UNI-FLAGELLUM — Next-cycle execution plan (OODA, ultracode 3-body)

Durable location: hierarchical-aif/docs/NEXT-CYCLE-EXECUTION-PLAN.md (this file is the single source of truth; the rendered companion is hierarchical-aif/docs/NEXT-CYCLE-PLAN.html, a self-contained, CPU-only, no-network standalone page — open it directly with file://). Anchor: HEAD e21747c, branch hierarchical-aif/motor-stack, nothing pushed. Method: OBSERVE→BOUND→PREDICT→ACT→VERIFY→FALSIFY→UPDATE→RECORD, processed by a 9-agent ultracode workflow with a 3-body adversarial review (2 new personas). This document is DESIGN — authorized to build, not a result. Nothing here has moved a P-level. It is a standalone plan document, not a served route.

0. Honest headline (lead with the outcome)

• The model math is done for this batch; the binding constraint is DATA, not math. The first missing rung is P4 transfer, which is irreducibly external — no in-repo work closes it.
• This cycle exactly 4 gates can move to durable green with in-repo OBSERVED evidence — and zero of them raise any P-level. They are: restore P0, restore H-AIF-G1, bring the D12 remediation gate PENDING→PASS, and close H-AIF-G9.
• Green now = 16 of 31 gates — but only ONE of those 16 (P2) is genuinely observational, and it is LIMITED (mark fields quarantined, raw archive not located). The other 15 are computational- integrity / equation-implementation / defect-closure gates. Do not read "16 green" as "16 biological observations."
• 10 gates are EXTERNAL-BLOCKED (P4, P5, P6, P7, P8, H-AIF-G8 raw archive, D5, D6, D9, and the new accessibility gate which ships NOT_ESTABLISHED). 1 is design-only (H-AIF-G6, the human L0..L12 stack).
• The adverse result rides at the top, co-equal: a simple lognormal (M2) still out-scores the flagship two-timescale UNI model (M3) on held-out data (M2 −3.013 vs M3 −3.050 nats/event; the M3−M2 interval crosses zero → unresolved at this sample size, not "M2 wins," not "no difference").
• P8 = FULL_PARITY = false. No parity, active-inference, life, consciousness, or "soul" claim is minted from passive data — at any layer, including the child-facing one.

Gate forecast: greenNow 16 · remediation-clearable this cycle 4 · external-blocked 10 · design-only 1 · total 31. oneCycleClearCount = 4.

1. OBSERVE — what the resample actually showed (durable receipts)

Live re-run at HEAD e21747c (not copied from the snapshot doc). Reproduce with the §1 commands in hierarchical-aif/docs/CURRENT-STATE-AND-NEXT-ACT.md plus the four gate commands below.

* Preserved discrepancy: the live-state doc §1 predicts 505 passed, 2 skipped; observed 504 passed, 3 skipped (same total 508). One test skipped in this environment that passed in the recorded run. Snapshot↔environment drift, flagged not hidden.

Two resonance disagreements to keep visible, with their receipts (do not launder):

1. publicArtifactMismatchDetected: true (scripts/independent-science-check.mjs:52) is the independent oracle asserting the expected G03_PUBLIC_ARTIFACT_PARITY = FAIL (lines 44–46) — i.e. science:verify PASS includes verifying that the public artifact correctly does not match and that parity is correctly false. It is a green that confirms an honest negative. Surface it with this receipt; never fold it into "P0 strengthened."
2. The working tree is DIRTY — M hierarchical-aif/docs/CURRENT-STATE-AND-NEXT-ACT.md (a 106+/109− uncommitted rewrite). This trips H-AIF-G1's own verbatim falsifier ("a non-clean working tree presented as clean"), so P0 and H-AIF-G1 are currently DIRTY-CLAIM-WITHDRAWN until the tree is clean. This file is USER-OWNED (per CLAUDE.md, treat unknown working-tree changes as user-owned) → committing or reverting it is PRINCIPAL-GATED (Michael's call), not an agent action.

Topology. Real repo UNI-FLAGELLUM/ @ e21747c. UNI-FLAGELLUM-math-workbench/ is a git worktree of the same repo on feature/scientific-math-workbench @ c23f686; it adds the /math-workbench route (6 views) the base app lacks. Three distribution archives sit on disk (see §3).

2. Independently verified defect facts (observed, not relayed)

3. Distribution-archive exposure record (Custodian requirement)

Three built, distributable objects are on disk at the wrapper root. Member-level D5-safety is NOT_CHECKED / UNVERIFIED (opening members is firewall-barred to design-only agents). Treat as UNVERIFIED, never "safe." Transmission is external and principal-only. These packets are NOT closure evidence.

4. The 3-body persona charter (odd, majority-adversarial, tension-holding)

Every decision is held by ≥3 perspectives, always odd, majority adversarial. Michael sits INSIDE as a present member; the two NEW personas hold the adversarial majority; the existing lab team plugs in for FE-touching changes.

3-body verdicts this round: Custodian REJECT (dirty tree; D12 narrative-only; archives on disk; mismatch flag laundering) · Devil's Pedagogue SIGN_WITH_CHANGES (teaching surface unscored; OBSERVED-class conflation; adverse result demoted) · Michael SIGN_WITH_CHANGES (accessibility unscored; teaching stack homeless; shortest-path-to-P4 unranked). All required changes are folded into §6; all preserved tensions into §7.

5. What the two adversaries added (beyond the authorization)

1. The teaching surface is unscored and unclamped. lib/walkthrough.js + spoken narration (the channel a child hears) is outside claim_guard. Resonance channel #5 (UI truth badge + species label) has zero gate coverage. → new UI-truth-badge runtime gate + species-label gate + a sibling *.js/narration clamp; until then annotate "claim_guard 0 EXCLUDES the teaching surface."
2. OBSERVED-class conflation. "16 green" must be split: 1 (LIMITED) observational vs 15 process/implementation/defect gates. No banner may let a reader hear "16 biological observations."
3. Adverse M2-beats-M3 result demoted. Promote it into the headline, co-equal with the green count.
4. Reconstruction-beside-microscopy (walkthrough step 2) and multi-species cutaway (step 3) need a runtime receipt that the truth badge + per-species label render, can't be desynced, and survive a screenshot/share — a doc table is not runtime enforcement.
5. publicArtifactMismatchDetected must be surfaced with its receipt (see §1).

6. Execution plan — 11 steps, strictly sequenced (containment first)

Sequencing rule (principal M25 authorization): CONTAINMENT precedes everything; METHOD guards land as additive sidecars; UI-WORKBENCH + LAYMAN-LAYERING add the honest teaching surface; SCIENCE-EXTERNAL is pre-registration ONLY. P4/P5/P6/P7/P8 stay external-blocked and unmoved. Nothing is pushed or transmitted — building in-repo is authorized; sending is principal-only.

Track A — CONTAINMENT (owner: Custodian)

Step 1 — Restore the integrity floor. [PRINCIPAL-GATED] Commit or revert the dirty user-owned hierarchical-aif/docs/CURRENT-STATE-AND-NEXT-ACT.md, then re-run science:verify / cross-study:verify / pytest (504p/3s/1xf) / claim_guard 0 / numeric_provenance 2070-0. Preserve e21747c — no amend/rebase/force-push. → Durable evidence: clean git status -sb + re-run PASS receipts → P0 and H-AIF-G1 re-established green. → Blocked by: the file is user-owned — Michael decides commit vs revert; agents do not touch it.

Step 2 — Build d5_distribution_guard (SEPARATE guard; do NOT overload numeric_provenance_guard). Runs on the STAGED tree AND the finished round-trip archive. Structurally parses JSON/JSONL string values (not just keys); scans md/txt/code/logs/manifests/generated reports; inspects or rejects nested archives; detects real event-id patterns, motor-ids co-located with holdout mark context, record-shaped stateN/nextStateN/jump/direction+holdout combos; distinguishes schema/declaration prose from value-bearing records; permits synthetic fixtures only with an explicit synthetic marker; fails on pre-redaction examples + mutation variants; emits a machine-readable report of every inspected artifact

• exception. Failing tests written first. → Release falsifier: any real event-id / associated motor-id / event-level mark tuple surviving any shipped artifact FAILS release.

Step 3 — D12 incident-remediation (build the receipts that don't yet exist). Redact event-level identifiers, motor-ids tied to the mark defect, and record-shaped mark tuples from EVERY distributable surface: reports/D6-INGEST-NEXTSTATE-RANGE-CHECK-DEFECT.md; protocols/MARK-PROCESS-TRANSFER-RESCUE-PROTOCOL.md; ledgers/HIERARCHICAL-AIF-DEFECT-LEDGER.md; tests/motor_stack_aif/test_nextstate_range_check.py; + every package-stage scan hit. Prefer "below-physical-minimum target-state marks" over the exact value; aggregate counts may remain only with no event-id/motor-id/reconstructable tuple. Narrow the handoff claim from "no raw mark-key values." Add a tracked WITHDRAWAL notice for UNI-FLAGELLUM-haif-closure-e21747c.zip (name + SHA-256, no records reproduced; states: WITHDRAWN_D5_UNSAFE / stop redistribution / request deletion / record acks without assuming them / name successor once it exists). Record D12 as incident=NEGATIVE, remediation gate=PENDING, historical archive=WITHDRAWN_D5_UNSAFE, residual=prior distribution cannot be recalled. → D12 never collapses to CLOSED_BY_REDACTION; the historical NEGATIVE remains even after the gate passes.

Step 4 — Build the D5-safe successor archive (this is what moves D12 PENDING→PASS). Pipeline: new anchor → d5_distribution_guard scan of staged tree → manifest → archive → unpack clean → rescan → recompute manifest+hash → confirm no withdrawn blobs/patches/bundles → include withdrawal notice

• D12 report. Use a SANITIZED ANCESTRY RECEIPT (commit ids/parents/timestamps/payloads/name-status/ commands/receipt hashes), classed GIT_ANCESTRY_RECEIPT_PROVIDED; FULL_HISTORY_RECONSTRUCTION_NOT_INCLUDED — do NOT place a git bundle in the archive (it would carry the withdrawn blobs). → Blocked by: re-distribution/transmission stays principal-gated. Building is authorized; sending is not.

Track B — METHOD (additive sidecars; owners: Devil's Pedagogue + Custodian; move NO P-level)

Step 5 — Fail-closed method guards. (a) score.py motor_equal_nlpd/per_motor_means: replace bare zip with checks for equal event-score & motor-id lengths, non-empty, finite scores, non-empty & finite paired motor arrays, valid bootstrap request, complete expected-motor coverage at aggregation. (b) Optimizer fail-closed: a fitted object whose optimizer reports non-convergence must not enter scoring/result assembly — one shared eligibility assertion across every scoring entry point. Mutation falsifiers: remove one motor-id or one event score → aggregation raises before any mean; force finite params + unsuccessful termination → scoring halts, no artifact written.

Step 6 — Canonical P-ladder metadata restoration in compare.py:546. → P4 transfer, P5 intervention, P6 structural/mechanistic, P7 independent replication, P8 full verdict. Issue correction SIDECARS with new hashes; do NOT overwrite frozen result/prediction JSON. status.py overrun-semantics stays PENDING (must resolve before external P4; need not delay D12).

Step 7 — Failure-count + D9 corrections. Add CORRECTION-NOTICE-e21747c.md quoting the false historical failure-count sentence (PHASE-…-CLOSURE.md:32), the corrected cell-specific status, and the superseding commit (original commit message stays historical). D9: separate historical_evidence_at_defect_open / current_repo_disposition / external_handoff_evidence_limit; add a status addendum; keep B4C10 & F-side NOT_SATISFIED; do NOT edit original B4C11 prediction bytes. → Blocked by: transmission of the correction notice is principal-gated (Michael only) — prepare, never send.

Step 8 — Reduce the disclosed G5 resonance disagreement in the same change. Correct the stale H-AIF-GATES.md "IN PROGRESS" to match the ladder-map "COMPLETE" (no frozen artifact touched). Closes H-AIF-G9. Moves no P-level.

Track C — UI-WORKBENCH + LAYMAN-LAYERING (owners: Michael + Organic Operator + Devil's Pedagogue)

Step 9 — Add a first-class ACCESSIBILITY/PEDAGOGY gate + honest OUTSIDE-IN-REPO rendering, bound to the workbench (route /math-workbench, named commit). Render the TRUE gate state (16 green / 4 clearable / 10 external-blocked / 1 design-only; P8=false; first-missing P4) where no simplification layer may relabel a status. Add the UI-truth-badge runtime gate + species-label gate + a *.js/narration sibling clamp (§5). CPU-only; no LLM/GPU/WebGL/WebGPU/Three.js/analytics/accounts/hidden-network. Ship the accessibility gate NOT_ESTABLISHED with a stated falsifier — usability is a claim, not an achievement. Flag app/chatgpt-auth.ts as a release risk (accounts/network the release forbids).

Step 10 — Declare the NON-science lane for the L0..L12 teaching stack. A reduced teaching model over a reduced/synthetic world, carrying NO P-level, NO expected_free_energy on passive data, NO mechanism claim. Keep H-AIF-G6's test-enforced absence of any EFE function in passive-data scope. This gives the platform ambition an honest home without laundering; it does not build L0..L12 content on passive data.

Track D — SCIENCE-EXTERNAL (owner: Experimentalist / AIF Theorist; pre-registration ONLY)

Step 11 — Pre-register the shortest path to P4 (D9 discipline, committed BEFORE any data is touched). Rank external blockers by expected information gain; commit the scoring rule for the single cheapest independent transfer cohort — an independent single-motor dwell series with enough independent MOTORS to beat the C01 nesting-blindness (M0 self-win 0.290). Do NOT execute. P4 stays external-blocked and unmoved; P4 execution is gated on containment + guards in place AND real external data acquired.

7. Preserved tensions (do NOT resolve by weakening)

1. M2/lognormal still leads the held-out leaderboard — kept alive by design; mechanistic ambition may not bury the adverse result.
2. P8 FULL_PARITY=false, first missing rung P4 transfer, irreducibly external.
3. Mechanism is RETROSPECTIVE-ONLY / TRANSFER-REQUIRED (D5/D6); identifiability and architecture are not mechanism; no aggregate P6 is asserted.
4. L0..L12 stack + any EFE stay OFF passive data via test-enforced absence (H-AIF-G6 design-only) — ambition vs falsifiability, preserved not collapsed.
5. Accessibility / 5-to-199 path ships NOT_ESTABLISHED with a falsifier — a hypothesis, not an achievement; any layer that relabels a status is rejected.
6. D10 sub-floor guard: the F-side ~2.5e-7-nat (~168,000× sub-floor) effect may never be "beats M7."
7. D12 incident stays NEGATIVE permanently even after the remediation gate passes; prior distribution cannot be recalled; un-scanned archive members are UNVERIFIED, never safe.
8. Reviewer / correction-notice transmission stays principal-gated (Michael only).
9. H-AIF-G8 raw archive not located; packaged event JSON may never be read as raw confirmation.
10. Species + units fences: E. coli behavioral kept distinct from Salmonella/Bacillus structural; nats (free energy, log score) never converted to joules (τ·Δθ work).

8. Algebra-up layering (the child-to-scientist ladder, badge-invariant)

Every load-bearing object is taught on the same 3-rung scaffold; the rungs climb only in MATHEMATICAL machinery, never in evidential authority. The truth badge is IDENTICAL across all three rungs of an object, and each object states exactly what the simplification drops. The full content is rendered in the companion page hierarchical-aif/docs/NEXT-CYCLE-PLAN.html. Objects: dwell-time first-passage survival (MODEL, gates G03/G05/G06 FAILED); held-out log score / NLPD (MODEL); M3 two-timescale vs M2 lognormal (MODEL, M2 leads); motor-cluster bootstrap CI (MODEL); the P0..P8 ladder itself (DESIGN); free energy in nats vs τ·Δθ work in joules (MODEL, no conversion, E-M01 open); the exact log-odds Bayesian update / UNI gear (MODEL); finite-lattice cooperativity falsifier (MODEL, X06/X11 FAIL); CV²>1 overdispersion (MODEL, the one clearly-resolved held-out result).

10 overclaim guards bind the scaffold (badge invariance; prediction-is-not-mechanism; no nats↔joules conversion; adverse results ride with every summary; resolved-vs-unresolved decided by the interval; conjunctive ladder; species/class separation; no consciousness/life/soul claim; D5 firewall in every teaching example; open defects stay visible even at the child rung).

9. Authorization, rollback, and what this cycle is NOT

• Authorized to build: steps 2–11 in-repo (containment guards, method sidecars, teaching gates, pre-registration). Principal-gated: step 1 (user-owned file), all transmission/redistribution.
• Not authorized: amend/rebase/force-push of e21747c; git bundle in the successor archive; editing frozen audits/phase-b|c|d/** or frozen result/prediction bytes; overwriting the user-owned working file.
• Rollback: every step is additive (new files / sidecars) or a scoped metadata edit; e21747c and all frozen artifacts remain byte-identical, so rollback = discard the new files.
• This cycle raises NO P-level. It is containment + method integrity + an honest teaching surface. The next scientific gain (P4) requires external data and begins only after containment + guards are in place.

10. Governance note — critical information lives IN this repo

This plan and its rendered page are version-controlled files inside the repository, not an external hosted surface. Critical UNI-FLAGELLUM information (plans, gate state, evidence, findings, decisions) is kept in the repo, reproducible from recorded in-repo artifacts, with no hidden external or network dependency — consistent with the truth/resonance contract and the CPU-only, no-network release rules.