable and fair.

Cursor rule for Cursor

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 rules/xclaw-bot/benchmark-task-authoring/50-verifier
Clone the repo
git clone --depth 1 https://github.com/Xclaw-bot/benchmark-task-authoring

Made for: Cursor.

Per session 1,016 This file is loaded in full into every session.
When invoked 1,016 The same file — it is already loaded in full.
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 $0.01016 $0.01016
Opus 5 $0.00508 $0.00508
Sonnet 5 $0.00203 $0.00203
Haiku 4.5 $0.00102 $0.00102

Measured 2d ago against content hash 95cac74d33cf, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

50-verifier 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 2d 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.

.cursor/rules/50-verifier.mdc · 91 lines

How it starts

The opening of the file, as written. The whole thing — 91 lines — stays where its author put it; the contents beside it link to each section on GitHub.


description: Verifier design: ground-truth protection, tolerance justification, anti-cheat, and the pre-submission self-check. Auto-attaches to test files. globs: /task/tests/ alwaysApply: false

Verifier craft

The verifier runs in the same image as the agent, after the agent stops. Harbor overlays tests/ at /tests only at verify time. Everything it needs must be baked into environment/Dockerfile (pinned) — test.sh installs and downloads nothing.

Ground truth must be unreachable

❌ truth read from a path the submission can also open
✅ truth lives where the submission cannot reach it — tests/, overlaid at verify time

Expected values belong in tests/, never in environment/, never in the image. Prefer values derived in the solution over hardcoded constants, so a non-specialist reviewer can follow the derivation.

Assertions

  • Every assertion traces to a requirement stated in instruction.md; every stated requirement has coverage. Grading on unstated criteria is an unfair stump.
  • One requirement per test function, each with a docstring naming the behaviour it checks in the instruction's terms.
  • Verify behaviour by execution, never by scanning source for keywords, imports or regexes — an agent can insert a keyword without implementing anything.
  • Existence and formatting checks should rarely fail for a competent solver. If they are what fails, the defect is in the instruction.

Tolerances

Exact comparison is preferred and needs no defence when the pipeline is integer-deterministic and the format is normatively pinned. Any inequality-based check — ranges, thresholds, percentile bounds, fuzzy matching — must have its calibration justified in verification_explanation: what the band brackets, which legitimate variation it absorbs (float precision, alternative valid algorithms, quadrature, rounding), and whether it was validated against an alternative correct method rather than only the reference. A band only the reference can hit is too tight; a band that admits obviously wrong answers is too loose. Bare statements like "accepted range [29, 31]" are insufficient.

Read the full file on GitHub · 91 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. 2d ago First seen · 91 lines · 1,016 tokens per session scan A 95cac74d33cf

Subscribe to this mod's changes

50-verifier is a cursor rule published in the GitHub repository Xclaw-bot/benchmark-task-authoring (2 stars, last pushed 18d ago), licensed MIT. It adds 1,016 tokens to every session, about $0.0051 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.