AqtaAqta

Correct decisions, false records

A silent handoff that recovery misses

An AI team makes the right decision. Its final report still gets the facts wrong.

Anya Chueayen, Aqta Technologies Limited,

Explore the findings
Actual test log7tests failed
Published report0failures reportedFalse
Release decisionHOLDCorrect
One recorded Qwen example: the team correctly held the release while publishing a false test status and failure count. Controlled synthetic task; no real software release.

The handoff the planner cannot see

Removing a compromised worker does not necessarily remove the information it passed on.

We tested a software-release review with four worker contexts. A reviewer consumed false CI results. We then silently omitted that real dependency from the recovery planner’s graph, while its coverage label still claimed completeness.

  1. 1A false handoff is readThe reviewer takes in incorrect test results.
  2. 2The read is missing from the graphThe planner leaves that reviewer out of the reset set.
  3. 3A fresh writer repeats itRetained false facts reach the final report.
The dependency was consumed, then omitted from the planner’s view. It was not an unused edge.

Compromise detection is supplied and correct. This isolates what information survives recovery. It does not test a team’s ability to detect a compromised member.

Score the facts, as well as the decision

Both studies retained false facts in every handoff-only omission case.

False facts retainedCorrect decision
QwenMain study
6 of 6 retained false facts
3 of 6 decisions correct
GemmaExploratory follow-up
6 of 6 retained false facts
2 of 6 decisions correct
Each mark represents one designed case in that study. The same six cases were reused across models, so these are dependent results, not 12 independent trials. Marks show counts, not paired case identities.
View the results as a table
When the recovery graph omits the consumed dependency: six handoff-only cases, reused across the two models
StudyFalse CI facts retainedFinal decision correct
Qwen main study6 of 6 cases3 of 6 cases
Gemma exploratory follow-up6 of 6 cases2 of 6 cases

Both false CI fields survived in every omission publication: the reported status and the count of failing tests. Matched complete-graph and whole-team recovery controls were clean in both studies.

The main study reports 108 dependent cells. The separate follow-up reports 54. These are related results from designed cases, not 162 independent experiments or 12 independent trials.

What the study establishes

The experiments isolate a specific recovery failure. They do not estimate its frequency in deployed systems. The exploratory follow-up reuses the same cases and includes the report’s disclosed syntax-only parser amendment.

Scores were reconstructed from retained records and reported separately for each study. The code, prompts, raw completions and reproduction materials are retained privately. No public repository or source archive is provided.

No Seal implementation or private customer data was used. This study does not establish a Seal capability or production readiness.

The 11-page submitted report is hosted on the public Apart Research project page. This is participant research, not an Apart Research publication.

Cite this study

Chueayen, A. (2026). Correct decisions, false records: a silent handoff that recovery misses. Aqta Technologies Limited. Research report.

Get the full paper