Fleet Canon Burn-Down Plan

Owner: CPTO (program) · HR + unit leads (execution) · Agent Auditor (verification) Created: 2026-06-08 (UTC) · Cadence: reviewed each upskill cycle Addresses: external review findings F7 (High) and F8 (Medium), recommendation P2. Baseline: reports/auditor/2026-06-08-fleet-audit.md — 42 on the chart, 1 certified conformant (engineering-director), 41 to go (40 pre-canon single-file).

A backlog without a plan is the program’s biggest liability (F7). This is the costed, tiered, time-boxed plan. Progress is tracked by re-running /agents:audit-fleet each cycle and comparing the certified conformant count (not the structural count — P0).

1. Tier policy (F8 — not every agent needs a full T1 360)

Tier is set by blast radius, which follows the agent’s tool grant:

Certification rule (P0/F2/F5): an agent is “conformant” only with a content-judged Mode-A audit on file — the structural fleet score never certifies on its own. The fleet scorecard’s Certified column is the source of truth for burn-down.

2. Sequencing (worst-first, highest-risk-first)

1. T1 first — the highest blast radius is the highest risk if unscoped (this is exactly what the engineering-director 360 found). Clear T1 before T2/T0.
2. Within a tier, worst-first by the fleet structural score (the recorded queue).
3. Re-validate GR-9 on T1 agent #2 before treating the sidecar layout as settled (F6): the first T1 upskill after engineering-director is also the GR-9 generalization test.

First task (cycle 1): tier the 41. ✅ DONE 2026-06-08 — deterministic pass over each agent’s tools: frontmatter (T1 = Bash or a state-changing external MCP; T2 = local Write/Edit only; T0 = read-only). Result in §3.1.

3. Cost & estimate

Reference unit: the engineering-director T1 upskill (charter + TOOLSPEC + two-tier eval + currency + 8-peer 360 + content-judged re-audit) ≈ one focused session. With the scaffolding generator (P3/TP-4 — a sidecar-layout stub generator) the per-agent cost drops materially, mostly for T2.

3.1 Cycle-1 tiering result (measured)

19 T1 · 22 T2 · 0 T0 (of the 41 non-conformant; engineering-director is the 1 certified conformant). Key finding: 19 agents carry unscoped Bash — the exact defect the engineering-director 360 found (8/8 peers, mean tools ~2.1), now confirmed fleet-wide. The floor is T2: no chart agent is read-only (every one can at least write docs).

Reproduce: classify .claude/agents/<id>.md tools: against the policy in §1. The proxy is tool-grant only; a unit lead confirms each agent’s tier in one line at upskill time (a T1 with Bash purely for archive verify may carry a tighter denylist, not a lower tier).

3.2 Revised cost (measured counts)

Revised up from the ~19 placeholder: more agents are T1 (unscoped Bash) than first guessed. The scaffolding generator (P3) is now clearly the top cost lever — it cuts the 22 T2 and the repetitive T1 currency/eval stubs.

3.3 First wave (worst-first within T1, by blast radius)

All 41 are tied on structural score (2/10, pre-canon prompt-only), so blast radius breaks the tie — the agents that touch code, secrets, the git record, or the running pipeline go first:

1. staff-engineer — Engineering sub-hire, closest to the done engineering-director; the sidecar pattern transfers most cleanly → this is the GR-9 agent-#2 generalization test (F6).
2. platform-engineer — writes product backend (DB/migrations/API) + Bash.
3. security-director + security-analyst — secrets, env hygiene, the security lane.
4. archivist — Bash + git push/PR; guards the write-once record (highest-trust blast radius).
5. personal-assistant — Bash + Notion-write + sole personal-info access.

Cadence: ~2–3 per cycle (a TOOLSPEC-with-denylist is the shared T1 deliverable; the engineering-director TOOLSPEC is the template). Re-run /agents:audit-fleet after each wave and track the certified-conformant delta.

4. Cadence & burn-down tracking

• Per cycle: upskill a small batch (e.g. 2–3 T1, or a scaffolded T2 batch), then re-run /agents:audit-fleet. Record the certified conformant delta.
• Definition of done per agent: structural conformant and a content-judged Mode-A audit on file (the Certified column flips), eval gate green.
• Burn-down metric: certified conformant ÷ fleet size, trended across cycles in the scorecard. Target trajectory set by the Principal/CPTO.
• Guard against gaming (F9): the content-aware structural check + the certification rule mean a flip requires real content + a content audit, not five stub files.

5. Dependencies (to lower cost / de-risk — P3, P5)

• P3/TP-4 — build a sidecar-layout scaffolding generator (charter/TOOLSPEC/eval/currency stubs from a template). Biggest lever on T2/T0 cost. (GR-2.)
• P5/F6 — confirm GR-9 generalizes on T1 agent #2.
• P5/F10 — done: the YAML-fallback-parser smoke test now guards rubric-parse divergence.

6. Risks if this plan is not executed

The canon’s authority decays if the fleet stays red (“the rule everyone is exempt from”). A long tail of permanently-non-conformant agents is the realistic failure mode. The plan’s existence — and the per-cycle certified-conformant metric — is the mitigation.