20-design-doctrine

20-design-doctrine is a cursor rule for Cursor from Xclaw-bot/benchmark-task-authoring. It costs 1,151 tokens per session, scanned A, original, MIT.

A set of measured findings for designing fair coding benchmarks, including which task ideas failed and what accepted tasks look like.

In plain words
What is it for?
Use it when proposing or selecting benchmark tasks and when reviewing whether their difficulty can be verified independently.
Why use it?
It prevents authors from relying on an AI agent simply failing to check its own work when judging task difficulty.

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/20-design-doctrine
Clone the repo
git clone --depth 1 https://github.com/Xclaw-bot/benchmark-task-authoring

Made for: Cursor.

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 20-design-doctrine

README.md
[![agentmods](https://agentmods.dev/badge/rules/xclaw-bot/benchmark-task-authoring/20-design-doctrine.svg)](https://agentmods.dev/rules/xclaw-bot/benchmark-task-authoring/20-design-doctrine)
Your own site
<a href="https://agentmods.dev/rules/xclaw-bot/benchmark-task-authoring/20-design-doctrine"><img src="https://agentmods.dev/badge/rules/xclaw-bot/benchmark-task-authoring/20-design-doctrine.svg" alt="Measured on agentmods" height="20"></a>
Per session 1,151 This file is loaded in full into every session.
When invoked 1,151 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 $0.01151 $0.01151
Opus 5 $0.00575 $0.00575
Sonnet 5 $0.00230 $0.00230
Haiku 4.5 $0.00115 $0.00115

Measured 3d ago against content hash b88186dab45e, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

20-design-doctrine 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.

.cursor/rules/20-design-doctrine.mdc · 86 lines

How it starts

The opening of the file, as written. The whole thing — 86 lines — stays where its author put it; the contents beside it link to each section on GitHub.


description: Measured dead ends and what the accepted-task corpus looks like. Load before proposing or choosing a task shape. alwaysApply: false

Design doctrine — what is measured, not theorised

Source: the FIELD NOTES sections of ../hardness-laws.md. Every claim here came from a build that was carried to measurement.

The decisive finding

The benchmark model does not depend on supplied samples to verify its work. It writes a second independent implementation and cross-checks against it. Observed directly in trial analyses: "it built two independent solvers … both agreed on all 8 sample positions and all 32 graded positions before the final artifact was written." It also derives correct semantics from a spec and self-corrects the exact bug a corpus was built around.

The structural consequence:

Any task specified completely enough to be fair is self-verifiable. Corpus and sample selection cannot create a gap, because the gap they exploit is "the agent stops while still wrong" — and this model does not stop.

Design accordingly. A difficulty story that depends on the agent failing to check its own work is already refuted.

Dead ends — do not retry these

Trap Measurement
Latent crux via sample selection (samples blind to the wrong method) Works on paper, irrelevant in practice — the model verifies independently of the samples.
Making the deciding rule rare Rare hides it from the evidence and from the answer; both move together. Measured twice: rarer scoring clause moved log-visibility 76%→93% while answer-divergence fell 17%→3%.
More interacting rules A 10-clause engine measured as a 2-clause one: 6 of 8 rule-omission variants diverged on 0% of games, because the clauses never fired.
Bigger state spaces 11× growth (6k → 67k states) changed nothing; still solved.
Value-identical error modes Provably blind value-checking still lost 2/2 — the model wrote a second solver.
Format reproduction from a corpus The corpus gives complete feedback: encode, diff bytes, localise, fix, repeat. Bit-packing makes it worse — any width or ordering error shifts every later bit, so all four wrong encoders were 100% visible.
Scale as a validation barrier Dead, and provably so from our own records: brute force was already infeasible at the shipped size and the agent still solved it 4/5 — because its second implementation was never brute force, it was a second retrograde solver. Two independent correct solvers agree at any size.

Read the full file on GitHub · 86 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 · 86 lines · 1,151 tokens per session scan A b88186dab45e

Subscribe to this mod's changes

20-design-doctrine is a cursor rule published in the GitHub repository Xclaw-bot/benchmark-task-authoring (2 stars, last pushed 19d ago), licensed MIT. It adds 1,151 tokens to every session, about $0.0058 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.