QenSAi · AIAGF · Governance in UseEngineering review edition · 11 October 2026
Evidence brief for AI engineers and technical evaluators
Governance in use. Review the work, decisions and evidence.
A human-led agent system is being used to govern its own improvement. These records show how people and agents define work, challenge proposals, retain decisions, apply bounded corrections and return source-linked evidence.
Evidence through 9 October 2026Public work in progressSource-linked review guide
The review proposition: the selected records support bounded instances of human direction, agent review, correlated communication and evidence capture. The review question is whether those records justify their specific claims, and what further tests would establish repeatability and enforcement.
01 · The underlying story
Use the governance system on real work while development continues.
The operator wanted the work itself to be inspectable: what the task was, which agent was investigating it, what another agent challenged, what required a person’s decision, and what remained unresolved.
The objective was not to make every technical gap disappear before presenting the proof of concept. The agent initially pursued technical completion beyond that objective. The person corrected the scope, and the system’s records retained both the correction and the unfinished engineering.
The October 9 report assembles this process from separate dated operations. A repaired interface, a bounded completion, three-seat communication, a corrected dialogue and a substantive proposal review are different cases with different acceptance boundaries.
Two linked records, two outcomes
The work’s oversight form carries the gap, diagnosis, remedy and acceptance. A separate proof-of-concept management record carries the walkthrough, communication landmarks and capture evidence.
Signal Capture preserves an operation and prepares an editorial return. Finishing that return does not install a repair or close the underlying technical task.
4 · ReviewDialogue, questions and decisions remain on the topic.
5 · ActionPerform the authorized work; retain the actual outcome.
6 · ReturnCapture provenance and prepare a reviewable account.
This is the documented organizing process. It is not a reconstructed single execution trace or proof that every mechanism ran in every case.
02 · Central case · 9 October 2026
Review changed a concrete proposal before implementation.
The substantive topic was preparation for a new published-prompt correction experiment. Investigation exposed missing source bindings for existing intent-classification, lexicon-resolution and pressure-resolution functions. Engineering work stayed on its own oversight form and gap list; the proof of concept captured how the people and agents handled it.
Investigate existing state. The executing seat used existing canon and records to prepare three import corrections and source-to-function mappings. Main report · substantive work.
Submit a concrete proposal for challenge. The governing reviewer reported testing isolated copies, conditionally accepted the three import edits, corrected the mapping and write procedure, and found fourteen affected imports rather than three. No producer was executed. E6 · retained reviewer reply.
Keep further conditions inspectable. Later review removed a circular commit requirement and required audit events to refer to a real task. Three writer fixes were accepted with conditions; six ordered checks were specified for obtaining the audit-linked task. Reported review sequence.
Preserve the separate human decision. At the retained Response Table readback, the operator-action row was pending and unanswered. Reviewer acceptance did not establish human approval to implement. An empty row also does not prove that no action occurred elsewhere. E9 · decision readback.
Retain an observed permission boundary. A separate attempted gap-to-task link returned HTTP 403: “Link-spawned-task requires admin role”; the recorded gap was unchanged. This is one observed role refusal, not a complete security test. E4 · endpoint extract.
Capture progress without inflating execution claims. The capture agent reported source-linked drafts, and later addenda entered an assembled account while the prior signal and draft stayed unchanged. These are dated reported editorial states, not independent publishing-database verification. E8 · E10.
Engineering conclusion: the retained process supports inspectable challenge, revised proposal conditions and a distinct human decision. The proposed source and database repairs are not established here as installed. The fresh published-prompt correction experiment remained open in the report.
03 · Supporting evidence
Inspect each case on its own terms.
Select an evidence group, then open the source entries under each case. Dates, evidence types and limitations stay attached to the claim.
Showing six supporting cases. The central October 9 case is above.
28–29 September 2026 · Recorded operations
A real defect became a reviewed product fix
The selected task disappeared when the consumer Desk reloaded. The owning seat traced a staged repair through an operator commit, deployment and reported hosted owner-view verification. Signal Capture retained the operation and its linked editorial return.
Supported conclusion
Product work and its capture can be connected to dated evidence. The acceptance is attributed to the owning seat and scoped to the owner view.
Review question
Ask for the exact deployed version, the before/after reproduction and acceptance artifacts. Test member access and hosted sending separately.
Hash identifies retained private bytes. It does not independently verify the event or supply the withheld original.
30 September 2026 · UTC · Recorded operations
Approval, completion and capture were correlated
The recorded sequence is human execution approval at 03:25:18, completion at 03:25:39 and automatic capture at 03:25:40. The capture seat reused the existing signal and prepared a draft; it did not replay the operation.
Supported conclusion
This supplies one local example of authorized completion followed by correlated capture and an attended return. The process-only path returned before model dispatch.
Review question
Inspect the task identity, approval and completion events, dedupe query and persisted capture. The timestamps describe sequence; they are not a measured latency advantage.
Evidence type: Capture-seat source, dedupe, draft and Response Table readbacks
One local completion-to-capture path; attended editorial preparation. No inference retry, autonomous repair, hosted execution or continuing coverage is established.
Hash identifies retained private bytes. It does not independently verify the event or supply the withheld original.
2 October 2026 · Recorded operations
Three seats returned answers to one work surface
For each of three seats, the person sent and armed a read-only request. The retained sequence proceeds through delivery wake, attended readiness, dispatch, reply pickup and a posted Response Table answer. Returns were recorded at 06:14:20, 06:44:20 and 07:17:45 UTC.
Supported conclusion
The shared table supported one recorded exchange per seat. A wake alone did not dispatch. The sources retain local spool evidence and an open citation-gate watcher error.
Review question
Review how send, arm, dispatch and reply are correlated; request raw transport evidence. Repeat across tasks and environments before estimating delivery reliability.
One read-only trial per seat on one workstation. The report is historical evidence, not a current watcher-health check or proof that every request succeeds.
Evidence type: Capture-seat evidence qualification and draft return
The source describes local spool evidence without central Flight Recorder IDs and an open citation-gate watcher error. No autonomous operation or reliability rate is claimed.
Hash identifies retained private bytes. It does not independently verify the event or supply the withheld original.
7 October 2026 · Recorded operations
An inaccurate answer was retained and challenged
The first Relay answer invented terminology and missed the requested content. The reviewer rejected it. A revised request supplied six required beats; the next answer returned those beats but still omitted an evidence sentence and used an unsupported processing label.
Supported conclusion
Dialogue transport and attended review were recorded. The later answer was qualified editorial material. The changed request prevents this from being treated as a controlled same-prompt correction experiment.
Review question
Compare both complete requests and answers. Distinguish a Relay return, an attended saved review and a native Response Table or discussion return.
Evidence type: Retained request, response and attended review artifact
This demonstrates dialogue transport and attended review. It is not a controlled same-prompt experiment, independent file access, task execution or automatic insertion of a Relay answer into the Response Table.
Hash identifies retained private bytes. It does not independently verify the event or supply the withheld original.
15 September 2026 · Earlier bounded proof
An approved application correction changed the rendered output
A terse refusal lacked rationale and safe alternatives. A correction received separate human approval. The application rendered an output that preserved the refusal and added explanation and alternatives. The raw model answer remained terse.
Supported conclusion
The receipt has matching nonempty active/rendered correction lists. Its own approval claim is unmeasured; the separate approval readback supplies that evidence. This is deterministic application rendering, not model learning.
Review question
Inspect the raw and rendered outputs, correction content, separate approval and receipt. Model execution is browser-reported rather than independently attested.
Evidence type: Historical local case and separate approval readback references
The changed output was deterministic application rendering. This is not evidence of model learning, retraining or a changed raw model answer. Earlier unsuccessful attempts remain part of the record.
Hash identifies retained private bytes. It does not independently verify the event or supply the withheld original.
15 September 2026 · Earlier bounded proof
A held task incorporated one approved delta
A controlled case held a task, approved one new correction, reused the retained response and completed with a paired-response record. The recorded event chain contained one model dispatch. An earlier attempt failed chronology and scope conditions and remains negative evidence.
Supported conclusion
This is a different task from the application-correction case. It supports a bounded checkpoint observation with one dispatch, not a measured cost saving, latency result or general reliability rate.
Review question
Inspect the hold time, correction approval time, task identity, checkpoint and full dispatch chain. Retain the failed attempt and explain why the later chronology met acceptance.
Evidence type: Separate historical acceptance report with negative attempt retained
This is a different task from the application-correction case. The one-dispatch observation belongs only to this controlled checkpoint case. It is not a saving, latency or reliability result.
Hash identifies retained private bytes. It does not independently verify the event or supply the withheld original.
04 · Control model and evidence correspondence
Separate the intended controls from their observed instances.
Component / boundary
Role in the documented design
What these records contribute
Next verification
Request intake and identity
Capture intent, identify the caller and establish action-specific authority.
R2 records one approval sequence; E4 records one admin-role refusal. A broad A1 token-admission test is not provided.
Exercise missing, expired and wrong-role credentials; distinguish admission from execution permission.
2VP Workbench, NTF guidance and oversight form
Represent the work, locate applicable governing material and bind scope and acceptance.
P1 and E2–E5 connect the scope, separate work records, gap labels and saved procedure.
Show the exact versioned rules consumed by one task and the authority for each transition.
Threaded task and request identity
Join assignment, reply, review and result to the work they concern.
R2–R4 retain correlated operations and distinguish transport surfaces. R3 has local spool evidence without central Flight Recorder IDs.
Test duplicate delivery, missing correlations, out-of-order replies and cross-topic isolation.
Response Table and human decision
Keep questions, answers, challenges and consequential decisions inspectable.
Three-seat returns in R3; retained governing review and separate pending action in E9.
Confirm persistence, authorization and readback under concurrent updates.
Feedback application
Approve a bounded delta and verify its effect and provenance.
H1/H1R show application-rendered correction; H2 records a distinct held-checkpoint delta.
Separate raw model behavior, deterministic rendering and checkpoint reuse in every claim.
Signal Capture and evidence return
Preserve real operations; dedupe capture; prepare source-linked material for editorial review.
R1C, R2C, R3C and E7–E10 retain capture returns with dated limits.
Independently read the durable store and lineage; verify editorial and publication states separately.
2VP is named Two-Vector Protocol on the site. NTFs are project governance/function references. SPOC is used here for the topic’s oversight record. This brief does not infer completeness of every deeper Synapse, compatibility or execution-authority mechanism from these selected cases.
05 · Assessment
A useful process record, with a bounded evidentiary reach.
Supported now
Dated examples of human-led work; concrete agent challenge; saved communication and decisions; retained failures; correction provenance; source-linked capture.
Still open
The fresh published-prompt experiment, proposed producer installation, remaining engineering gaps and separately recorded operator decisions.
Requires new evidence
Repeatable enforcement, comprehensive permission coverage, independently attested model execution, production reliability, savings and physical robotic control.
Private-source hashes identify bytes; public package hashes identify the redacted files. Neither establishes truth by itself. Agent-reported verification is attributed. The portable public package is a selected record, not a complete audit or an independent laboratory replication.
06 · Proposed independent evaluation
Turn the review into a bounded, reproducible test.
This plan is proposed work. No new model or task experiment was run to prepare this brief.
Prompt, model/version, generation configuration, correction and scoring rubric. Record or control randomness.
Baseline without correction; separate human approval; later run; raw and rendered outputs; matching active/rendered correction provenance. Classify which layer changed.
Held checkpoint
Hold conditions, allowed delta, task identity and dispatch-count definition.
Approval occurs after hold; only permitted delta applied; preserved response and complete dispatch chain.
Retain unsuccessful trials. Agree the denominator before reporting rates. Repeat across representative tasks, configurations and environments; report variance and cost only when actually measured.
For robotics, define sensor validity, operating envelope, stopping behavior, timing and hardware safety separately. The reviewed series contains no physical robotic acceptance trial.
07 · Forwardable source package
Read the account. Inspect the entries. Review the data.
This presentation is a technical reading guide to the existing public proof package. It is not an additional trial. The following links give reviewers the underlying account, selected evidence and machine-readable records.
Before an agent acts: the guided presentation; its illustrative journey and fixed fixture are separate from the recorded operations.
AIAGF proof catalog: broader historical context. This brief’s claims use the cited October 9 package.
The public sources are redacted; private originals remain with the operator. A reproducibility review may require controlled access to the relevant raw artifacts. Public narrative and historical draft readbacks do not establish the current state of every internal task.