ORCHESTRATE Rules — FLOW, STOP conditions, lanes, firewall

Binding for every agent working in hierarchical-aif/.

1. FLOW — and the NEXT_ACT requirement

PERCEIVE -> BOUND -> PREDICT -> CHOOSE -> ACT -> OBSERVE -> UPDATE -> RECORD -> NEXT_ACT

Every RECORD step must end with NEXT_ACT. If no immediate ACT is possible, write the exact blocker and the exact receipt needed to unblock it.

Alignment is not a stop state

Do not stop after:

• saying you are aligned
• writing a report
• creating a ledger
• tests passing
• discovering a defect
• the next step being long

"Awaiting" is permitted only for a genuinely running process, and while awaiting you must work a safe non-interfering lane (§4).

Append a FLOW card to reports/FLOW-JOURNAL.jsonl for every major action:

{"flow_id":"F0NN","gate":"H-AIF-G#","step":"PERCEIVE|BOUND|PREDICT|CHOOSE|ACT|OBSERVE|UPDATE|RECORD",
 "question":"...","state_before":"...","alive_hypotheses":["..."],"falsifier":"...",
 "action":"...","receipt":"...","state_after":"...","next_action":"..."}

Externalize decisions as audit-ready records. Do not dump private reasoning.

2. STOP conditions — the only legitimate stops

Everything else is a FLOW action, not a stop.

3. Lanes — scope every impact

Forbidden: "P6 is weaker". Required: "P6 for C11 U4 is withdrawn until the corrected C11 run lands; P6 for duration-only B3/B4 is unchanged."

A defect in one lane never licenses a global retreat.

4. Safe non-interfering work while long runs execute

Allowed:

• complete F-side implementation and tests
• prepare prediction records and launch scripts for queued cells
• prepare the F-side scoring harness
• update the defect closure ledger
• run the claim guard
• update the biological-parity receipt map

Forbidden while runs execute:

• new broad exploratory audits
• any new held-out read
• new mark-process analyses
• new parity claims
• unscoped weakening statements

5. The D5 held-out data firewall

Reading held-out data is irreversible. Read-only is not consequence-free.

D5: a read-only track was briefed to report "empirical marginals of direction/jump per state." The brief carried no split boundary. The agent read the held-out mark channel to answer it. Nothing was written, no file was modified, and the channel was destroyed anyway — no mark-process claim on Wadhwa-2022 can ever be PROSPECTIVE again. One study, no second holdout, unrepairable.

Rules now in force:

1. Every analysis declares a split boundary: TRAIN_ONLY · HOLDOUT_ALREADY_SPENT_DURATION_ONLY · HOLDOUT_ALREADY_SPENT_DIRECTION · HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY · INDEPENDENT_TRANSFER_REQUIRED · PROSPECTIVE_NEW_DATA_ONLY · NO_DATA_ACCESS_NEEDED
2. Consult docs/DATA-CHANNEL-SPEND-LEDGER.md before touching any field.
3. Any task that might touch held-out fields first writes protocols/<TASK>-DATA-ACCESS-PROTOCOL.md.
4. Every subagent brief pastes the firewall from docs/SUBAGENT-BRIEF-TEMPLATE.md.
5. NOT_CHECKED — would require holdout access is a correct and valuable answer, never a failure.

The firewall is also enforced in code: events.load_events(mode=MARK_RETROSPECTIVE) raises HoldoutMarkAccessError unless the caller explicitly acknowledges retrospective status.

6. Defect closure discipline

A defect is a completed FLOW result only when routed into repair, quarantine, rerun, transfer, or falsification. Every defect row must carry closure_status, last_action, next_action, blocked_by, and not_blocked_lanes.

Closure vocabulary:

CLOSED
CLOSING_BY_TESTED_REPAIR
QUARANTINED_WITH_TRANSFER_ROUTE
OPEN_AWAITING_CORRECTED_RUN
FALSIFIED_AND_RETIRED

If a defect is not closing, name the exact missing receipt.

7. Evidence discipline (inherited, still binding)

• Partial logs are not results. Never interpret a running job's stdout as a finding.
• A crashed run is FAILED_RUN, not a scientific negative.
• actual_N = 0 → NOT_RUN; 0 < actual_N < planned_N → PARTIAL_NOT_ESTABLISHED unless a prospective sequential stopping rule was declared before the run.
• Point estimates are never verdicts. A CI crossing the threshold is NOT_ESTABLISHED.
• The experimental unit is the motor. Bootstraps resample motors, never events.
• Underpowered is not equivalence. But note: a small yet consistent paired effect is resolvable — "small" and "unresolvable" are different claims.
• No floor anywhere. A non-finite log density HALTs.
• Prediction records are committed before the run they describe.
• Never overwrite a historical artifact. Corrected results take new names.

8. Strong inference

Keep alive until a predeclared discriminator removes them:

• H_PARITY as target hypothesis
• the current design as control
• M0/M1/M2/M5/M8 as adversaries — M2 currently out-predicts the mechanism on held-out data, and that adverse result is retained
• negative controls: invalid censoring, event-level bootstrap, orientation/label scrambling

Nature may motivate the architecture. Only a gate moves a status.