The Receipts
Our platform mines its own failures, writes its own fixes, ships them through gated deploys, measures the result, and reverts regressions. This page is the log.
Every entry below is a real production record — dates, ship times, and root causes as the system wrote them. No human approved these ships; a human audits them afterward.
30-day window ending Jul 13, 2026: 126 tickets — 28 self-healed code fixes, 30 questions answered, 17 resolved, 8 with verified outcomes; 7 needed a human.
Self-shipped improvements
Eight improvements the system identified, specified, and shipped itself — four of them in a single 92-minute window on the evening of July 5.
Repeat-ask guardrail
Detected that users sometimes ask for the same change twice because the previous build didn't actually apply it (zero diff) or the change wasn't visible in the preview. Shipped a gate that verifies a build changed what the user asked for before declaring it done.
Autofix retry-loop breaker
Found its own retry loops: fix attempts that didn't change the code kept being retried. Shipped a breaker that requires a root-cause diagnosis before another attempt is allowed.
Zero-diff rebuild gate
Noticed rebuilds that produce no delta (the change was already present, or the feedback wasn't actionable) waste compute and erode trust. Shipped an abort gate for zero-diff refines.
Thread-artifact grounding for email replies
Caught itself answering “this” and “it” on long email threads from the latest message only. Shipped grounding that resolves pronouns against the artifact the thread was originally about.
Product catalog lookup
Couldn't answer “which existing apps match this capability profile?” — so it specified the missing internal lookup, filed its own work order, and shipped it. Fourteen minutes from self-filed ticket to healed.
Test-run history lookup
Couldn't check whether a requested test had already been executed, or retrieve its results. Specified and shipped the query capability for its own test-run history.
Spreadsheet formula tracing
Was asked how a number in a proposal was derived from a linked Google Sheet and couldn't trace cell values and formulas. Shipped the tracing capability.
Attachment content analysis
Couldn't read a .docx attachment well enough to extract a calculation methodology. Shipped document-content extraction for attachments — 33 minutes from self-filed ticket to healed.
How the loop works
The system never edits code directly. It names failure classes, proposes fixes, and files work orders into the same gated pipeline every human change goes through.
Mine
Clusters 30 days of escaped failures — tickets a human had to touch, retries, repeat asks — into named failure classes. 39 classes mined so far.
Propose
For each recurring class, drafts a guardrail or missing capability, grounded in how the actual tickets were resolved — with an adversarial self-review and a confidence score.
Gate
Every ship goes through the same pipeline as human changes: compile checks, a gold test-suite floor, and smoke tests before traffic flips.
Ship
Files its own work order and implements through the standard pipeline. One ship in flight at a time. No human approval since July 4 — a human audits after the fact.
Measure
Each shipped guardrail gets a verified reward: recurrence of its failure class in equal windows before vs. after ship.
Revert
A guardrail that regresses its own metric is automatically reverted, and its failure class is marked human-only forever.
What keeps it honest
- Correctness-anchored circuit breaker: if resolution quality degrades week-over-week, autonomous shipping pauses itself and emails the founder.
- Every self-shipped change rides the same smoke-gated deploy pipeline as human commits — nothing skips the gates.
- Per-guardrail verified reward; regressions auto-revert.
- Security-sensitive surfaces (auth, tokens, access control, billing) are fenced off from autonomous changes.
- One-line kill switch, and a daily audit briefing to a human.
Don't take the log's word for it
Send us any back-office workflow. The same system builds you working software in 24 hours.
Intuist is raising a $3M seed. Investors: write to [email protected] or open in Gmail.