Built by its own night shift

Dagpilot is developed by Dagpilot. Every number below is a live aggregate from the production instance that builds this product — the same pipeline, the same gates, no demo data.

60
tasks run to date
10
completed with passing acceptance
$46
median task cost in model tokens
4.0h
average duration, accepted tasks
14
contract assertions evaluated
14
intent-reading semantic checks

Live data, cached for five minutes. Last refreshed 2026-08-21 03:06 UTC. Aggregates only — no task content leaves the pipeline. Completed counts only tasks whose acceptance evidence passed; a completion the system cannot prove is not shown as proof. One of them, published by its owner, is opened below, item by item.

A run of our own, opened

the pair's dogfood run, 2026-08-10

Orch probe 2026-08-10 - clamp, aimed at the two verdict-path atoms of today

Verified per contract — 18/18 contract items

Duration12m
Steps2
Self-healed1
Contracts18/18
Cost$9.35

The contract, item by item — cycle 1 of 2

#CriterionKindVerdict
CLAMP-01clamp.js must exist at the repository root as specified by the plandeterministic · file_existsPASS
CLAMP-02clamp.test.js must exist at the repository root as specified by the plandeterministic · file_existsPASS
CLAMP-03clamp.js must export a named function clamp with the exact parameters (value, min, max)deterministic · file_containsPASS
CLAMP-04clamp.js must not contain stub markers — the function must be a real implementation, not a placeholderdeterministic · file_not_containsPASS
CLAMP-05clamp.js must use a finiteness check (Number.isFinite or isFinite) to implement the NaN guard for non-finite inputsdeterministic · file_containsPASS
CLAMP-06Core correctness check: the function must implement all four rules with the guard clause ordered before the clamping arithmeticsemanticPASS
CLAMP-07The test file must contain at least 4 test cases (using either test() or it() from node:test) to cover all four behavioral rulesdeterministic · regex_match_countPASS
CLAMP-08Test file must import from node:test to use the Node.js built-in test runnerdeterministic · file_containsPASS
CLAMP-09Test file must import clamp from the local module to actually test itdeterministic · file_containsPASS
CLAMP-10Test file must import node:assert to make real assertions; without it tests pass vacuously with no verificationdeterministic · file_containsPASS
CLAMP-11Rule 1 requires dedicated test coverage: value below min must return minsemanticPASS
CLAMP-12Rule 2 requires dedicated test coverage: value above max must return maxsemanticPASS
CLAMP-13Rule 3 requires dedicated test coverage: value in range must be returned unchangedsemanticPASS
CLAMP-14Rule 4 covers all non-finite inputs; testing only one category (e.g. only NaN) could miss bugs with Infinity, strings, etc.semanticPASS
CLAMP-15package.json must declare type:module for .js files to use import/export ES module syntax under Node.js; without it node --test will throw ERR_REQUIRE_ESM or SyntaxErrordeterministic · file_containsPASS
CLAMP-16The plan explicitly forbids dependencies; package.json must not introduce any dependency blocksdeterministic · file_not_containsPASS
CLAMP-17clamp.js must not import any external packages — only local relative imports are allowed, ensuring zero dependenciesdeterministic · file_not_containsPASS
CLAMP-18The plan requires the module to be readable to a newcomer with self-documenting namessemanticPASS

cycle 0: 17/18 — CLAMP-16 failed → a fix step → cycle 1: 18/18. The checks themselves cost $0.52, apart from the work.

Acceptance checks

1 check recorded, skipped: framework_detection.

Steps

  1. Implement clamp.js and clamp.test.jsclaude-sonnet-4-6 · 11s · $0.27
  2. Fix: Contract item CLAMP-16 (file_not_contains) failed evaluation.claude-sonnet-4-6 · 12s · $0.12

What changed

The patch lives in the person's repository — this server keeps the contract, the checks and the cost, not the diff. Files the contract looked at: clamp.js, clamp.test.js, package.json.

Read this run's full public report

Give Dag a task. He works the night shift.

What counts as proof here

Most tools report that checks passed. We report what the checks were capable of seeing. Syntactic assertions confirm code is present; semantic checks read the change and judge whether it does what the task meant. That distinction is on this page because it is the honest one — a bare pass count can be inflated by shallow checks.

Every task runs the same gauntlet: decompose into a dependency graph, execute in isolated worktrees, verify and heal each subtask, compile a machine-checkable acceptance contract, run end-to-end acceptance, then a final whole-task verification with a minimum-score gate. Red never lands.