The Failure That Looks Like Success
Table of Contents
A draft for the proposal shelf. Not submitted anywhere. Status: Keep.
1. Pitch
A check that has never failed has told you nothing. Run every gate against known-bad input before you trust its green.
2. Abstract
The failure worth testing for is the one that exits zero. One failure across 710 tests on a single seed, never reproduced, and nearly dismissed as flake. A drift sentinel going red on 1 of 69 tests because a vendored contract had skewed from its canonical copy. A pipeline that, given nothing to work on, wrote a plausible output file. The habits that came out of them: calibrate each gate in both directions, pin the shrunk counterexample when it fires, and return a reason rather than a boolean.
3. Audience, format, venue
- Audience
- Anyone who maintains CI, test suites or data pipelines. Needs no shared context.
- Format
- Lightning or 20 minutes. Both reviewers say do not stretch it to 45.
- Where
- ELS, Babashka Conf and Dutch Clojure Days lightning slots; MinuteCon 20-minute slot; a SPLASH workshop on agent reliability.
4. Review
The strongest draft on both reviews, and the presentation review's first pick: it fits any room and nothing on the Clojure/conj 2026 program competes with it. The L7 review would fold Visibility Is Verification into it as framing. The question a committee will ask: "test your tests" is mutation testing, and RacketCon 2024 had a talk on exactly that. The answer is the incident catalogue and the agent setting. Demo: break a tracked file, watch the lint stay green, fix the gate, watch it go red.
5. Backing
Public repositories:
Notes on this site: