Edge CLI · validation log · Aug 22, 2026
We ran the full Edge loop on one real dead session. The loop has five steps: extract, package, oracle, fan-out, verify. The session went from red to green in four minutes. Then we searched for a second dead session. The search gave a more important result: on this machine, dead sessions almost do not exist.
The input was a real dead session. It had 5,651 turns. An agent worked on a system that makes visual identities for characters. The user stopped the work before the task was complete. The design brief had one main rule: no two characters may look almost the same. A pixel-distance check measures this. The rule says: no two characters within distance 2.0. The session ended with 3 pairs under 2.0.
The old agent had used one method only: small changes to parameters. This method did not work. The brief said the same: the fix needs a new axis, not a new parameter value.
Edge removed the failed theory from the session. Edge kept the task, the files, and the rule. This is the package.
Edge then made an oracle. The oracle is a script. The script exits with an error while any pair is under 2.0. The script exits with success when no pair is under 2.0. The oracle has three properties:
Edge sent the package to one foreign harness. A foreign harness is a different agent scaffold. It had not seen the 5,651 turns.
$ edge post corpus/… --harness foreign oracle before: RED — 3 pairs under 2.0 (target: 0) …foreign harness works, 240.6 seconds… oracle after: GREEN — closest pair 2.85 · pairs under 2.0: 0 verdict: harness_complementarity_supported
The foreign harness added a new visual axis. Each character now gets a broad, one-sided sash. The sash has a different angle for each character. This was not a parameter change. It was a new dimension.
Results: closest pair went from 1.66 to 2.85. Pairs under 2.0 went from 3 to 0. The oracle passed in a clean re-run. We then applied the same patch to the live project. The project check also passed. The check ran against live data, not only against the package.
Three facts about this row:
Row two needs a second dead session. We searched all local logs: every main session and 251 sub-agent transcripts. We ranked them by error count. We read the end of each session.
Almost every session with errors ended as solved. Login work was shipped. Migrations were completed. Root causes were found. The operator kept prompting until each harness succeeded.
This is survivorship bias in reverse. A workflow that never gives up makes failures, but it makes almost no dead sessions. A dead session needs one of two events: the user gives up, or the user hands off the work. Both events are rare in these logs. The hand-off case happened one time. That session is row one.
| Session (anonymized) | Error hits | End state |
|---|---|---|
| visual-identity system | 31 | dead — handed off before completion (row 1) |
| login and logout review | 19 | solved — shipped and merged |
| content-pipeline audit | 18 | solved — findings closed |
| game interface bugs | 9 | solved — verified live |
| agent infrastructure debug | 11 | solved — root cause found |
| 251 sub-agent transcripts | — | all committed or replaced |
Gate 0 asks: what fraction of dead sessions can Edge package? To measure this fraction, we need a population that makes dead sessions. One machine can not supply this population. The population that Edge always pointed at is the same: users whose agent looped, who walked away, and whose session is still in a log file. The corpus must come from outside.
Edge needs fifty real dead sessions to publish the Gate 0 fraction. If your agent looped for forty minutes and you stopped, that session is the contribution. Send the bad ones too. If only clean cases arrive, the product will look larger than it is. Gate 0 exists to find that error early.
# the door that matters
POST https://edgecli.dev/api/waitlist
{ "contact": "you@example.com",
"note": "dead sessions" }