THE HANDS · conductor plan · r1 — for red-pen

The Baseline Sweep

Measurement only: per-class truth today — works, works-slow, refuses, lies — with stage clocks, on scratch copies of the flagship. Dispatches only after the census is ratified.

RATIFIED 2026-08-14 · gate satisfied · L1 dispatched (engine deviation named in Sequence)
01

Objective — measurement only

This sweep changes no trunk, cures nothing, merges nothing. It banks the arc’s denominator: for every ratified class, what is actually true today — works · works-slow · refuses · lies — with stage-attributed timings, on scratch copies of the flagship book at the pinned trunk. It dispatches only after the census is ratified.

Companion: the class-universe census · instrument facts cite the banked receipts (lifecycle & rails especially). The orchestration charter binds in full.

02

The four verdicts — defined so they can fail

worksEnd-to-end
Correct target, scope-complete diff, truthful record (claimed spans equal measured byte deltas, zero bytes outside), undo restores byte-identical, under 60 seconds.
works-slowOver the clock
All of the above, over 60 seconds — and the stage attribution names the slow layer.
refusesStops
Every refusal binned against the recorded shapes (physics · route · rails). A refusal outside the bins — or a WALL: “found nothing” where a recorded reason exists — is its own finding class.
liesRecord vs. reality
A record disagreeing with a measurement: claimed spans the bytes deny, green over wrong, done over partial, undo not byte-identical, stale verdict cited fresh. Lies outrank everything — the sweep’s most valuable harvest.
03

Discipline — the laws, applied

  1. Judgment never on-box. Legs run mechanics and bank RECORDS — outcomes, hashes, timings, transcripts. No leg classifies works/lies; classification is the home reader’s act, recomputable by any non-author from the banked records.
  2. Isolation. One leg, one scratch clone, one custody unit; never two mutating probes in one clone; disk pre-check before dispatch; a disposal line flows the moment records bank. Clones come from my pinned copy — never the keeper’s checkout. Claims filed per clone before dispatch.
  3. Measurement only. No leg pushes a branch; nothing lands on any trunk; agents never delete — clone disposal is the keeper’s hands.
  4. Sealed-gauntlet independence. Probe asks are measurement instruments authored from the census classes — not rehearsal asks, never the sealed twelve.
  5. Receipts, never summaries. Each leg delivers a records bundle at a named path plus a mechanical index; home verdicts cite record paths.
The behavior dimensions are lenses, never rubrics (inherited from the census — Robert’s red-pen, 2026-08-14, binding). The five dimensions are applied by the home reader to banked records, after the fact, to grade the system’s delivered experience. No leg builds, prompts, or wires any runtime check that grades or constrains an agent’s judgment mid-ask from these dimensions — no code rubric, no prompt-injected behavior gate. Code holds facts and refusal-physics; the environment — exits, bounds, receipts — is the legible part; intelligence stays free inside it.
04

The probe matrix — three families, ~90 records

P1
Resolution probes — read-only, widest
The wish door has no dry-run, so P1 calls the resolution layer directly (exactly how the ancestor census probed) for every class member with a real value from the book. Banks the decision per member. Re-verifies all 58 ancestor rows AT TRUNK before anything mutates.
P2
Execution probes — scratch clones, landed classes
One-to-three asks per landed class through the class’s REAL lane (door where admitted, terminal otherwise — the lane is part of the record). Banks: exits + slots, ledger rows, witnesses, tree hashes before/after, the undo and its byte-identity result, and the run’s own timing rows — the recorder is wired but has never fired on this book; our probes write its first rows on their clones. The driver also stamps wall-clock around preview/publish and records how many pages reprinted — incremental-vs-full reprint is THE clock question. No new harness is built; the instruments exist.
P3
Behavior probes — the five dimensions
Two ambiguous asks (must return questions carrying the neighborhood) · two bounded asks (must voice the recorded reason) · one wall ask at C18 (expected “found nothing” today — banked as the wall’s baseline) · one multi-part ask (scoreboard completeness) · consent choreography on the booking-link class, on a clone, zero world effect. The clock rides every P2 probe.
05

Legs and engines — every leg names its engine and fit

L1 · P1 probes
deepseek (sandbox lane)
Bulk mechanical: a fixed probe harness over ~58 inputs, zero judgment — the cheapest honest hands. Runs first and alone; its output locks the ask cards.
L2a-d · P2 executions
grok-build · bay-1
Build workhorse; shell-heavy end-to-end driving of real verbs on a real clone — one leg per class group (words · media · structure-landed · SEO), one scratch clone each.
L3 · P3 behavior
grok-build · bay-1
Same shape as L2; transcripts banked verbatim for the home reader.
Classification
fable · at home
The meaning seat — judgment never on-box. Every verdict cites record paths a non-author can recompute.
Red-pen
keeper
Robert’s proxy; the plan round-trip before any dispatch, and the denominator review after.
Parallelism is bounded by verification: at most 3 mutating legs live at once — the width one home reader can genuinely classify same-day without thinning verification.
06

Sequence

  1. Gate: census ratified. Robert’s eyeball via the keeper. Any class-list change reshapes the matrix before anything runs.
  2. L1, then lock the ask cards. Half a day, sandbox, read-only.
  3. L2a-d + L3, at most 3 live. Each with a claim row, a disk pre-check, and a disposal line on bank.
  4. Home classification → the denominator. One row per probe: class · ask · lane · verdict · stage clocks · basis paths. Keeper red-pen, then Robert, with the burn-down proposal: worst honest class first.
07

Escalation triggers

Any write outside a clone
Stop the lane, report immediately.
A lie in the trust rails themselves
Witness/receipt/round-trip machinery lying → keeper immediately, before classification finishes — frame territory, not a row.
A stage that resists attribution twice
Two honest rounds failing to name a latency stage → frame question to the keeper, never a third patch.
Teammate-facing words
Anything requiring one → keeper. Held, per charter.
08

Open items for the red-pen

1
Probe-ask authorship.
This plan has the orchestrator authoring probe asks (measurement instruments). Stricter separation — a non-builder authoring them from the census table alone — costs a day, buys independence. Your call on where the line sits for a baseline.
2
Breadth.
Representative (~24 executions) now, exhaustive (~58) only where results disagree with the ancestor — or exhaustive from the start.
3
Consent probes on clones.
Exercising the full consent choreography (booking-link class) on a scratch clone mints consent-record fixtures that never leave the clone. Confirm comfort.