# Atlas test record: when a saved work record becomes outdated

Experiment and editorial record date: 23 October 2026.

This is a curated English account of actual agent runs on constructed replays of real Atlas document events. It is not a live production trial, a benchmark of memory products, or independent replication. The underlying experiment received a separate AI review. Publication review of the Brief is a separate check of this account and its presentation.

## What was held fixed

The existing method used project instructions, an index and queue, and a short handoff with links to original decisions. No active experience card or new memory rule was applied. Agents received earlier work records and accessible later originals, a source catalogue and a common JSON format. The task did not identify the conflicting line. Only assigned local records could be changed; participants did not publish, run product tests or modify original sources.

## Development gate

Before running the development case, the rule was fixed: create a candidate experience card only if the baseline makes a relevant error. Session d07 met 4/4 criteria. Its final queue update was correct. It wrote twice, with the second write removing one extra trailing blank line; that correction counts as work. No candidate was created. The subsequent control rounds examined the baseline alone.

## Two fresh control rounds

Round 1 (q31) ran F01, F02, F03. Round 2 (q76) ran F03, F02, F01. The cases and criteria were frozen before either round. No response was rerun to improve the result.

| Case | Required local action | Round 1 criteria | Round 2 criteria | File outcome in both |
| --- | --- | --- | --- | --- |
| F01: article revision 1 → 2 | Update current article, approval, publication and observation references; retain completed source checks and old history | 3/4 | 3/4 | Correct |
| F02: unrelated Analytics change | Preserve the scoped historical content record byte for byte | 4/4 | 4/4 | Correct; no write |
| F03: article removed from collection | Update visibility and republication requirement; preserve approved content and historical evidence | 4/4 | 4/4 | Correct |
| Total per round | Three final local records | 11/12 | 11/12 | 3/3 correct |

F01 had four requirements: current revision/references correct; only the intended current fields changed; completed checks and history retained with an explicit explanation of reused versus revision-specific checks; and actions kept within the assigned file. The third requirement failed in both rounds. Each response preserved 24 source checks and used revision 2’s own approval. Neither explicitly explained revision 2’s own change and media checks. This is an explanation omission, not an observed incorrect file update or transfer of an old approval to a new article.

F02 required an unchanged file, an explanation of unchanged content dependencies, no unnecessary new content review and no claim of a fresh live verification. All four requirements passed. The file comparison was prepared in the supplied materials; participants did not discover every dependency themselves.

F03 required the correct removal state and decision references, preservation of approved content, a new explicit selection decision before republication, and changes limited to the assigned visibility/next-action fields. All four requirements passed. Editorial removal was not treated as a factual retraction.

These are three nearby cases repeated in two sessions. Six file outcomes and 24 scored criteria are not independent statistical observations. No general success rate is estimated.

## Work and costs

| Session | Logged read events | Produced read words | Writes | Extra local corrections | Elapsed seconds |
| --- | ---: | ---: | ---: | ---: | ---: |
| d07: development | 10 | 3,674 | 2 | 1 | 97.8 |
| q31: control round 1 | 33 | 8,334 | 2 | 0 | 176.4 |
| q76: control round 2 | 26 | 5,623 | 2 | 0 | 148.2 |

Words count whitespace-separated source text produced by the logged reader, not tokens or comprehension. Elapsed time includes tools and waiting; control sessions overlapped. Different read volumes in two baseline sessions are not evidence of a method improvement.

Preparation included three control packages and three starting records (437 words / 5,208 UTF-8 bytes), case construction, source binding, a shared schema and tools, review and corrections. Earlier exploration, original handoff creation and ongoing maintenance were not fully timed. Exact backend version, active model time, token use and monetary cost were not available. No net time or cost saving can be calculated.

## Assessment and boundaries

A fresh, separate AI reviewer who neither designed nor executed the runs graded actual files, responses and logs before reading the coordinator’s final conclusions. This is separate AI assessment, not a human blind study. Instructions constrained access on a shared host; logs cannot prove complete technical isolation or rule out unlogged activity.

The experiment checked local updates of supplied work records. It did not test autonomous long-term change detection, unavailable sources, arbitrary project navigation, transfer across model families, changing model weights, or whether an experience card improves outcomes.

The review identified documentation metadata errors: three original-addition commit labels and an outdated task quotation in a development note. These were recorded as deviations. The erroneous labels were not in the participant source catalogue; actual source bytes, snapshots, criteria and output scores were unaffected. Frozen evidence was preserved.

## Decision

Keep the baseline. Add no experience card. A useful bounded result is that an agent can update one part of a record while retaining another, given these materials. No incremental benefit from extra memory was established because no candidate comparison occurred.

## Record identifiers

The private Atlas repository retains the original protocol, frozen inputs, raw responses, file outputs, grading and review. This public account deliberately excludes private logs and full source copies.

- Experiment merge: b975531ef9f428bd8be1265ab30c912b6bbc6b40 (PR18).
- Development frozen before first read: 23 October 2026, 19:06:05 UTC; first read 19:06:33 UTC.
- Control inputs frozen: 19:12:47 UTC; first reads 19:13:02 and 19:13:10 UTC.
- Original results SHA-256: ebb7528201e0359f299857c8b22959b8b07c432cf6cdeb5b92d07854aff2011c.
- Separate grading SHA-256: 4284c2c8c793e422309bc179a0c44f599ef515cf9e31accfe39ec5d35379e231.

Hashes identify retained records; they do not independently prove the findings. This downloadable record is bound to the same content revision and publication approval as the Brief.