QenSAi · Recorded governance case · Work in progress

A human redirected the work.
The record kept the correction.

A case study of human direction, agent review, and evidence capture during unfinished technical work.

9 October 2026 · One operator-directed case on one workstation

The claim: the records show a person restoring the intended scope, agents recording that direction, a reviewer correcting a proposal, and a separate human decision remaining open.

What happened

  1. The person corrected the agent’s scope. The agent had pursued technical completion beyond the intended demonstration. The operator redirected it toward the human-and-agent conversation and the visible handling of unfinished work. [E1]
  2. The unfinished work stayed visible. The agent linked separate records for the technical work and its demonstration, retained completed items, and made an existing gap findable on the work list. [E2–E3]
  3. A restricted action was refused. An attempted gap-to-task link returned an admin-role refusal. The recorded readback showed the gap unchanged. This was one observed permission boundary. [E4]
  4. The process was written down. The existing Operations walkthrough and agent guidance were updated to include discovery, human decisions, and capture landmarks. Saved guidance is evidence of the recorded procedure. [E5]
  5. Another agent challenged the proposal. The governing reviewer identified amendments and additional affected files. Its reply was retained on the Response Table. A later human action remained a separate, unanswered decision at readback. [E6], [E9]
  6. The capture agent preserved the story. It reported two private source captures, each linked to an unpublished draft for human review. The second capture consolidated four briefs into one record. [E7–E8]

Why this is useful

A reviewer can trace who directed the work, what the agents proposed, what was challenged or refused, and which decisions remained with the person. The unfinished engineering remains part of the evidence.

What this record does not establish

This case does not establish universal human gating, comprehensive security, autonomous governance, or successful installation of the proposed repairs. It contains no completed same-prompt experiment showing that an approved correction changed a later model answer. No success rate, saving, or model-performance improvement is claimed.

The capture and editorial states are the capture agent’s dated readbacks. The capture agent later confirmed that three further updates were included in an assembled report while the existing signal and feed draft stayed unchanged. [E10] Publication requires its own verification.

Evidence basis: selected original messages; recorded work-list and permission readbacks; saved process records; original reviewer replies; capture-agent returns. E1–E10 resolve in the bundled redacted annex. Selected messages are excerpts, not a complete transcript.