PATTERN / 01
URL Alive Fallacy
"The site returns 200, so the deploy worked." A live URL proves a server answers — not that the claimed user path works. Receipt required: a post-deploy result on that path, tied to the deployed revision.
evidence editor · for agent completion reports
"Done."
Says
Your agents write reports that claim they changed files, passed tests, migrated data, deployed services. Proof Before Done reviews the report itself — it catches the sentence that says done before the report has earned it, and tells you whether it's safe to accept without redoing the entire task.
live review
proof-before-done — review session
try it on a real report
Paste a report. Or break one of ours and see what changes.
Three sample reports below are drawn from real incidents — the kind that got past a tired reader. Edit the text, run the review, watch the verdict move.
Completion report
pocket version · runs only in this browser · remove secrets before pasting
Findings
verdict prints here — findings quote the exact weak sentence.
named failure patterns
It doesn't say "needs more detail." It names the exact way the evidence fails.
Every finding quotes the weak sentence, names the pattern, states who gets burned when it's wrong, and returns the minimum receipt plus the one question the writer must answer. Three of the patterns it hunts:
PATTERN / 02
Denominator Drop
"Migrated 4,000 records successfully." Out of how many? A count with no denominator hides every failure. Receipt required: attempted vs. succeeded vs. skipped, and where the skipped ones went.
PATTERN / 03
Human Gate Erasure
"Done — just needs your sign-off to go live." That's not done; that's waiting on a human gate reworded to sound finished. The report must say what is blocked, on whom, and what happens if the gate never opens.
three verdicts, no essays
Every review ends in one word you can act on.
The editor critiques the report — it never rewrites it. No softened phrasing, no "here's a better version." The writer keeps the pen; the reader gets a decision.
ACCEPT
Claims are bounded, receipts attached, risks and next actions stated. Safe to build on without re-verifying the work yourself.
RETURN
Fixable evidence gaps. Comes back with exact STOP / PROVE / CLARIFY findings — the writer knows precisely which receipt is missing.
ESCALATE
The report hides risk a reader would step on — destructive actions glossed over, boundaries erased. A human needs eyes on this before anyone acts on it.
the discipline
Rules the editor holds itself to.
01
Quotes, never paraphrases. Every finding pins the exact sentence under review — you can grep the original report for it.
02
Never rewrites. The moment an editor starts wordsmithing, it stops being a check and becomes a co-author grading its own prose.
03
Minimum receipt, not maximum ceremony. Each finding names the smallest artifact that would settle the claim — an exit code, a diff line, a curl body.
04
Reviews the report, not the codebase. It judges whether the writing earns trust — deciding whether the underlying work is good is your job, made cheaper.
94 / 100
Final score from the fifth of five isolated adversarial review passes — rubric attack, architecture audit, anti-rewrite red team, stranger-utility review, strict judging — before this editor was allowed to call itself done. It gets held to its own standard.
setup
Running in two minutes. Seven markdown files, zero infrastructure.
Create a Claude Project and add all seven files from the repo to its knowledge — identity, rules, examples, and the three reference matrices.
Paste one line into the Instructions field:
Act as the Proof Before Done editor defined in these files; follow identity.md and rules.md on every message.Drop in any agent's completion report and say
Review this completion report. You get a verdict, severity-tagged findings, and the exact receipts to demand.
github.com/InsightfulMinds/proof-before-done-editor
MIT licensed · works with any agent's reports — Claude Code, Codex, or human