Real work. Visible decisions. Inspectable outcomes.
A person and several agents use a shared work record to investigate, challenge proposals, decide what may happen next, and preserve the results.
Evidence through 9 October 2026 · Separate dated cases, assembled into one process account
What this POC demonstrates: recorded instances of human-led, multi-agent work moving through communication, review and evidence capture. The system’s unfinished work stays visible within that process.
The objective
Use the system to govern its own improvement.
The operator wanted to carry real work through the system while showing where the work was, what the agents were doing, why they were doing it, and what still required a human decision. The live feed and presentation preserve dated records of that activity as work progresses. Completing every engineering gap was never the admission price for showing the process.
The original presentation frames the value as “Before an agent acts: Orient it. Bound its authority. Inspect the outcome.” This report connects that intended journey to the operations already recorded. It separates observed behavior, agent-reported verification, illustrative design and unfinished work.
Set the purposeOperations and the oversight form establish the task, intended outcome and boundaries.
Find the governing materialUse the existing canon, indexes, function references and 2VP workbench before proposing a change.
Sequence and assignThe thread records the prescribed path: task scope → routing node → assembly → worker.
Exchange and reviewTAG Relay carries dialogue. The Response Table holds questions, answers, review and decisions.
Act within the decisionThe authorized action and its observed outcome are recorded. Unresolved gaps retain their own work items.
Capture and returnEvidence returns to the same topic. Signal Capture prepares a source-linked draft for human editorial review.
This is the organizing process. The cases below show which parts were actually exercised. They are separate operations, not a reconstructed single run.
Who does what
The human directs and decides. The governing reviewer challenges specifications and authority. The executing seat investigates and carries out authorized work. The capture seat preserves source material and prepares editorial drafts.
Two linked records
The work’s oversight form holds its gap, diagnosis, remedy and acceptance. A separate POC-management form holds the walkthrough, communication landmarks and capture evidence. Finishing a draft does not close the underlying technical task.
Work: keep the selected task when a consumer Desk reloads.
Record: the owning agent supplied the defect, staged repair, operator commit, deployment and hosted owner-view verification. Signal Capture retained the source and prepared a linked draft.
Outcome: a concrete product correction and its evidence were available to inspect.
The reported hosted acceptance covered the owner view. Member access and hosted sending were not established. Capture intake for this case was attended.
Work: complete one bounded, process-only continuation.
Record: human execution approval at 03:25:18; completion at 03:25:39; automatic capture at 03:25:40. The capture seat found the existing signal and prepared one draft, without replaying the operation.
Outcome: the completion, captured source and editorial return were correlated in the shared work record.
This record-only path returned before model dispatch. Automatic capture was followed by attended editorial work; it does not demonstrate autonomous repair.
The person asked. Three seats answered in the table.
Work: bring questions and agent replies into one visible work surface.
Record: for each of three seats, the human sent and armed a request. The recorded sequence proceeds through delivery wake, readiness, dispatch, reply pickup and a posted Response Table answer.
Outcome: one read-only exchange per seat was recorded, with posted returns at 06:14:20, 06:44:20 and 07:17:45 UTC. Signal Capture prepared the account “One desk instead of six windows.”
One trial per seat on one workstation. A wake alone did not dispatch. Local spool evidence lacked central Flight Recorder IDs, and a citation-gate watcher error remained open.
A bad answer stayed in the record. A correction followed.
Work: obtain guidance for presenting six supplied steps of the system journey.
Record: the first local Relay answer invented terminology and missed the requested content. The reviewer rejected it and retained it. A revised request supplied the required context; the next answer returned all six beats.
Outcome: dialogue transport and the subsequent attended review were recorded. The correction was usable as an editorial draft with qualifications.
The later answer still omitted an evidence sentence. The request changed, so this is not the same-prompt correction experiment. The review was saved by the attended seat; the Relay answer did not automatically become a Response Table answer.
Review changed the proposal before implementation.
The technical topic concerns a new published-prompt correction experiment. Its preparation exposed missing source bindings for existing intent-classification, lexicon-resolution and pressure-resolution functions. The work remained on its own oversight form and gap list; the POC captured how the people and agents handled it.
Investigate the existing design. The executing seat used the canon and existing records to prepare three import corrections and source-to-function mappings.
Submit the concrete proposal for review. The governing reviewer independently tested isolated copies, accepted the three import edits conditionally, and corrected the mapping and proposed write procedure.
Retain the challenge and its consequences. The reviewer found fourteen affected imports rather than three. The additional work remained explicitly tracked. Later review removed a circular commit requirement and required audit events to reference a real task.
Keep distinct decisions distinct. The reviewer accepted three writer fixes and specified six ordered checks for obtaining the audit-linked task. At the retained readback, the separate human action was still unanswered. This report does not convert a reviewer’s conditional approval into implementation approval.
Capture the progress. Signal Capture recorded discovery and review, prepared drafts, and later incorporated additional review steps in its assembled account. Those capture returns were retained on the POC Response Table.
Value demonstrated: another agent could inspect and challenge a proposal; the changes to that proposal and the remaining human decision stayed findable. The proposed source and database repairs are not presented here as installed.
Supporting incident: the operator corrected the agent’s scope
The agent had made technical completion the objective. The operator restored the intended demonstration: use the conversation and shared records to govern real work, with unfinished engineering tracked separately. That correction is one episode within the process.
Read the preserved October 9 case study. An observed admin-role refusal in that case is evidence of one permission boundary; it does not stand in for every governance control.
Separate September exhibits
Earlier proofs keep their own claims.
15 September · Application correction
An approved correction changed the governed output.
A model’s terse refusal lacked an explanation and safe alternatives. A held correction received separate human approval. The application then rendered a response that retained the refusal and added the required explanation and alternatives.
The retained receipt names the same correction as active and rendered. Its own approval field is unmeasured; a separate approval readback supplies that evidence.
Deterministic application rendering changed the output; the raw model answer remained terse. Model execution is browser-reported. This is not a model-training result or the new published-prompt experiment.
A held task incorporated one newly approved correction.
A separate controlled case retained a response during a hold. One new correction was approved, the checkpoint incorporated that delta, and the task completed with a paired-response record. The recorded chain showed one model dispatch.
An earlier attempt with the wrong chronology and scope remains negative evidence.
The one-dispatch observation belongs to this checkpoint case alone. It is not a measured cost saving, a general rate, or evidence about the application-correction case.
The original presentation also includes a fixed induction fixture and an illustrated authority boundary. They explain the intended journey and are labeled as such; they are not counted as additional live trials.
The records support bounded examples of real work, human authorization, agent-to-agent review, replies on a shared Response Table, and correlated capture. They also preserve an inaccurate answer, a failed historical attempt and unfinished decisions. That makes the process inspectable while development continues.
The architecture describes deeper checks involving system state, prescribed processes and execution authority. These selected cases do not certify all of that machinery. The communication trial, completion case, correction rendering and held checkpoint have separate identities and evidence.
Approved application correction and held-checkpoint continuation, separately identified.
Illustrative or specified
The original guided journey, fixture, general architecture and broader authority behavior.
Still open
The fresh published-prompt experiment, proposed producer installation, remaining gaps and operator editorial decisions.
There is no claimed success rate, saving, latency advantage, independent model-process attestation or universal human-gating result. Historical draft states are reported at their source readbacks; this assembly does not claim current publication status.
The invitation
Evaluate the process with a bounded task.
The original partner proposal invites an evaluator to select an agent or model configuration, non-sensitive task and agreed boundary conditions. Observe whether the agent finds applicable guidance, completes permitted work, stops when authority is missing, responds to changed scope, recognizes incomplete evidence and applies an approved correction with traceable provenance.
Those are proposed evaluation cases. The separate new published-prompt experiment remains open: freeze prompt/model/correction; run without the correction; obtain a genuine separate human approval; run again and inspect the changed answer plus matching active/rendered correction identifiers. A failed pair is retained as negative evidence.
The present report assembles evidence already available for review. Any new experiment, technical implementation or publication keeps its own authorization and acceptance.