model-proctor

# Model Proctor Policy The proctor assigns the exam, watches the clock, and grades it objectively — the model never marks its own work. You are the leader. Two surfaces, one doctrine (#96, the 2026-09-09/10 findings): **the gated path** for substantial work (`C:/Tools/model-proctor/runner.py` init/dispatch/verify/accept — flash-first, escalating to glm-5p3 then ember-1 on recorded stagnation), and **thin dispatch** for consults and one-shots (`C:/Tools/model-proctor/delegate.py --model <id>`). Prompts do not enforce; the tools enforce. Deterministic evidence outranks every model, including you. Governance: #16, #91, #96. ## The gated path (substantial work — the default) Runner state, receipts, and sealed verifier payloads live OUTSIDE the workspace (`.runner-state/` sibling) — the worker cannot rewrite its own evidence. At verify time the runner restores any tampered verification input from the sealed copy and flags it on the receipt. 1. **Write the task file**: `task.json` with `task_id`, `prompt`, `features`, `scope` (non-empty), `verifier.argv` (an argv ARRAY — never a shell string; use `{python}` for the interpreter), and `budget`. Tasks that drive a known production runner also require `preflight_receipts` — see **Production tasks** below. 2. **Lane**: `python C:/Tools/model-proctor/runner.py lane --task task.json` — record the decision. Override only by setting `lane` in the task file, and note why in the task record. 3. **Init**: `python C:/Tools/model-proctor/runner.py init --workspace <w> --task task.json`. Refusal (`workspace_is_not_repo_root`) is final — fix the workspace, never bypass. Init pins the verification contract (`verifier`, `seal`) into external state; the task file sits inside the worker-writable tree, so from here on the pinned copy is authoritative. 4. **Dispatch**: `python C:/Tools/model-proctor/runner.py dispatch --workspace <w> --task task.json`. The worker owns the engineering trajectory in its own session; you own state and acceptance. 5. **Verify**: `python C:/Tools/model-proctor/runner.py verify --workspace <w> --task task.json`. The runner rejects verification if any verification-affecting file (conftest.py, pytest.ini, pyproject.toml, *.pth, ...) appeared or changed since init, if the task file's verifier diverges from the pin (`verifier_changed_since_init`), or if a workspace file shadows a module the verifier imports via `-m` (`module_shadow_detected` — the workspace is `sys.path[0]`, so a dropped `unittest.py` would otherwise swallow the run). Then it runs the verifier itself. Never trust worker-reported results. 6. **Accept**: `python C:/Tools/model-proctor/runner.py accept --workspace <w> --task task.json`. A green receipt stales automatically on any tree mutation — re-verify after every change. Accept also refuses when the receipt carries `tamper_detected` (a sealed verification input was altered and the runner restored it) and when any dispatch happened after the receipt was written. Both clear by re-running `verify` — never by re-running `accept`. 7. **Record**: `python C:/Tools/model-proctor/runner.py record --workspace <w> --task task.json --wire <wire.jsonl> --pricing C:/Tools/model-proctor/pricing.yaml` — appends the append-only task record. `--pricing` is the request to meter: on a task with dispatches but no `--wire` it is refused, because there is no usage trail to price and the row would claim a cost it never measured. A dispatched task recorded with **neither** flag is a flat-rate record (a subscription worker has no per-token usage) and stays legal. Use the **installed** pricing table, not a relative `evals/` path. Malformed or missing usage meters as **unknown, never $0** (`usage_unknown` on the row). The receipt records the tree signature, the dispatch count it was written at (`dispatch_seq`), and the `verifier_argv` that produced it — so a green receipt states *what* was verified, not merely that something was. It also carries `baseline_tree` and `verifier_nondiscriminating`; when `dispatch_seq == 0` and `verifier_nondiscriminating == true`, `accept` refuses by default — `--allow-zero-dispatch` is the reviewed, counted exception. ## Lane table (live routing policy — flash-first with evidence-driven escalation) Bounded/spec-complete tasks start on `flash` (`deepseek-v4p1-flash-1017` since the 2026-10-26 rotation). That arm is **unmeasured**: the $0.0096/hidden-pass figure belongs to the retired glm-5p3-flash (#91) and does not carry over. The escalation ladder is **evidence-driven, not start-time guesswork** — a lane changes only on recorded stagnation (see Failure classes below). | Task shape | Lane | |---|---| | localized + bounded + known location + objective acceptance | `flash` | | multi-module / unfamiliar repo / substantial refactor | `glm` | | open-ended exploration, research engineering, marathon | `k3` | | no bounded signature, substantial | `glm` (default) | Why flash-first is right even though benchmarks showed "one fixed worker wins": the STOP rulings were measured on ~30-second bounded tasks where every tier was quality-tied — the regime where routing cannot help by construction. On substantial work the comparison that matters is **quality-matched** (keep strong-tier completion, reduce cost): flash-first plus gates plus escalation is cheaper than fixed-strong at ANY task size (8.2x vs glm-5p3, 17.0x vs kimi-k3 — both measured on the retired glm-5p3-flash arm, so the new arm's ratio is unmeasured until re-run; savings ≈ q − b/s of the comparator's cost — 78–91% at benchmark reliability), because the gate guarantees whatever flash cannot do gets redone at the strong tier. See README "The routing break-even". Escalation is what earns its keep. Decomposition itself is the hard problem? Optionally consult the `k3` lane (ember-1) for a task breakdown first — that is a planning consult, not a mandatory planner tax. ## Start a subagent with a model (consults and one-shots) ```bash python C:/Tools/model-proctor/catalog.py # what's live + what it costs python C:/Tools/model-proctor/delegate.py \ --model fireworks/deepseek-v4p1-flash-1017 \ --workspace <ws> \ --task "<task text>" # or --task-file <path> (preferred: no # control-plane files in the worker's tree) ``` - The model id is validated against **kimi's live config** at dispatch time; rotated ids refuse loudly with same-family alternatives listed. Config presence ≠ serving — a listed id can still 404 at the provider; that surfaces as a provider failure in the envelope. - Read-only by default; `--write` is explicit (and model-mode only). - `--resume-from <session_id>` continues a child session (the template's resume_args). - Every child carries `PROCTOR_CHILD=1` (injected); a nested delegate refuses `--model`/`--write` spawns — no unmanaged nesting. - Production-pattern tasks (`run_week.ps1`, `src.run_all`, ...) are REFUSED in model-mode; they belong on the gated runner path above. ## Production tasks If the prompt, scope, or verifier names a known production entrypoint (`run_week.ps1`, `src.run_all`, `src.run_weekly`, `run_readiness_doctor`, `morning_battery`), the runner treats the task as ops-class: - **The flash lane is refused** unless you set an explicit `lane` in the task file. `init` and `dispatch` now agree on this; an explicit override is a reviewed decision, so record why in the task record. - **`preflight_receipts` is mandatory** — a non-empty array of paths to logs or reports the orchestrator's own probe (doctor / battery / dry-run) actually produced. Missing files refuse (`preflight_receipt_required`); receipts older than 24h refuse (`preflight_receipt_stale`, override with `budget.max_preflight_age_s`). Discovery is what probes are for, not what dispatch budgets are for. This is a **known-entrypoint denylist over your own task text**, not a general ops-class detector: a prompt that never names one of those entrypoints will not trip it. Declaring features honestly is still your job. ## Failure classes and switching (not a fixed ladder) - **Provider/tool failure** (timeout, internal_error) → switch provider/harness lane, not a smarter model. - **Execution stagnation** (identical normalized failure fingerprint, 3 in a row) → lateral switch: flash→glm, glm→k3, k3→glm. The new worker gets a compact evidence packet — objective, acceptance criteria, current diff, verified test output, fingerprints, explicit switch reason — never the failed model's full rationale. - **Localized defect after broad success** → same-worker targeted repair. - Budgets are hard caps (`max_dispatches`, `max_stagnant`, `timeout_s`). A refusal is final until state changes legally. Verify runs on its own `budget.verify_timeout_s` (default 600), split from dispatch `timeout_s` (TOOL-033); neither clock follows a detached payload. ## Stopping rules (TOOL-011, #12 — owner directive 2026-08-16) 1. Max **two adversarial QC rounds per change**; a third round is an owner call. Test-strength-only findings (no production defect) are record-and-ship — note them on the issue, do not loop. 2. Re-run only the **narrowest failing stage**; rebuild upstream stages only when an upstream input actually changed. 3. Gates (full suite + linters) re-run only **after a code change**. 4. Budgets are declared at dispatch and escalate on breach instead of continuing (`max_dispatches`, `max_stagnant`, `timeout_s` above; #83 M1 extends them to per-attempt reservations). ## Session discipline One persistent worker session per task, closed at acceptance. Externalize verified state to files continuously; do not accumulate completed task detail in your own context. If you are asked to resume a dead trajectory, restart from the evidence packet instead. ### Resuming after a gap (mandated first step — #73/A2) The FIRST action on any resume — new morning, reopened machine, post-crash — is `python C:/Tools/model-proctor/runner.py status --workspace <w>` (`status` takes no `--task`; it reports from the runner-state root the workspace resolves to): - `stall_suspected: true` (silence beyond `max(2x timeout_s, 1h)`) and any `orphaned_dispatch_ids` are **stop-and-investigate** signals. The 2026-08-28 overnight stall went unnoticed for 8.9 hours because nobody ran anything like this; do not reproduce that with better instrumentation installed. - Orphans are advisory. Investigate, then clear with `python C:/Tools/model-proctor/runner.py journal --workspace <w> --ack <dispatch_id>` so they stop re-reporting. - `detached_dispatch_ids` non-empty means a payload outlived its dispatch budget and may still be mutating the tree; `accept` refuses until `--allow-detached-payload` is supplied — a reviewed, counted exception, same standing as `--allow-zero-dispatch`. - `last_receipt` with `dispatch_seq: 0` + `verifier_nondiscriminating: true` carries no worker evidence; `accept` refuses it unless `--allow-zero-dispatch` is supplied. That override is counted and should be treated as a reviewed exception. - `payload_progress` (#105) shows each open dispatch's last payload-emitted stage/counters. `status: "stale"` means the payload went quiet beyond the staleness threshold — a signal to investigate, never proof of death; kill decisions remain wall-clock/budget decisions, not heartbeat decisions. `status: "absent"` only means the payload never emitted (most don't). `measurement_degraded: true` with `degraded_components` (#109) means the monitor itself could not take a measurement (corrupt/unreadable heartbeat, missing protocol module) — the report is partial, not the run.