From 8dd311cbdf202cad0f592843b1cfa6261e7cd38c Mon Sep 17 00:00:00 2001 From: CryptoJym Date: Fri, 25 Sep 2026 20:08:20 -0600 Subject: [PATCH 1/2] Stardate: deep dream day one finished, with the final numbers The near-duplicate wave resumed at 16:09 and finished at 20:03. Final: 3,618 of 117,506 live memories retired (3.1%), all reversible; 300 recall samples returned no retired copy; no memory changed except its marker and labels; 9,584 JEV calls for $0.281. Adds an open question on sharing one short-lived permission proof across lanes. Co-Authored-By: Claude Opus 5.5 --- .../stardate/2026-268-deep-dream-day-one.html | 24 ++++++++--------- .../stardate/2026-268-incoming-filter-on.html | 2 +- .../stardate/2026-268-marker-chains.html | 2 +- .../stardate/2026-268-memory-reach.html | 2 +- site/assets/stardate/index.html | 10 +++---- site/index.html | 10 +++---- .../entries/2026-268-deep-dream-day-one.md | 26 +++++++++++-------- 7 files changed, 40 insertions(+), 36 deletions(-) diff --git a/site/assets/stardate/2026-268-deep-dream-day-one.html b/site/assets/stardate/2026-268-deep-dream-day-one.html index 9c6c91c..decd2a4 100644 --- a/site/assets/stardate/2026-268-deep-dream-day-one.html +++ b/site/assets/stardate/2026-268-deep-dream-day-one.html @@ -4,7 +4,7 @@ - + Stardate 2026.268: Deep dream, day one — BORG Collective @@ -25,11 +25,11 @@

STARDATE 2026.268

Deep dream, day one

-
Date
America/Denver
-
Status
Running
+
Date
America/Denver
+
Status
Finished
Format
Did · Learned · Open · Log
-

Our first full clean-up pass over the main studio's memory retired 3,133 of 117,506 live memories by 09:36, all reversibly. The near-duplicate wave paused at the day's JEV budget, and we resumed it the same afternoon.

+

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

@@ -39,23 +39,23 @@

What we d

What we learned

-

By 09:36, 3,133 memories were retired, 2.7% of the plan.

-
WaveCandidatesRetiredWhy the rest stayed
Machine-state snapshots1393 kept by JEV; 1 not yet 7 days old
Tool echoes49837190 kept by JEV; 3 held back for privacy; 34 pointed at by another marker
Exact duplicates535 extra copies114 protected decisions; the rest refused by the planner's compatibility rules
Near-duplicates9,642 pairs2,7426,010 pairs judged so far; 142 contradictions logged, not merged; 103 held by the chain rule
+

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.

+
WaveCandidatesRetiredWhy the rest stayed
Machine-state snapshots1393 kept by JEV; 1 not yet 7 days old
Tool echoes49837190 kept by JEV; 3 held back for privacy; 34 pointed at by another marker
Exact duplicates535 extra copies114 protected decisions; the rest refused by the planner's compatibility rules
Near-duplicates9,642 pairs3,2279,297 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. Every changed memory changed only in its marker and labels. 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 since 07:05 cost $0.176, for 6,173 calls. Deep dream was set to stop at $0.23 of a rolling 24-hour budget of $0.25 that it shares with the nightly. The near-duplicate wave stopped at the day's JEV budget. The remaining 3,632 pairs were first set for the next morning; instead we raised the rolling budget to $1.00, with an undo, and resumed them at 16:09.

-

What surprised us: exact duplicates, which sound easy, retired only 11 of 535 extra copies. And on most near-duplicate pairs JEV 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.

+

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, which caused 9 safe stops when the hub was slow. Can it check less often, safely?
+
  • 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

-
  1. review

    Three independent code reviews, then a rehearsal on a copy; every stage undone byte for byte.

  2. Live plan made over 117,506 live memories.

  3. Cost window opens; $0.176 over 6,173 JEV calls from here.

  4. 3,133 retired, 2.7% of the plan; near-duplicates paused at the day's JEV budget.

  5. Rolling budget raised to $1.00, with an undo; the remaining near-duplicate pairs resumed.

+
  1. review

    Three independent code reviews, then a rehearsal on a copy; every stage undone byte for byte.

  2. Live plan made over 117,506 live memories.

  3. Cost window opens.

  4. 3,133 retired, 2.7% of the plan; near-duplicates paused at the day's JEV budget after 6,173 calls ($0.176).

  5. Rolling budget raised to $1.00, with an undo; the remaining near-duplicate pairs resumed.

  6. A slow hub stopped the wave safely; it resumed 20 seconds later.

  7. Finished: every remaining pair judged; 3,618 retired in total, 3.1% of the plan; 9,584 JEV calls ($0.281).

  8. Final checks: 300 recall samples returned no retired copy; no memory changed except its marker and labels.

- +