audit

Principle Audit

One-shot audit of current context through your curated principles, verified by a stronger model (advisor). Unlike /grill (interactive, asks questions), /audit is non-interactive — it loads principles, calls advisor, and returns findings.

Usage

Invocation	Behavior
/audit	p0: auto-pick top 3 most relevant principles (single advisor() call)
/audit p1	audit through P1 only (single advisor() call)
/audit p0	same as bare /audit
/audit p1 p5	audit through P1 + P5 in one combined advisor() call (shorthand)
/audit --all	audit through ALL 5 principles via 5 SEPARATE Agent(subagent_type=general-purpose) calls (one per principle, fresh-context isolation per call)
/audit --team [<context-source>]	audit through ALL 5 principles via a subprocess-isolated agent team — TeamCreate + N×5 principle-auditor-p{1..5} teammates (default N=2) running a 3-round protocol (initial verdict → within-principle vote → cross-principle DM-debate signature). Variance dampened by within-principle consensus; verdict drift dampened by per-agent fresh context (each teammate re-reads its own agents/principle-auditor-pN.md definition + spawn prompt, no orchestrator interpretation in the loop).

/audit --all <context-source> accepts the same optional context-source argument as the other invocations. When omitted, each per-principle Agent receives the current conversation transcript as the audit context (the orchestrator embeds it in the per-call prompt; the Agent itself starts with a fresh context).

/audit --team <context-source> accepts the same optional context-source argument. When omitted, the lead embeds the current conversation transcript into each teammate's spawn prompt verbatim. The team primitive is more expensive than --all (10 teammates × ~3 rounds vs 5 Agent calls × 1 round), so --team is reserved for high-stakes audits where within-principle variance dampening matters — see ## When to pick below.

How It Works — single / shorthand modes (no --all)

1. Read principles from {vault}/thinking/principles/tracker.md
2. Pick principles: p0 auto-selects top 3 most relevant to current context, or use specified IDs
3. Print selected principles into conversation (so advisor sees them)
4. State: "Audit this conversation context through these principles. For each principle: does the current approach pass or fail? What specifically needs to change?"
5. Call advisor() ONCE — stronger model applies the selected principles in a single review
6. Format advisor findings for the user (single ## Audit Summary table)
7. Log principle usage to tracker: one row covering the run, increment Uses once per principle applied

/audit p1 and /audit p1 p5 are unchanged under this mode — they continue to work exactly as before. --all is additive, not a replacement.

How It Works — --all mode (5 isolated Agent calls)

When --all is passed, run principles sequentially as 5 separate Agent(subagent_type=general-purpose) calls — one call per principle, each spawned with a fresh, independent context window. NOT one call with all 5 principles in the prompt; NOT 5 advisor() calls. The per-call isolation is the load-bearing semantic; collapsing to a single batched call OR using advisor() (which forwards the full conversation transcript per call, leaking earlier verdicts) defeats the purpose of this mode (see ## Why per-principle isolation matters below).

Steps:

1. Read all 5 principle definitions (P1..P5) from {vault}/thinking/principles/tracker.md.
2. Resolve audit context: the named context-source arg if provided, otherwise the current conversation transcript.
3. Print all 5 principle definitions + the resolved context summary into the conversation so the user can see what is about to run.
4. For each principle in P1, P2, P3, P4, P5 — sequential, deterministic order, not parallel:
   1. Compose an Agent prompt that contains ONLY that single principle's text + the resolved audit context + the standard "does this pass or fail this principle, what specifically needs to change?" framing + an instruction to return its verdict as a single ^\| P{N} \| (PASS|FLAG|FAIL) \| <key finding> table row. Do NOT include the other four principles in the prompt. Do NOT include results from prior principle calls in the prompt.
   2. Call Agent(subagent_type="general-purpose", description="audit P{N}", prompt=<composed prompt>). The Agent runs with its own fresh context — none of the orchestrator's transcript is forwarded.
   3. Parse the returned summary for its row using the regex ^\| P[0-9]+ \| (PASS|FLAG|FAIL) \|. Capture the verdict (PASS/FLAG/FAIL) and the key-finding cell.
   4. Capture the full Agent response text for later use in the per-principle findings section.
5. Consolidate the 5 captured rows into ONE merged ## Audit Summary table — rows ordered P1 → P5 regardless of Agent return latency or other ordering signals. The orchestrator builds the table; individual Agents return only their single row + per-principle prose.
6. Append a ## Per-principle Findings section containing the full Agent response for each principle, separated by ### P{N} subheadings (also P1 → P5 order).
7. Append one tracker log row per principle (5 rows total) and increment the Uses count once per principle (5 increments). See ## Tracker Integration below.
8. Append the outcome-log row per ## Outcome log below.

Sequential (not parallel) ordering is intentional: it keeps debugging straightforward (failures attribute to a single principle's call) and the deterministic P1→P5 row order makes diffs across runs trivially comparable. Sequential is also fine for correctness: each Agent already gets fresh context, so there's no cross-contamination risk that parallel would prevent — it just makes the run linear.

Output consolidation contract

The --all mode output MUST follow this exact structure so downstream consumers (notably /slice's gate parser) keep working unchanged:

## Audit Summary
| Principle | Verdict | Key Finding |
|-----------|---------|-------------|
| P1 | PASS|FLAG|FAIL | one-line finding |
| P2 | PASS|FLAG|FAIL | one-line finding |
| P3 | PASS|FLAG|FAIL | one-line finding |
| P4 | PASS|FLAG|FAIL | one-line finding |
| P5 | PASS|FLAG|FAIL | one-line finding |

## Per-principle Findings

### P1
<full text from P1's Agent response>

### P2
<full text from P2's Agent response>

### P3
<full text from P3's Agent response>

### P4
<full text from P4's Agent response>

### P5
<full text from P5's Agent response>

Format contract (do not alter):

• The heading is exactly ## Audit Summary. Same heading as single-mode output.

• Verdict cell values are exactly PASS, FLAG, or FAIL — no synonyms, no decoration.

• Every verdict row matches the regex ^\| P[0-9]+ \| (PASS|FLAG|FAIL) \| so /slice's gate parser keeps working unchanged:

  echo "$AUDIT_OUTPUT" | grep -E '^\| P[0-9]+ \| (FAIL|FLAG) \|' && AUDIT_BLOCK=1
  

• Rows always appear in P1 → P5 order, regardless of per-principle Agent return latency. Consumers may rely on this ordering.

• ## Per-principle Findings is a sibling H2 to ## Audit Summary, with ### P{N} H3 subheadings (P1 → P5).

• Backward compat: single-principle invocation (/audit p1) and combined-shorthand invocation (/audit p1 p5) continue to emit the original single-advisor() output and the original ## Audit Summary table layout — --all is additive, not a replacement. Existing call sites that invoke /audit p1 or /audit p2 from /slice keep working unchanged.

How It Works — --team mode (subprocess-isolated agent-teams, ai-brain#252 Phase B)

When --team is passed, the lead spawns a TeamCreate-managed team of N×5 principle-auditor teammates (default N=2 → 10 teammates). Each teammate is a fresh-context Agent spawned via Agent(subagent_type="principle-auditor-pN", team_name="audit-...", name="pN-{a|b}", ...). Each teammate re-reads its own subagent definition (agents/principle-auditor-pN.md) — the orchestrator does NOT compose principle text into the spawn prompt, eliminating the run-to-run prompt-drift class of variance that --all showed during 2026-05-07 P2/P5 reproducibility runs.

Steps (lead-orchestrated):

1. Read all 5 principle definitions from {vault}/thinking/principles/tracker.md for the lead's own context (the teammates load from their agent definitions, not from the lead).
2. Resolve audit context: the named context-source arg if provided, otherwise the current conversation transcript.
3. Pre-spawn token-budget estimate. Subscription model: estimate per-teammate token use (≈ AUDIT_CONTEXT size + agent-definition size + per-round overhead × 3 rounds × N×5 teammates). If the estimate would consume more than ~30% of the current 5-hour session usage budget OR push the weekly limit past 80%, abort with a clear message naming the budget breach (do NOT silently proceed). Default N=2; if estimate exceeds budget, downgrade to N=1 and document in the run output.
4. Generate RUN_ID="audit-$(date +%s)". Create ~/.claude/audit-runs/${RUN_ID}/ for per-teammate JSON output.
5. TeamCreate({team_name: RUN_ID, description: "audit --team for <context>"}). The audit-* prefix is load-bearing — the TeammateIdle hook scopes its enforcement to this prefix.
6. For each principle P ∈ {1..5} and each within-principle slot S ∈ {a, b} (when N=2):
   • Spawn Agent(subagent_type="principle-auditor-p${P}", team_name=RUN_ID, name="p${P}-${S}", prompt=<spawn prompt below>).
7. Spawn prompt (embedded verbatim per teammate):
   • AUDIT_RUN_DIR = absolute path to the run dir.
   • AGENT_NAME = pN-a or pN-b.
   • WITHIN_PRINCIPLE_PEER = pN-b if this is pN-a (and N=2), or pN-a if this is pN-b, or null when N=1.
   • CROSS_PRINCIPLE_LEADS = the 4 names of the OTHER 4 principles' lead agents (convention: pN-a is the lead of principle N).
   • AUDIT_CONTEXT = full text of the artifact under audit, embedded verbatim. Do NOT summarise.
   • ROUND_BUDGET = 3 (default; configurable for this run).
   • PROTOCOL_PATH = absolute path to references/audit-team-protocol.md (single-source-of-truth for the 3-round protocol, JSON output schema, polling pattern, round budget, shutdown protocol, DM hygiene, and forbidden actions). Resolve from the active plugin root (e.g. ${CLAUDE_PLUGIN_ROOT}/references/audit-team-protocol.md). Each per-principle subagent reads PROTOCOL_PATH as its first action — the agent definition file (agents/principle-auditor-pN.md) carries ONLY the principle text + per-principle calibration, NOT the workflow protocol. This split eliminates 5x duplication and gives a single edit-point for cross-principle workflow changes.
8. Wait for all N×5 teammates to write their final round-3 JSON to ${AUDIT_RUN_DIR}/agent-${name}.json. Each teammate runs all 3 rounds within one agent turn (filesystem-polling between rounds); idle = round 3 done OR wait-loop timeout. The TeammateIdle hook (hooks/audit-team-consensus.sh) observes each idle event, logs progress to ${AUDIT_RUN_DIR}/hook.log, and escalates (osascript banner + Supaterm tab + escalation.md) when an agent idles with unresolved dissent (round<3, within-principle split, missing signatures, or dissent_acknowledged:true).
9. Aggregate to ${AUDIT_RUN_DIR}/consensus.json:
   • For each principle P, collect the within-principle agents' verdicts. If unanimous → that's the principle's verdict; record variance_within=converged. If split → use the majority verdict (or the first agent's verdict on a 1-1 tie when N=2), and record variance_within=split:P${P}.
   • Build the Audit Summary table rows P1 → P5 from the principle-level verdicts.
   • Build a Variance Notes section listing every variance_within=split and any P2 scope_flag that came back as ambiguous.
10. Output the lead synthesis (Audit Summary table + Per-principle Findings + Variance Notes) — see ## Output consolidation contract below; --team extends --all's contract additively.
11. TeamDelete() after all teammates have shut down (lead sends shutdown_request to each, awaits shutdown_response: approve=true).
12. Append outcome-log row per ## Outcome log below, including the variance= field.

Round protocol (run by each teammate within a single agent turn):

• Round 1: Initial verdict. Apply own calibration note (P2 scope, P3 scope-sufficiency, P5 severity rubric, P4 temporal ordering, P1 simplicity). Write own JSON.
• Round 2 (skip when N=1): Within-principle vote. Poll within-principle peer's round-1 JSON, compare YOUR round-1 verdict (and scope, where applicable) to PEER's round-1 verdict. Verdicts are LOCKED at round 1 — agents do NOT mutate verdict in round 2 or 3. This eliminates the within-principle stale-read race that fired in the 2026-05-07 dogfood run on the 2026-05-04 fixture (both p2-a and p2-b silently swapped verdicts mid-round-2 and each reported agreed: true against the other's pre-swap state). Round 2 records agreed: true|false based on frozen round-1 verdicts. DMs are for cross-principle context-sharing only; no waiting for response.
• Round 3: Cross-principle signature. Poll the 4 other principles' lead JSONs. Sign whether each finding is internally consistent with that lead's stated principle. On substantive conflict, DM the lead with one reconciliation question; wait up to 60s. Record 4 signatures.

Hard cap: 3 rounds. Past round 3 without consensus → teammate sets dissent_acknowledged: true and idles; the hook escalates to user.

Race-safety contract: two layers, both load-bearing.

1. Per-teammate file ownership. Each teammate writes ONLY to its own agent-${name}.json. The lead aggregates after all teammates idle. There is no shared-write consensus.json (the lead synthesises it post-aggregation). Multiple teammates writing to one file would race.
2. Round-1-locked verdicts. Each teammate's verdict (and scope_flag, where applicable) is FROZEN at round 1. Round 2 (within-principle vote) and round 3 (cross-principle signature) do NOT mutate verdict. The within_principle_vote.agreed field is computed strictly from round-1 verdicts of both within-principle peers. Without this lock, round 2 reconciliation can produce a stale-read race where both peers swap verdicts mid-round and each silently reports agreed: true against the other's pre-swap state (observed in the 2026-05-07 dogfood run; fixed by encoding the verdict-lock invariant into the agent definitions).

Shutdown protocol contract. Each principle-auditor-pN.md definition mandates that teammates respond to {"type": "shutdown_request"} SendMessage events with {"type": "shutdown_response", "approve": true} IMMEDIATELY. Without this, the lead's TeamDelete() waits indefinitely for shutdown approvals from idled teammates, hanging the entire /audit --team invocation (and any wrapping smoke-test subprocess). The agent definition is the single source of truth for the shutdown response — the lead does not need to enumerate teammates.

Subprocess-isolation contract: the principle text + calibration notes live in the agent definition file (agents/principle-auditor-pN.md), NOT in the lead's spawn-prompt composition. The lead embeds only AUDIT_CONTEXT + agent metadata; the principle interpretation layer is read by the teammate from its own definition file with a fresh context. Future maintainers who fold principle text into the lead's spawn prompt re-introduce orchestrator-prompt-drift (the variance source --all exhibited on 2026-05-07).

When to pick (single advisor / --all Agent-spawn / --team subprocess-team)

Mode	Use when	Cost (subscription token estimate)	Variance dampening
/audit p1 (single-mode advisor)	Sharpest single-principle audit; quick gut check; in-context decision review	Lowest — one advisor call, ~3-10K tokens	None (single shot)
/audit --all (5 Agent calls)	Default for grill-end audit + /slice gate; covers all 5 principles with fresh-context isolation per call; cheap enough to run on every grill	Medium — 5 Agent calls × ~5-15K tokens each = 25-75K tokens	Per-principle isolation only (no within-principle vote)
/audit --team (N×5 subprocess team)	High-stakes architecture audits where verdict-drift between runs has been observed (per-principle calibration drift, scope ambiguity, severity rubric drift); audits that lock load-bearing decisions for downstream stories	Highest — N×5 teammates × ~10-25K tokens × 3 rounds = 150-750K tokens (default N=2). Pre-spawn budget gate before running.	Within-principle vote (N=2) + cross-principle DM-debate signature. Variance metric in audit.log (variance=converged, variance=split:P2,P5, variance=scope-flagged:P2-local-vs-meta).

--team is additive to --all, not a replacement. /audit p1 and /audit --all continue to work unchanged. --team is reserved for the high-stakes case where the marginal cost of within-principle consensus is justified by the cost of locking a flawed decision.

Why per-principle isolation matters (load-bearing — do NOT collapse to one batched call OR revert to advisor())

The 2026-05-03/04 grill that fell into the P4 trap had /audit p1 and /audit p2 invoked in separate sessions; both passed. The hypothesis /audit --all tests is: "5 fresh-context-isolated calls catch what 2 separate sessions missed, because all 5 principles run on the same content within one /audit invocation." A single batched call with all 5 principles in the prompt would give the LLM permission to rationalize across principles silently — exactly the failure mode this gate exists to prevent.

Why Agent(general-purpose), not advisor() — the empirical finding from #253 AC#7 (2026-05-07). When /audit --all first ran end-to-end against the 2026-05-04 replay grill, the runner discovered that advisor() forwards the full conversation transcript to the reviewer on every call. By P3, the advisor was reading P1+P2's verdicts before forming its own — the transcript leakage defeats the "fresh context per principle" contract even though the calls were structurally sequential. The runner switched to Agent(subagent_type="general-purpose", ...) per principle: each Agent spawns with a fresh context, sees only the per-call prompt the orchestrator embeds, and cannot read the orchestrator's transcript. P4 then FAILed cleanly on the in-doc lock-before-validate violation that the 2-session 2026-05-05 audit had missed (tracker rows 89-90 — both P1 PASS + P2 PASS individually, neither saw the cross-principle gap).

Future-maintainer guard. Two failure modes silently regress this contract:

1. Batched call: "simplifying" by collapsing the 5 principle prompts into one call. The LLM rationalizes across principles. Catch: any change that produces fewer than 5 distinct calls for --all is a regression.
2. Reverting to advisor(): advisor() reads cleaner in skill prose and seems equivalent to Agent, but it leaks transcript. Catch: skills/audit/evals/check-all-flag.sh grep-guards the SKILL.md text for both Agent + subagent_type=general-purpose + an explicit transcript-leak warning. The CI script must stay PASS or --all is broken.

Per-call fresh-context isolation is the empirical fork that determined whether agent-teams (Phase B, ai-brain#252) was needed: AC#7 caught the original P4 failure mode (--all does dampen lock-before-validate), so the cheap --all primitive stayed live. BUT a separate empirical signal — per-principle verdict variance across two reproducibility runs of --all against the same 2026-05-04 fixture (P2 PASS↔FAIL on scope drift; P5 FLAG↔FAIL on severity rubric drift) — surfaced 2026-05-07 and unblocked Phase B (#252 SHIPPED 2026-05-07). The diagnosis: even with fresh-context Agent isolation, the orchestrator composes per-principle prompts from its own context, introducing an interpretation layer that drifts run-to-run. --team (#252) eliminates this by reading principle text + calibration from the agent definition file (agents/principle-auditor-pN.md) via subprocess isolation — the lead never composes principle text into the spawn prompt.

Three iterations, three contracts (history — protect against silent regression to a prior iteration):

1. Iteration 1 (#253 D9 / #255): advisor() per principle. Failed because advisor() forwards the full conversation transcript per call; by P3 the reviewer reads P1+P2 verdicts. Discovered AC#7 morning smoke 2026-05-07.
2. Iteration 2 (#260): Agent(subagent_type="general-purpose") per principle. Each Agent has fresh context, but the orchestrator still composes per-principle prompts from its own context — orchestrator-prompt drift produces verdict variance. Discovered AC#7 afternoon re-run 2026-05-07.
3. Iteration 3 (#252, this story, SHIPPED 2026-05-07): Agent(subagent_type="principle-auditor-pN") via TeamCreate. Each Agent re-reads its own principle definition + calibration from the agent file with a fresh context; the lead embeds ONLY audit context, never principle text. Within-principle N=2 consensus dampens residual variance; cross-principle DM-debate signs each finding's internal consistency with its principle.

Future-maintainer guard. Three failure modes silently regress these contracts:

1. Batched call: "simplifying" by collapsing the 5 principle prompts into one call. The LLM rationalizes across principles. Catch: any change that produces fewer than 5 distinct calls for --all is a regression.
2. Reverting to advisor(): advisor() reads cleaner in skill prose and seems equivalent to Agent, but it leaks transcript. Catch: skills/audit/evals/check-all-flag.sh grep-guards the SKILL.md text for Agent + subagent_type=general-purpose + an explicit transcript-leak warning in --all mode.
3. Folding principle text into the lead's spawn prompt for --team: future maintainer "simplifies" by inlining principle text into the spawn prompt instead of relying on the agent definition file. This re-introduces orchestrator-prompt-drift (Iteration 2's failure mode) into Iteration 3. Catch: check-all-flag.sh grep-guards that the --team workflow in this SKILL.md says principle text lives in the AGENT FILE, not in the spawn prompt.

A future maintainer who collapses the 5 Agent calls in --all, swaps them back to advisor(), OR folds principle text into the lead's --team spawn prompt defeats the corresponding mode — the --all and --team invocation sugar is incidental; per-call isolation (advisor → Agent → subprocess-team) is the load-bearing contract per iteration.

Principle Selection (p0 logic)

When no specific principle is given (/audit or /audit p0), pick the top 3 most relevant to what's currently being discussed. Pick order:

1. Principles that directly challenge the current approach (flag candidates first)
2. Principles that guard the most expensive failure mode (cost-of-wrong matters)
3. Principles most recently applied to a similar context

Single-principle audits produce the sharpest results. Default p0 to 3 but respect user's specific picks.

/audit --all overrides p0 entirely — it always runs all 5 principles regardless of relevance scoring (relevance is what the per-principle Agent calls evaluate; the user opted into the full sweep).

P3 scope check: P3 (Value Equation) requires ≥50 words of content AND structural differences between options (not phrasing-only). Skip P3 for micro-edits — it won't discriminate. Under --all and --team, P3 still runs; expect a PASS verdict with a "scope insufficient" key-finding when the content is too thin for P3 to bite. Under --team the P3 agent's scope_flag field captures the content-vs-architecture choice and feeds the variance metric.

Output

Single-mode output: present the advisor's findings directly, then append a single ## Audit Summary table:

## Audit Summary
| Principle | Verdict | Key Finding |
|-----------|---------|-------------|
| P1 | PASS/FLAG/FAIL | one-line finding |

Logged to tracker.

--all mode output: see ## Output consolidation contract above.

--team mode output: extends --all's contract additively. Same ## Audit Summary table (P1 → P5) for downstream consumer compat (/slice gate parser keeps working). Same ## Per-principle Findings section. Plus a NEW ## Variance Notes section appended after Per-principle Findings:

## Variance Notes
- P2: split (a=PASS local-scope, b=FAIL meta-scope) — variance=split:P2
- P5: converged on FAIL (severity rubric applied; both agents counted same load-bearing invariants)
- P3: scope=architecture (insufficient-content PASS — recorded for /improve calibration)

The Variance Notes block is empty when all within-principle votes are unanimous and no scope flags surfaced. The block's presence + content is what /improve reads to detect rubric drift across runs (e.g. recurring split:P2 would signal that the P2 scope calibration note in agents/principle-auditor-p2.md needs sharper rules).

Format contract: The ## Audit Summary heading and table structure (| P{N} | PASS|FLAG|FAIL |) are consumed programmatically by downstream skills (e.g. /slice audit gate). Verdict values must be exactly PASS, FLAG, or FAIL — do not alter heading, column structure, or verdict text in any of the three modes.

Tracker Integration

After each audit, append rows to the Log table in {vault}/thinking/principles/tracker.md:

| {date} | {principles} | {context} | {outcome} |

Also increment the Uses count for each principle applied.

Single / shorthand mode: one log row covering all selected principles (e.g. principles=p1,p5), increment Uses once per principle.

--all mode: one log row per principle (5 rows total, one per principle, P1 → P5), and increment the Uses count once per principle (5 increments). Each row carries {principles}=p{N} (single principle ID per row) so per-principle benefit attribution stays clean across --all runs vs single-principle runs.

--team mode: same as --all for the tracker — one log row per principle (5 rows), Uses incremented once per principle (5 increments). Each row's {outcome} cell additionally carries the within-principle agreement state (e.g. P2 FAIL within-agreed=true cross-signed=4/4, or P2 FAIL within-agreed=false split-on-scope) so per-principle drift across --team runs is visible in tracker history without needing to read the run dir.

Outcome log

Follow skill-spec.md § 11. Append to {vault}/daily/skill-outcomes/audit.log:

{date} | audit | audit | ~/work/brain-os-plugin | thinking/principles/tracker.md | commit:N/A | {result} | args="{principles}" findings={summary}

• result: pass when audit completes and produces findings (regardless of verdicts); partial if some principles couldn't be fully evaluated (e.g. one of the 5 --all Agent calls returned no parseable summary row, one or more --team teammates idled with round<3, or single-mode advisor() returned no parseable row); fail if the per-principle call failed (Agent / advisor / TeamCreate) or tracker.md unreadable
• findings: verdict breakdown capturing content quality separately from skill execution — e.g. 3FLAG+2PASS, 1FAIL+1FLAG+1PASS. Under --all and --team, the breakdown reflects all 5 principles (so 1FAIL+2FLAG+2PASS totals to 5).
• args=: principle IDs used (e.g. p1 p5, --all for the 5-call mode, or --team for the agent-team mode; under --team also include N=<n> to record the within-principle slot count, e.g. --team N=2)
• variance= (--team mode only): comma-separated variance signals from the run. Format:
  • variance=converged — all within-principle votes unanimous AND no scope flags surfaced.
  • variance=split:P2,P5 — comma-separated list of principles where within-principle votes split (verdict disagreement OR scope disagreement).
  • variance=scope-flagged:P2-local-vs-meta — when P2's scope calibration surfaced an ambiguous artifact and the agents picked different scopes; the lead must pick one before locking the verdict.
  • Multiple signals can compose: variance=split:P5,scope-flagged:P2-local-vs-meta.
  • The variance= field is REQUIRED on --team rows — check-all-flag.sh grep-guards its presence.
• tokens= (--team mode, optional but recommended): post-hoc total token usage across all teammates, recorded for AC#7 baseline calibration. Format: tokens=in/out (e.g. tokens=180000/45000). Lead reads from session metadata if available.

--all and --team modes write a single outcome-log row each (not 5) — the per-principle granularity lives in the tracker log; the outcome log captures the run-level result + variance.