UNI-FLAGELLUM Agent Operating Contract

⟢ RESUME POINT — read this before anything else (marked 2026-07-28)

We are in PHASE 9 — RETURN TO RESONANCE, mid Stage 4. This block is navigation and measured state only. It amends no law. Every law below it stands unchanged.

If you have no context, read UNI-FLAGELLUM/docs/control-plane/AGENT-CALIBRATION-PROMPT.md first. It is self-contained and carries every trap that has already caught someone.

• The plan is not a document. UNI.Minecraft/evidence/remediation/phase9_plan.json is the single source of truth. UNI TRACK http://127.0.0.1:8102/ renders it live and Gaia projects it verbatim. When a step completes, edit that file; the surfaces follow. viewer/verify_plan_consistency.cjs now holds it to its own vocabulary — it exists because the plan carried two different next acts at once.
• Register: UNI-FLAGELLUM/docs/control-plane/phases/PHASE-9-REMEDIATION.md
• Resume detail: UNI-FLAGELLUM/docs/control-plane/RESUME.md

THE STATE BELOW IS GENERATED, NOT WRITTEN. Every number between a BEGIN GENERATED and its END GENERATED is produced by node viewer/generate_state_blocks.cjs from the artifact it describes. Do not edit inside a block; edit the artifact and regenerate. This banner used to carry these numbers by hand and was measurably wrong in six places on 2026-07-29 — 25 gates in one paragraph and 23 in another (both wrong, 28), a ledger of 31 (32), "six" lab gates (seven), and a next act that had shipped six hours earlier.

NEXT ACT: RESUME-THE-LIVE-FRONTS — AGENTS

Two live fronts, both agent work, both measured live since 2026-08-19: (1) the whiteboard defect programme -- docs/whiteboard/DEFECTS-AND-REPAIRS.md, every repair proven on the real engine before it ships; (2) the public University assembly -- labs, course, falsification wall and contribution surfaces served from the documentation estate, every page pinned to the commit it was read from. The checkpoint this key used to name was ruled out by the operator; see was.

Declared at docs/whiteboard/DEFECTS-AND-REPAIRS.md and the UNI.Public estate (the University routes). Blocked on: nothing -- both fronts accept agent work today; operator dependencies inside them (per-repair co-signs) are queued as they arise, and none gates the next step

Retired: L6 (Stage 4 step 4.6 -- build L6, THE GAUNTLET THEN THE CO-SIGN, shipped 6234f3d); CHECKPOINT-E (WITHDRAWN by operator ruling 2026-08-10: the two-image shot surface at the lab viewer was ruled a fatal hallucination -- never resurface it.).

Plan: 7 stages · 43 steps (31 DONE · 1 BLOCKED · 8 PLANNED · 3 OPERATOR) · 7 builds under step 4.6, 7 DONE.

Gates: 36 registered, of which 33 ci:true and 3 ci:false (colony, hud, overlays — listed, never run, never a fabricated pass). 7 lab gates (lab-l0, lab-l1, lab-l2-shot, lab-l3, lab-l4, lab-l5, lab-l6).

Both numbers are stated because both were written before without saying which was which: one banner paragraph said 25 and another said 23, and a single file said 23 at one line and 25 at another. Neither was the registered count.

Gate ledger evidence/gates.ndjson — ca8fd61ab5380994..., 212 rows / 112 unique names. Last row per name: 94 PASS · 5 PARTIAL · 12 PENDING · 1 FAIL.

The per-name tally is stated as such because the per-ROW tally is a different set of numbers, and a count whose derivation is unstated is how a backlog and the history of a backlog came to be reported as one word.

Registry vs. the canonical ledger: of 36 registered gates, 1 appear in evidence/gates.ndjson and 35 do not (0 of those carry a glob gate_row, which no kebab-case row can ever bear). gate_row.schema.json says every gate the project claims MUST be represented there.

The intersection is NOT empty, and four governing documents said it was. They declared "EVERY registered gate has ZERO rows" and "the intersection is empty by id and by gate_row" for two weeks after a row landed for one of them on 2026-07-17 — inside the paragraph that says these numbers are generated. It was hand-written. It is not any more.

Authoring the missing rows is S4 — the operator's, but the blocker is not his signature: desk.preRegistration() reports most of them blocked on an empty receipt_path the schema requires, which is a pre-registration document an agent owes him. He could not append them today even if he wanted to.

Control-plane ledger: 32 entries, tip b90b74980f47b93a... at seq 32. Anchor declares length 32, head b90b74980f47b93a... — they agree.

Three things are deliberately NOT stated here, because no committed file can hold them honestly. They are facts about a run or about now, not about the tree:

question	the command
Are the trees clean?	git -C <tree> status -sb
Does the Elixir suite pass?	mix test
Do the gates pass?	node viewer/gate_runner.cjs

This banner used to answer all three. The gate-runner answer was measured at 06:01:09 on 2026-07-29 and was false by 06:04:06 — a half-life of 176 seconds — and it was committed reading as present tense. Run the commands.

Four things that must not be softened:

1. The off-box witness is COMPROMISED — node2 accepts the writer's key, independent_custodians: 0. The anchor stands on git alone: tamper-evident, not unforgeable. Removing that key is S1 — the one repair an agent must not perform.
2. Phase 7's witness clause still FAILS. Its other failing clause — "two fixtures distinguishable with no text read" — is closed: the renderer exists (viewer/lab/, and the lab-gate count is in the generated block above, not restated here), and verify_shot.cjs --mutate proves it bites in greyscale. Measured 2026-07-29: both images are real and differ — /api/lab/shot?swap=0 is 3371 bytes and ?swap=1 is 3375, both valid PNG with different sha256, embedded side by side at viewer/lab/l6.html:52-53.

3. Most registered gates have no row in the canonical ledger — the count is in the generated block above and is NOT restated here. This entry used to read "EVERY registered gate has ZERO rows... the intersection is empty by id and by gate_row" — false since 2026-07-17, and it was hand-written inside the banner that says its numbers are generated. gate_row.schema.json says every gate the project claims MUST be represented there. Appending is S4 — the operator's — but the blocker is NOT his signature: most rows are blocked on an empty receipt_path the schema requires, which is a pre-registration document an agent owes him. The desk at /lab/l5 prints each exact line.
4. The go-live guard is real and refuses all seven paths, and it is presence_evident, NOT unforgeable. F31 has code, a gate and an operator's prover. It binds this codebase's paths — the OBS WebSocket on :4455 still has no authentication (S2, his) — and it is bound to ::, ALL INTERFACES, reachable from the LAN and the tailnet. Every prior line here said 127.0.0.1:4455, which was FALSE and understated the exposure (LIMITATIONS.md f31.obs-unauthenticated, measured 2026-07-29).

The road to air runs THROUGH the science: colony_on_program is blocked on forage-pureworld-graduation, whose runner runs/pureworld_qa_gate.exs still raises @scaffold. And no verdict has yet been authored about a real scientific claim.

Corrected 2026-07-28 (this banner was false in seven places): it said 42 steps (43), 964 tests, "the renderer was never built", "nothing detects running-but-not-the-committed-bytes" (step 1.1 built boot identity on all four bodies), "the ledger stopped recording at Phase 5", "F31 has no code and no test", and "NEXT ACT: Stage 1 step 1.1" — four stages stale.

Corrected AGAIN 2026-07-29, and this time the fix is structural. Within six hours of that correction the banner was stale again, in six places: it said 25 registered gates in one paragraph and 23 in another (28), a ledger of 31 (32), "six" lab gates (seven), gates.ndjson unchanged at 964ea25c… (it moved to 1daac912… when a probe row landed), and a NEXT ACT of "build L6" six hours after L6 shipped at 6234f3d — while the plan itself said the next act was Checkpoint E. AGENT-CALIBRATION-PROMPT.md tells every fresh agent to obey the next act before verifying anything, so a fresh agent would have rebuilt a finished build. That is why the numbers above are now generated and no longer written. A hand-written number is a claim with a half-life; these had a half-life of six hours, and one of them — a gate-runner tally — had a half-life of 176 seconds.

Corrected A THIRD TIME 2026-07-30, and the third correction was of the second one's own claim. The paragraph above used to end by saying all three copies of this block were byte-identical, md5 003fc92d…. They were not, and the structural fix itself is what made them differ. The generated blocks were installed in the two TRACKED copies and not in the third, so for a day the count was 6 blocks, 6 blocks and ZERO — and that third copy, Documents/UNI-Flagellum/CLAUDE.md, the file an agent starting there reads first, still declared "NEXT ACT: Stage 4 step 4.6 — build L6" with L6 finished and receipted. The precise failure was in UNI.Minecraft: viewer/state_blocks.cjs defined an OUT_OF_TREE root and exported it, but made it the root of no declared document — so verify_claims.cjs, the gate whose entire reason for existing is a stale NEXT ACT: declaration, could not see the one document that still carried one. That blind spot also means the 07-29 audit's own count was an undercount: it reported five documents carrying the stale declaration, and the true number was six — the sixth being the copy no instrument could reach. All three copies now carry all six blocks and drift under the same gate. The untracked copy is still tracked by no git repository, so no diff and no CI run can ever reach it; only the gate can, and only when someone runs it. That remains a standing hazard, not a fixed one.

This repository is a living science instrument. Work as one coherent organism, but never erase the boundaries that make the organism scientifically honest. The world process, recorded observation, reconstruction, reduced model, UNI analogue, inference state, human interpretation, and release claim are distinct.

Mission

Build the strongest reproducible account of bacterial flagellar-motor behavior that survives serious alternatives. Map where the model works, where it fails, and which observation would most reduce uncertainty. Passing software tests is necessary but is not proof of biological parity, general intelligence, human parity, or scientific significance.

The released product must remain CPU-only and must contain no LLM inference, GPU computation, WebGL, WebGPU, Three.js, analytics, accounts, or hidden network calls. Development agents may use their own tools, but those tools may not become undeclared runtime dependencies.

Resonance and flow

Resonance is a testable consistency condition across:

1. source observation and provenance;
2. declared variables, units, equations, and assumptions;
3. implementation and deterministic runtime state;
4. prospective prediction and later observation;
5. UI truth badge and species label;
6. gate criterion, result, uncertainty, and limitation;
7. report, notebook export, reproduction command, and artifact hash.

If these disagree, preserve the disagreement and reduce it through evidence. Do not create apparent harmony by weakening tests, retuning a holdout, changing labels, suppressing adverse results, or rewriting history.

Follow this Active Inference execution flow for every bounded change:

OBSERVE -> BOUND -> PREDICT -> ACT -> VERIFY -> FALSIFY -> UPDATE -> RECORD

• Observe: inspect repository state, instructions, evidence, and current gates.
• Bound: name the claim, system boundary, experimental unit, species, scale, dependencies, authorization, and rollback.
• Predict: write the expected result and falsifier before changing code or revealing held-out evidence.
• Act: make the smallest reversible change that can test the prediction.
• Verify: run the narrow test, then the relevant full gates.
• Falsify: test serious alternatives, negative controls, mutations, leakage, misspecification, and edge cases.
• Update: revise confidence and next action from results, including null or adverse results.
• Record: preserve commands, outputs, hashes, uncertainty, limitations, and reproduction instructions.

Truth contract

• Only source-pinned recorded measurements may be labelled OBSERVED.
• A reconstruction, simulation, derived field, inferred latent state, or model output may never be relabelled observed.
• Keep E. coli behavioral evidence separate from Salmonella and Bacillus structural evidence. Never imply that they came from one measured specimen.
• Keep thermodynamic work tau * delta_theta separate from variational free energy and preserve their units.
• Keep priors, likelihoods, approximate posteriors, exact model posteriors, policies, predictions, actions, observations, and residuals distinct.
• Never silently mix mean-field, Bethe, marginal, or message-passing schemes.
• Do not count frames, time points, or repeated events as independent biological replicates when the experimental unit is a motor, cell, or culture.
• Calibration, training, holdout, and prospective evidence must remain separate.
• A prediction is prospective only if it was committed before its observation.
• Failed, blocked, external, and not-run gates remain visible.

Repository discipline

Begin by reading this file, README.md, relevant files in docs/, protocols, evidence manifests, gate ledgers, and tests. Then inspect:

git status -sb
git log -3 --oneline
git remote -v
node --version
npm --version
python --version

Treat unknown working-tree changes as user-owned. Do not reset, discard, delete, or overwrite them. Use isolated branches or worktrees for destructive tests and mutation experiments. Never push, publish, deploy, change access, or mutate an external system without explicit authorization.

Use documentation-first TDD:

1. state the claim and acceptance criterion;
2. identify evidence, assumptions, units, and falsifier;
3. add a test that fails for the correct reason;
4. implement the smallest correction;
5. run focused and full validation;
6. update the gate ledger and documentation;
7. preserve pre-fix evidence and rollback instructions.

Required validation

Use a genuinely clean clone for release claims. The deployment environment must remain compatible with npm 10 as well as the declared local runtime.

npm ci
npm test
npm run lint
npx tsc --noEmit
npm run science:verify
npm run cross-study:verify
npm run cross-study:verify-raw
npm audit --omit=dev --audit-level=moderate
npm audit

If a required dataset, archive, instrument, credential, or service is absent, mark its gate BLOCKED, NOT RUN, or EXTERNAL VALIDATION REQUIRED; never report a pass. Report production and development dependency risk separately.

Numerically rederive important results through an independent implementation. Audit tests for vacuous success, circular oracles, shared implementation, unjustified tolerances, leakage, pseudoreplication, and missing negative controls. Mutation tests must demonstrate that truth laundering, species swaps, hash changes, split leakage, sign inversions, normalization errors, and removal of adverse results are detected.

Deep falsification after green gates

All existing gates passing starts deeper work; it does not end it. Run serious model competition, ablation, parameter recovery, posterior predictive checks, negative controls, robustness sweeps, leave-one-unit/study/condition-out tests, and frozen prospective predictions. Compare against simpler and flexible alternatives on identical splits and scoring rules.

Map the validity domain across species, strain, motor, cell, load, PMF, stator state, temperature, viscosity, CheY-P condition, apparatus, timescale, study, and model formulation. Classify every region as supported, tentatively supported, contradicted, unidentifiable, unobserved, or extrapolation-only.

Prioritize experiments by expected information gain and their ability to make competing models disagree. A potentially transformative discovery requires risky prospective prediction, independent replication, explanatory reach, and survival against serious alternatives. Do not use prestige as a gate.

Communication

Lead with observed outcomes. Separate fact, inference, hypothesis, and ambition. Always include adverse results and limitations. Every reported delta needs:

• severity and affected claim;
• file and line;
• reproduction command and evidence;
• root cause and minimal correction;
• required failing test;
• validation, rollback, and scientific impact.

Use docs/CLAUDE_ULTRACODE_INDEPENDENT_AUDIT_PROMPT.md for the full independent audit and paste-back contract.

The parity ladder is the spine; state it honestly every time

Every status report names where the work sits on P0..P8 (computational integrity, equation/implementation, observational, held-out predictive, transfer, interventional, structural/mechanistic, independent replication, full verdict) and gives the receipt for each: the exact artifact, gate id, and status. P8 is conjunctive — any single required FAIL, CONTRADICTED, NOT_ESTABLISHED, BLOCKED_EXTERNAL, NOT_RUN, or INVALID_PROVENANCE makes FULL_PARITY=false, and the report names the first unsatisfied level. This is a statement about what the current evidence licenses, not a limit on the pursuit: every gate that real evidence moves, moves, and the goal of a faithful, fully-validated model stays live. Per plan §1 the boundary is on CLAIM WORDING — the released product may not assert an unqualified claim of "full and exact parity with nature" or call itself "digital life"; the strongest allowable positive claim is bounded parity at the declared observational and intervention resolution, over the frozen validity domain, within the frozen tolerances, supported by independent evidence and reproducible from the recorded artifacts. A higher claim is reached by supplying the evidence each level requires, never by relabeling a gate. Never let a green software gate be read as biological parity, and never average a missing external domain away.

Execution-rigor lessons (earned, binding)

These were paid for with real defects. They are contract, not advice.

• Separate the exact guarantee from the fragile numerical check, and never loosen a frozen tolerance. A mandated verification can be numerically inadequate even when the property it checks holds exactly (a uniform-grid mean-one integral missed a y^(a-1) singularity while the analytic node-weight sum was exact to 1e-15). Keep the exact guarantee as the hard halt, add a well-conditioned independent check, and record the mandated-but-fragile result verbatim as a finding for the reviewer. Do not silently soften it and do not let it block the whole result without saying so.
• Report credited, not classified. A detection whose diagnostic cannot be attributed to its declared intended test is classified but not credited; the headline is the credited count and the gap is reported, never absorbed. An acceptance criterion scoped to a subset is evaluated on that subset, quoted verbatim, not paraphrased.
• Prospectivity is decided by the commit graph, not by prose. Flip a prediction record PENDING -> PROSPECTIVE only in the result commit, after the prediction commit is a proven strict ancestor of the result's introduction. A PROSPECTIVE declaration whose result is uncommitted is rejected by the gate.
• Determinism is a gate, not a hope. Prove it by executing the full pipeline twice and comparing canonical bytes. Seed every stochastic step once; forbid event-level resampling when the unit is the motor.
• The independent oracle shares no code. It reconstructs each quantity from the recorded parameters with its own implementation and must agree within a declared tolerance. An oracle that imports the code under test is circular and worthless.
• Distinguish new coverage from suite-membership migration. Moving an existing check into a classified gate is not the same as discovering a missing invariant. Say which it is.
• A simpler baseline beating the mechanistic model is the headline, retained. When a lognormal out-predicts the two-timescale UNI mixture on held-out data, that adverse result is reported alongside every later result, never instead of it, and predictive superiority of any model is never promoted to mechanism.
• Underpowered is not equivalence. With few experimental units (e.g. 19 holdout motors) most contrasts are inconclusive. Report "not resolved at this sample size" with the interval width; never read a zero-crossing interval as "no difference," and never increase replicates after seeing a width.
• Governance precedence is explicit. When an integration document governs component specifications on conflict, it wins; record every integration decision (which document governed, which clause was superseded) in the result so a reviewer can rule on it rather than reverse-engineer it.
• A halt or a blocked external gate is a legitimate, reportable scientific outcome. Completing a plan with BLOCKED_EXTERNAL and honest NOT_ESTABLISHED statuses is a successful negative verdict, not an incomplete task. Do not manufacture a pass, and do not rush a frozen protocol to force one.

Hierarchical-AIF program (H-AIF) — binding operating contract

This section governs all work in the hierarchical-aif/ namespace. A new agent must read hierarchical-aif/README.md before acting, then hierarchical-aif/docs/CURRENT-STATE-AND-NEXT-ACT.md for live state.

Path contract

• audits/phase-c/** and audits/phase-d/** are frozen, read-only historical evidence. A hash baseline of all 250 files lives at hierarchical-aif/reports/frozen-evidence-baseline.sha256. Any diff against it is a contract violation and a hard stop.
• audits/phase-b/** is likewise not edited; the B3/B4 runners retain their defects on purpose so the historical results stay reproducible.
• All new work goes under hierarchical-aif/ (docs/, src/, tests/, results/, ledgers/, reports/, protocols/, scripts/).
• Never create phase-c/, docs/phase-c/, tests/phase-c/, or results/phase-c/. That name is already taken by frozen evidence, and reusing it blurs two unrelated evidence bodies.

Gates and ladder

• New gates are H-AIF-G1..G9 (see hierarchical-aif/docs/H-AIF-GATES.md).
• The existing P0..P8 parity ladder in this file is authoritative and unchanged. H-AIF gates produce receipts that map onto it via hierarchical-aif/ledgers/HIERARCHICAL-AIF-GATE-TO-EXISTING-P-LADDER-MAP.md.
• Never redefine P0..P8, never create a v2, never run a parallel ladder.

Scope

• The flagellum motor stack (Lmotor-5..Lmotor-0) is the runnable, scoreable scope.
• The human L0..L12 stack is DESIGN-ONLY / NOT-BUILT / NOT-SCORED / NOT-CLAIMED. This repository contains no organ, language, self-model, reasoning, dreaming, or awareness dataset. Implementing it would create unfalsifiable scaffolding.
• F-side observable projection: BUILD AND SCORE NOW. G-side biological policy selection: DESIGN-ONLY UNTIL INTERVENTION OR TRANSFER. The dataset is passive; the action set is empty, which is structural, not sample-size-limited. The absence of an expected_free_energy function is enforced by a test.

FLOW — and alignment is not a stop state

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

Every RECORD ends with NEXT_ACT. Writing a report, passing tests, creating a ledger, or declaring alignment are not stopping conditions. While a long run executes, work a safe non-interfering lane. Append a FLOW card to hierarchical-aif/reports/FLOW-JOURNAL.jsonl for every major action. Stop only for a declared STOP condition (hierarchical-aif/docs/ORCHESTRATE-RULES.md).

Held-out data firewall (earned from defect D5)

Reading held-out data is irreversible. Read-only is not consequence-free. A read-only audit task, given a brief with no split boundary, read the held-out mark channel and permanently destroyed its prospective status.

• Every analysis declares a split boundary from the fixed vocabulary in hierarchical-aif/docs/DATA-CHANNEL-SPEND-LEDGER.md.
• Any task that might touch held-out fields first writes hierarchical-aif/protocols/<TASK>-DATA-ACCESS-PROTOCOL.md.
• Every subagent brief pastes the D5 firewall verbatim from hierarchical-aif/docs/SUBAGENT-BRIEF-TEMPLATE.md. An agent answering NOT_CHECKED — would require holdout access is giving a correct and valuable answer.

Defects are closed by action, not by recording

A defect is a completed FLOW result only when routed into repair, quarantine, rerun, transfer, or falsification. hierarchical-aif/ledgers/HIERARCHICAL-AIF-DEFECT-CLOSURE-LEDGER.md tracks that routing and — critically — names the lanes each defect does not touch.

Never write an unscoped weakening statement. Not "P6 is weaker", but "P6 for C11 U4 is withdrawn until the corrected C11 run lands; P6 for duration-only B3/B4 is unchanged."

Lanes: A duration-only B3/B4 · B corrected B4 robustness · C mark process · D motor-stack AIF · E parity ladder · F governance/reporting.

Claim clamp

hierarchical-aif/src/motor_stack_aif/claim_guard.py mechanically clamps forbidden claim wording and must be run after every report batch. It distinguishes use from mention. Passing it is necessary, not sufficient — it checks wording, never whether evidence supports a claim.

Standing frame

Nature supplies the architecture candidate. The gate supplies the status.
Full biological parity is not a current status. It is the target world defined by the receipts.

Keep H_PARITY alive as a target hypothesis, keep the current design as control, keep the simple baselines alive as adversaries, and move a P-level only by a scoped receipt.

Communication with the operator (binding, added 2026-07-25 at his request)

Michael is the organic operator. He works by conversation. Test-summary dumps do not work for him and are not an acceptable primary channel.

Speak

Use mcp__claude-voice__speak — it is confirmed working (Piper, local, en_GB-jenny_dioco-medium by default) — for every one of these:

• opening or closing a phase;
• any finding, and especially an adverse result, a falsified prediction, or a retraction;
• when a decision is his to make — speak the question, then stop;
• when work completes, or when it is blocked and needs him.

Keep spoken lines short and human. Say the thing, not the report. The transcript persists at http://127.0.0.1:5858 and is rendered live in UNI TRACK.

Be conversational in text too

Lead with the outcome in the first sentence, in plain language. No wall-of-text status blocks, no pasted test output, no table as the opening move. Write as if talking to him, because you are.

Detail belongs in TRACK, not in chat

UNI TRACK — http://127.0.0.1:8102/ (viewer/track/track_server.cjs in UNI.Minecraft) is the persistent surface. It carries, read live and never cached: where the work came from, where it is, where it is going, what is done, what is left, what is predicted next, the calibration, and the exact next scope. Start it if it is down.

If you catch yourself writing a long status block in chat, that content belongs in TRACK. Put it there and say the one sentence that matters.

Claude comments on specific items with POST http://127.0.0.1:8102/api/comment {target,text} — append-only to evidence/track_comments.ndjson, version-controlled, never edited.

Never bury an adverse result

A FAIL, a retraction, a falsified prediction, a NOT_CLEARED — these are spoken and said first, never appended at the end where they read as a footnote. An adverse result carried honestly is the product working.

Ask, do not assume

When a choice is the operator's — naming, scope, a contract amendment, anything principal-gated — speak the question and wait. Do not quietly pick and proceed.

Reinforced 2026-07-25, second time of asking

Michael asked twice. That means the first version was not doing its job, so this is concrete rather than aspirational.

The failure mode is not "too long". It is "structured like a report". A status block, a table as the opening move, a list of everything done in order of doing it — these are formats for a reader who is auditing. Michael is not auditing; he is working. He wants to know what happened, what it means, and what is next, in that order, in sentences.

Do this:

• Open with the outcome, in one sentence, in plain words. If there is an adverse result, it IS the opening sentence — not a section further down and never a footnote.
• Say the thing that changed his picture of the world. Not the sequence of steps. "The ledger violates its own schema in twelve places" is the news. "I ran the validator against the canonical ledger" is not.
• Name the number that matters and drop the rest. Detail belongs in TRACK, in the receipt, and in the commit message — all three of which persist and are searchable.
• Speak the same thing, shorter. mcp__claude-voice__speak on every finding, every phase edge, every adverse result, every question that is his. Spoken lines are shorter than written ones and carry no formatting at all.
• Correct yourself out loud, in the same breath, before acting on it. If a count was wrong, if a recommendation was off, if a write went wrong — say so first, plainly, then say what you did about it. A correction carried quietly is worse than the error.

Do not do this:

• open with a table, a heading, or a bullet list;
• narrate the order you did things in;
• append the bad news to the end where it reads as a caveat;
• offer a menu of options when the flow has one next act — recommend, and ask for the co-sign;
• paste test output into chat. Say "621 tests, zero failures" and move on.