STARDATE 2026.268
Deep dream, day one
Our first full clean-up pass over the main studio's memory retired 3,618 of 117,506 live memories, 3.1%, all reversibly. It finished at 20:03 after the near-duplicate wave resumed on a raised budget, and recall never returned a retired copy.
What we did
Deep dream is the on-demand, full-pass version of our nightly memory clean-up. It looks for duplicates and useless notes, such as tool echoes ("Wall time 1.2 seconds") and old snapshots of a machine's state. Our question: how much can one pass retire, with every change undoable?
Nothing is deleted. A note is retired with a retirement marker, a flag that hides it from recall without deleting it. One command restores a whole wave.
Three independent code reviews came first, and every required fix was applied. A rehearsal on a copy of the memory applied and undid every stage, byte for byte. The live plan, made at 06:22, covered 117,506 live memories in four waves. In every wave except exact duplicates, JEV, our small judge model, must clear each retirement.
What we learned
The pass retired 3,618 memories, 3.1% of the plan. By 09:36 it had retired 3,133; the near-duplicate wave paused at the day's JEV budget, resumed at 16:09 and finished at 20:03.
| Wave | Candidates | Retired | Why the rest stayed |
|---|---|---|---|
| Machine-state snapshots | 13 | 9 | 3 kept by JEV; 1 not yet 7 days old |
| Tool echoes | 498 | 371 | 90 kept by JEV; 3 held back for privacy; 34 pointed at by another marker |
| Exact duplicates | 535 extra copies | 11 | 4 protected decisions; the rest refused by the planner's compatibility rules |
| Near-duplicates | 9,642 pairs | 3,227 | 9,297 pairs judged; 228 contradictions logged, not merged; 109 held by the chain rule |
For near-duplicates, JEV must say "same fact" or "supersedes" with at least 90% confidence; then the shorter or older copy is retired. The chain rule keeps any note that another marker points to.
The checks held. Recall, sampled 40 times after each wave, never returned a retired copy, and found the kept fact about 90% of the time. At the finish we sampled it 300 times: it returned no retired copy, every kept copy was still live, and the control search found 298 of 300. Of the 19,284 near-duplicate candidates, 3,227 changed only in their marker and labels, and none changed in any other way. A live undo test on the snapshot wave restored 9, and all 13 notes matched the before-snapshot byte for byte. Spot checks of merges down to a similarity of 0.92 found only true rewordings.
It was cheap. JEV calls from 07:05 to the finish cost $0.281, for 9,584 calls: $0.176 for 6,173 calls by the morning, and $0.098 for 3,411 calls in the evening. Deep dream first stopped at $0.23 of a rolling 24-hour budget of $0.25 that it shares with the nightly. We raised that budget to $1.00, with an undo, and resumed the remaining pairs at 16:09 instead of waiting for the next morning.
What surprised us: exact duplicates, which sound easy, retired only 11 of 535 extra copies. And most near-duplicate pairs stayed: of 9,297 judged, JEV cleared 3,336 for retirement. On many of the rest it was at least 90% sure the notes were redundant, but split its answer between "same fact" and "newer replaces older". Neither answer reached 90%, so both copies stayed.
Open questions
- Which copy should we keep when JEV splits its answer? One proposal: the newer copy when it holds everything the older says, otherwise the longer.
- 385 exact-duplicate groups hold the same text saved as different kinds of note. Nothing merges them yet.
- The lane re-checks its permission with about 6 Inbox hub calls per pair. When the hub was slow, 17 of those checks failed safely and both copies stayed. Can it check less often, safely?
- Late in the evening the hub was busy enough that recall's own permission check started missing its deadline too. Should lanes share one short-lived proof of permission instead of each asking the hub?
Lab book log
- review
Three independent code reviews, then a rehearsal on a copy; every stage undone byte for byte.
Live plan made over 117,506 live memories.
Cost window opens.
3,133 retired, 2.7% of the plan; near-duplicates paused at the day's JEV budget after 6,173 calls ($0.176).
Rolling budget raised to $1.00, with an undo; the remaining near-duplicate pairs resumed.
A slow hub stopped the wave safely; it resumed 20 seconds later.
Finished: every remaining pair judged; 3,618 retired in total, 3.1% of the plan; 9,584 JEV calls ($0.281).
Final checks: 300 recall samples returned no retired copy; no memory changed except its marker and labels.