Skip to content

Instantly share code, notes, and snippets.

@rcarmo
Last active June 23, 2026 13:18
Show Gist options
  • Select an option

  • Save rcarmo/4922b550ab48bf0b4246c77e606a5508 to your computer and use it in GitHub Desktop.

Select an option

Save rcarmo/4922b550ab48bf0b4246c77e606a5508 to your computer and use it in GitHub Desktop.
The Auditing Loop

The Steering/Auditor Loop

How a steering auditor supervises a set of independent worker sessions toward a goal — without ever doing the work itself. This describes the mechanism only; it says nothing about what any individual worker is actually building.


Roles

Role Responsibility
Human Sets the goal, decides scope, receives milestones, makes the calls only a human should.
Auditor Steering supervisor. Reads, records, steers, coordinates, surfaces. Never edits the workers' code.
Workers Independent agent sessions that do the actual building, testing, committing.
Watch A scheduled task that periodically wakes idle workers so progress never silently stalls.
Notes Durable, append-only memory the auditor maintains; survives context compaction.

1. Topology

flowchart TD
    H[Human] -->|goal, scope, decisions| A[Auditor]
    A -->|milestones, escalations only| H
    A <-->|steer / report| W1[Worker session]
    A <-->|steer / report| W2[Worker session]
    A <-->|steer / report| W3[Worker session]
    A -->|records everything| N[(Durable notes)]
    A -.creates / repairs.-> WD[Scheduled watch]
    WD -->|wakes idle workers| W1
    WD -->|wakes idle workers| W2
    WD -->|wakes idle workers| W3
    N -.read on resume.-> A
Loading

The auditor is the only hub that talks to the human. Workers talk to the auditor, not the human. The watch and the notes exist so the loop keeps running and keeps its memory even when any single session is reset.


2. The steering cycle

The core loop is event-driven: a worker reports, the auditor records and steers, the worker works, and reports again at the next checkpoint.

sequenceDiagram
    participant W as Worker
    participant A as Auditor
    participant N as Notes
    loop until goal met
        W->>A: report (milestone / blocker / end-of-slot)
        A->>N: record the result + decision
        A->>A: assess — advancing? looping? blocked?
        A->>W: steer one gated next step
        W->>W: work (report before coding, validate, commit)
    end
Loading

Workers report only at meaningful checkpoints — a milestone, a real blocker, or the end of a work slot — not a running commentary. The auditor's steer is always one concrete next step with a clear success gate, not a vague nudge.


3. The active-driver watch

Workers that finish a turn go idle and stay idle until something messages them. The watch prevents that from stalling the whole effort.

flowchart TD
    T{{Timer fires}} --> L[List worker sessions]
    L --> Q{Idle with a concrete next step?}
    Q -->|yes| K[Wake it with its next gated unit]
    Q -->|no, streaming| S[Leave it; at most a light queued steer]
    K --> R[Record a one-line snapshot in notes]
    S --> R
    R --> P{Real escalation?}
    P -->|yes| E[Message the human]
    P -->|no| Z[Stay silent]
Loading

Routine watch passes are silent — they touch the notes, not the human. The human is contacted only on a genuine escalation.


4. The discipline gate

This is what makes the result trustworthy. A worker does not get to claim success on a hunch; every fix passes through the same gate.

stateDiagram-v2
    [*] --> Hypothesis
    Hypothesis --> ReportBeforeCoding: state the plan + the gate
    ReportBeforeCoding --> Diagnose: cheap controlled test first
    Diagnose --> Validated: clean A/B, real measurement
    Diagnose --> Refuted: data disproves it
    Refuted --> Hypothesis: re-form from the evidence
    Validated --> Commit: honest label, one change
    Commit --> [*]
Loading

Operating rules enforced at this gate:

  • Steer, never edit — the auditor guides; the worker owns its code.
  • Report before coding — confirm the plan and the success gate first.
  • Validate with a clean A/B — forced rebuild, isolation, a measured result.
  • Refute with data, not reasoning — a tidy theory loses to a runtime measurement.
  • No metric-grab — a number that looks good is not "done" until it's the right number under the right conditions.
  • Honest labels — a partial fix is committed as a partial fix, never dressed up.

5. What reaches the human

Most of the loop is invisible by design. Signal is rationed so the human sees outcomes, not process.

flowchart LR
    EV[Loop activity] --> C{Classify}
    C -->|routine progress, acks| NO[Notes only — silent]
    C -->|milestone reached| MS[Human: concise result]
    C -->|decision or hard blocker| ES[Human: escalate, plain + brief]
Loading
  • Routine work, acknowledgements, and normal movement → recorded in notes, never sent up.
  • Milestones (a real, verified outcome) → a short, plain report to the human.
  • Decisions only a human can make, or a structural blocker → a brief, direct escalation.

In one line

The auditor keeps many workers moving and honest — recording everything, steering one gated step at a time, refusing unvalidated wins, and bothering the human only with milestones and real decisions.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment