Jury — panel judgement for contested visual decisions

Use this reference when one critic is not enough and averaging several critics would be worse than either. A jury ranks materially different directions and widens defect coverage. It does not produce a more confident number.

Muriel convenes the panel. She is not on it. SKILL.md makes her a visual collaborator, not a detached critic; the jury is a body she summons, and its finding is evidence she carries back into the build. The calling agent still owns product truth.

Contents

1. Load gate
2. Jurisdiction
3. Two bodies, two aggregation rules
4. Seat definition
5. The roster
6. Ballot mechanics
7. The sealed round
8. Splits are the product
9. Delphi round
10. The ledger
11. Unanimity smell test
12. The chair
13. Worked example
14. Naming decisions
15. Anti-patterns
16. Prior art

Load gate

Load when all three hold. Contested — the question survives after the deterministic checks ran and their results were stipulated. Consequential — the answer changes implementation, not decoration; Compare-mode direction selection is the clearest earn. Committed — the artifact enters a repo, paper, published page, or shipped surface this session.

Cost is per seat per option, not per panel. Budget one critique per seat per option, plus that seat's lens cost: the Squinter renders a four-level blur ladder per option, the Thumbnail reads two downscales per option, the Loupe reads up to 16 magnified tiles per option, the Forger builds a forgery per option it audits. A three-seat panel on one artifact is roughly 4-6× a single critique. The five-seat, three-option panel in the worked example below is 15-25×. Scale the gate to that number, not to a flat 5×.

Do not load for: a test render or intermediate frame; anything a compute tool answers; a settled direction needing only execution notes; backend, body copy, or spacing work; a decision the calling agent has made and only wants ratified. A name is the one non-visual decision in scope; see Naming decisions. A single artifact with no alternatives is out of scope only when its premise is settled — run muriel-critique alone. A single artifact whose premise is contested is in scope, and votes on the single-artifact ballot defined under Ballot mechanics. When in doubt, run one critic: one with a clean ledger beats five with correlated priors.

Jurisdiction

The jury may not rule on facts. Findings of fact come from deterministic tooling and enter as stipulated evidence, never as motions:

• python -m muriel.contrast <file.svg> — WCAG ratios per text pair;
• python -m muriel.oklch <color> — sRGB / OKLCH / in-gamut;
• python -m muriel.devibe <path> — AI-default design tells by regex over source: untouched shadcn/Tailwind defaults, AI-purple as primary, gradient headings, purple-to-blue gradients, neon glow, emoji-as-icons, the centred hero + three-feature-card skeleton;
• recorded seeds, dimensions, and export paths;
• conditional, may return nothing — python -m muriel.tools.impeccable_bridge <path> is a no-op when Node, npx, or the third-party impeccable CLI is absent, and degrades silently; the polish-rule regex checks (hit-area minimum, transition: all, will-change: all, missing tabular-nums, scale < 0.95 press) are not built — they are a queued TODO item. A source that produced no output is not a stipulated clean bill. Record it as unavailable, and the question stays in jury jurisdiction rather than being struck.

"Is this measurable?" depends on the substrate. devibe reads HTML, CSS, JSX, and SVG source and is decisive there; it is silent on a raster, a PDF, or a screenshot, where there is no source to scan. Scope the Forger accordingly: when the artifact has scannable source, devibe's tell list is stipulated fact and a Forger ballot claiming any of it is struck as noise; the Forger's jurisdiction is what survives the scan — whether the composition is reproducible from a generic template, which no regex answers. When the artifact is a raster, the whole tell list returns to the Forger.

A juror may not vote a measured contrast failure down to "fine, it's decorative." The 8:1 floor on readable text is on no ballot, in no round, under no severity discount. Reproducibility is the same: a missing seed is a fact, and facts are repaired, not debated.

Jury jurisdiction is only the genuinely contested: composition and focal dominance; hierarchy and reading order; which of N directions earns implementation; brand voice; whether an encoding earns its complexity; whether the framing itself is wrong. The chair's first duty is kicking measurable questions back to the compute tools. A ballot claim a tool could have settled is struck, the tool is re-run, and the strike is logged as noise.

Two bodies, two aggregation rules

One panel, two jobs, two rules. Never mix them.

Body	Question	Aggregation
Defect finding	What is wrong here?	Union, max severity, no vote
Direction selection	Which earns build?	Comparative, split reported

Defect finding — worst critic wins. Union every seat's findings; where two name the same defect, keep the higher severity. Nobody votes a defect away, because the loss is asymmetric: a false positive costs one wasted fix, a false negative ships. Severity vocabulary is CRITICAL | HIGH | MEDIUM | LOW, identical to agents/muriel-critique.md, and the rollup is unchanged — any CRITICAL → FAIL, any HIGH → NEEDS REVISION, MEDIUM/LOW only → PASS.

Direction selection — the jury proper. There is no "worst" when ranking viable options, so worst-case aggregation is meaningless. Use comparative judgement, report the ranking with its scale values, report the split. A seat may serve both bodies in one sitting; the chair routes each half of its ballot to its own rule.

A CRITICAL does not eliminate an option from the ranking. This is the one place the bodies touch, so the rule is normative here and not inferable from an example. A CRITICAL sets that option's verdict to FAIL, which gates committing it; the comparative ranking is computed over every option that was seated, including that one. The winner ships only after its CRITICALs are repaired and the re-render confirms the repair.

The single exception: a CRITICAL is form-fatal when its fix cannot be applied without turning the option into a different option — the fix changes the encoding, not the execution. A form-fatal CRITICAL eliminates the option; the chair drops it and recomputes the ranking over the remainder, and records which finding was form-fatal.

Why this rule and not elimination-on-any-CRITICAL. Elimination lets one seat veto the comparative body: a single repairable defect on the Bradley-Terry winner would hand the build to the runner-up, which is a defect-finding seat deciding direction selection. The two bodies answer different questions — "what is wrong here" and "which form earns build" — and a repairable defect is evidence about the first, not the second. The form-fatal carve-out is where the defect genuinely is evidence about the form, and it is narrow on purpose: the chair must be able to name the encoding change the fix requires.

In a naming sitting the option is the string, so a Forger CRITICAL is form-fatal every time and the carve-out stops being narrow. There the chair reports the ranking with and without the elimination instead; see Rule (b): no Forger veto.

A jury and muriel-critique are mutually exclusive. Run one or the other on an artifact, never both. muriel-critique is a generalist whose coverage overlaps four of the six seats — hierarchy (Squinter), the numeric-claim probe list (Pedant, which this reference delegates back to it), AI-tell (Forger), and multi-scale evaluation (Thumbnail). Under union aggregation, running both double-counts by construction: the same defect arrives twice from correlated estimators and looks like corroboration.

Seat definition

N copies of one critic is not a jury. Same base model, same priors, correlated errors — you paid per seat per option for noise reduction on a biased estimator. Prompt flavor ("harsh critic", "warm critic") changes tone, not the loss function, and decorrelates nothing. A seat is three things, all stated when it is defined:

• Lens — the transformation applied before the seat sees the artifact. The real source of diversity: it changes the input, not the prompt.
• Loss function — what it optimizes against, and so what it is willing to be wrong about.
• Evidence access — what it may see, including what it is denied. A seat that read the brief and one that did not are different estimators.

Two seats sharing a lens and a loss function are one seat. Cut one.

The roster

Six seats, all six shipped as agents. Seat them as subagents. Running one inline as a scripted prompt is a weaker thing than seating it — say so on the ballot rather than pretending otherwise.

Evidence denial requires context isolation, and only a subagent provides it. A scripted prompt runs inside the chair's context. Anything the chair has read — the brief, the option set, an unsealed ballot — is available to the model producing that "seat's" output whether or not the prompt mentions it. The denial is nominal: it is an instruction not to use evidence, not an absence of evidence. Every ballot therefore carries Isolation, and an inline seat may not be the sole source of a CRITICAL. Its findings still enter the union at HIGH or below; promoting one to CRITICAL needs a second seat naming the same defect.

The Stranger may never be scripted inline. It is the only seat defined by denial of the brief. Run inside a context that has already read the brief, it is the chair's own opinion wearing a persona — the exact failure the seat exists to prevent, and worse than not seating it, because the ballot claims an independence it does not have. It ships as a subagent and enforces this at its own boundary: it audits its invocation for briefing material and aborts rather than filing a contaminated ballot. If it aborts, reseat it cleanly or record the Stranger as unseated. Do not argue with the abort.

The Squinter — SHIPPED as an agent. Lens: heavy Gaussian blur (sigma ≈ 2% of the long edge), plus a luminance-only pass. Loss: misses in hierarchy; willing to be wrong about detail it can no longer see. Access: blurred renders only; denied the sharp artifact, all text, the brief. Judges: focal dominance, reading order, figure/ground, whether the composition has one center of mass or three. The actual test for hierarchy, not a proxy, and the blur is deterministic, so the lens replays.

The Thumbnail — SHIPPED as an agent. Lens: the artifact at 1/8 scale, at 16 px, and at 16 px zoomed 8× for inspection — the three downscale companions muriel.squint already writes on every run, so the lens replays. Loss: misses in scale survival. Access: downscaled renders plus the intended use sizes; denied the full-size render and the blur ladder. Judges: whether the primary signal survives an OG preview, favicon, slide thumbnail, or paper figure at column width.

It is not a smaller Squinter, and the distinction is what keeps the two seats from correlating. Blur preserves position and area while removing high spatial frequency, so a large pale mass survives heavy intact. Downscale destroys sub-pixel features outright — hairlines, thin strokes, small type — and aliasing invents structure that was never in the source. An artifact can pass the hierarchy test cleanly and die at 16 px, and the reverse.

The Loupe — SHIPPED as an agent. Lens: magnification. muriel.loupe cuts fixed 256 px cells, in the pixels the artifact ships at, with a 32 px overlap, ranks them by edge density, and writes the top 16 at 4×. It re-renders an SVG at 4×, crops an @4x capture without resampling, and magnifies a 1× raster nearest-neighbour and records that it did. Loss: misses in local finish; willing to be wrong about everything that needs the whole frame. Access: the tiles and the manifest's coordinates and coverage; denied the full-size render, the blur ladder, the downscales, any overview or grid map, and the brief. Judges: line ends that land, joints that meet, fills that abut without a hairline gap, text that is not clipped, crossings drawn in the right order, and repeated patterns that end on an edge instead of stopping dead. Empty space is never a finding: whitespace is a decision, and the lens does not write an empty cell.

The cell is fixed rather than scaled to the frame, which is the opposite of the Squinter's ladder and on purpose. Hierarchy is relative, so the blur scales with the long edge. Local finish is not: a 1 px gap is the same gap on a sprite and a poster, and a proportional cell would make a poster's loupe a 1× view. The tile is 1024 px because a vision model downsamples larger inputs and would undo the magnification.

The Loupe exists because contour completion hides these defects at viewing scale. A rule that stops 2 px short of an outline reads as touching it, and a reviewer who sees the whole composition fills in what the joint is meant to be. A seat denied the composition cannot. It is not the Thumbnail inverted: the Thumbnail asks whether the signal survives shrinking, the Loupe whether the execution survives enlarging, and an artifact can pass one and fail the other. The practice comes from the anatomy isometric-figure skill, whose rejected figures were mostly faults found by zooming in on joints that had passed at 1×; see Prior art.

The Stranger — SHIPPED as an agent; never seat it inline. Lens: a single Read of the artifact, no re-read, followed by three questions answered from that one pass — what is this for, what is the one number, where do I click. Loss: misses in premise; willing to be wrong about execution entirely. Access: one pass over the artifact; the brief is withheld from the invocation, as are the option set, every other ballot, and project context. The only seat that can catch that the panel is grading the wrong thing well, because every other seat is anchored to the brief. Fails on premise, not execution.

The earlier form of this lens — "5-second exposure, then the artifact is withdrawn" — is unexecutable in any harness and was replaced. You cannot withdraw an image from a model's context, and a model has no exposure duration; a prompt saying "you saw this for five seconds" is a costume. Single-read, no-re-read, brief-withheld is the executable version of the same denial, and it is enforceable only across a subagent boundary.

What the Stranger actually measures. It is not a naive human. It shares a base model with every other seat, so isolation removes brief anchoring and does nothing about model priors, training-set familiarity, or genre knowledge. Its ballot reports what this model concludes from the artifact alone — a real and otherwise-unavailable signal, and not the same as first-time human comprehension. Its ballot therefore carries a mandatory Prior strength line: high where the model has strong genre priors (a bar chart, a login form), low where the artifact is novel enough that the reading approaches first contact. Weight the seat by that line. Do not let a quiet Stranger on a familiar genre be read as evidence that real strangers will understand it.

The Forger — SHIPPED as an agent. Lens: none on input; the lens is on output. The seat asserts "I can reproduce this from a generic template in four minutes," then actually attempts it and reports how close it got. Loss: misses in distinctiveness. Access: the artifact and a generic component library; denied the brand tokens, so it cannot cheat by copying them. Converts the anti-slop check from rubric to proof obligation: "this looks generic" is unfalsifiable, "here is the forgery, 4 minutes, differs only in accent hue" is not. Reports: elapsed time, forgery path, and a named list of what it could not reproduce. An empty list is CRITICAL.

The Pedant — SHIPPED as an agent. Lens: text extraction only — labels, legend entries, ticks, captions, units, callouts, flattened with source locations. Loss: misses in labeling. Access: extracted text plus any sibling data file; denied the rendered composition, so a handsome layout cannot charm it. Judges: unlabeled numbers, missing units, terminology drift between legend and caption; applies the numeric-claim probe list from agents/muriel-critique.md exhaustively. Unglamorous, highest hit rate of the first five seats — most shipped defects are a missing unit, not a broken layout.

Its denial depends on substrate, and the ballot must say which it got. On source artifacts — SVG, HTML, Markdown, LaTeX, a plotting script — the denial is real: the seat greps text nodes and never reads the render. On rasters there is no OCR path that returns strings (muriel.detectors.text returns bounding boxes only and is not a subcommand), so either the chair supplies extracted text and the denial holds, or the seat reads the raster itself and records Isolation: subagent (denial partial — raster read). Its findings stay valid either way; the claim of compositional blindness does not.

Note also that its probe list comes from muriel-critique, which is why the two bodies may not both run — see Two bodies, two aggregation rules.

Ballot mechanics

Randomize option order per juror. LLM judges have strong position bias; first and last presented options are systematically advantaged. Cheapest available fix, and where most naive panels leak. Derive the permutation from a named digest, per seat, so the panel replays:

printf '%s' "<decisionId><seatId>" | shasum

Read the hex digest left to right, two characters at a time, and use each byte to pick the next option out of the remaining list (byte % remaining). Record the result as Order seen. A ballot without Order seen is invalid. Do not substitute a language-level hash — Python's hash() is salted per process, so it does not replay across sessions, which is the entire point.

Comparative judgement over absolute scoring whenever there are options. Absolute LLM scores are range-restricted — nearly everything lands 6-8/10 — and badly calibrated across sessions. Pairwise forced choice builds a scale out of binary decisions and yields an inter-rater reliability statistic for free. Run the full pairwise set when N is 4 or fewer (6 comparisons per seat at most); sample above that. Fit Bradley-Terry over the pooled comparisons for the ranking.

Two statistics, both raw proportions the chair computes by counting. The chair is an LLM with no fitting library; anything it cannot get by tallying the ballots will not be computed. Bradley-Terry strengths are reported for the ranking, but neither threshold below is defined on them, because a strength means three different things under sum-to-one, log-odds, and probability normalization.

• Separation — P(top1 > top2), from the pooled pairwise tally: count the comparisons between the top two options across all seats, divide the wins for the higher-ranked option by the number of non-tied comparisons between them. 0.50 is a coin flip; 1.00 is unanimous.
• Concordance — the fraction of agreeing seat-pair judgements. For every unordered pair of seats and every option pair, score 1 when both seats gave the same strict judgement or both tied, 0.5 when exactly one of them tied, and 0 when they chose opposite options. Divide by (seat pairs × option pairs). This is the reliability figure.

Reasoning before the vote. The ballot emits reasoning above its comparisons, ranking, and ship decision. A vote written first anchors the rationale that follows, and you get justification, not judgement.

Single artifact, no alternatives. In scope when the artifact's premise is contested; see Load gate. Do not fake a comparison. Require a forced rank of the issues — no ties, which stops a seat calling everything equally critical — plus one binary ship call scoped to that seat's own loss function. Binary decisions carry more than five absolute scores clustered at 7. Ship (this seat's loss only) is not a panel ship decision and never aggregates into one: the Squinter answering no means hierarchy fails, not that the artifact is blocked. The chair reports the spread of these binaries as part of the split; it does not tally them into a verdict.

Every juror agent emits exactly this block and nothing outside it:

Juror ballot
Seat: <squinter | thumbnail | stranger | forger | pedant>
Decision ID: <stable task-local id>
Isolation: subagent | inline (denial nominal)
Order seen: <option ids, in the order presented>
Evidence seen: <what this seat was given>
Evidence denied: <what this seat was refused>
Reasoning: <prose; must precede Pairwise, Ranking, Issue rank, Ship>
Pairwise: <A>B, C>A, B~C>          (omit if no alternatives)
Ranking: <option ids, best first>  (omit if no alternatives)
Findings:
  - id: <seat-local id>
    target: <option id or "artifact">
    rule: <named rule or framework>
    severity: CRITICAL | HIGH | MEDIUM | LOW
    evidence: <concrete: where, what, measured or observed>
    fix: <one sentence, actionable>
    form_fatal: yes | no    (required on CRITICAL; see Two bodies)
Issue rank: <finding ids, worst first, no ties>  (single-artifact only)
Ship (this seat's loss only): yes | no           (single-artifact only)
Confidence: low | medium | high
Recused: <criteria this seat did not judge>

Recused is mandatory and non-empty for any seat whose access denies it something. The Squinter recuses from typography; the Pedant from composition. A seat that recuses from nothing is a generalist, and it will correlate with every other generalist. Isolation is mandatory on every ballot: subagent only when the seat ran in its own context, inline (denial nominal) otherwise. A seat that reports subagent while running as a scripted prompt in the chair's context is the one lie this format cannot survive.

The sealed round

Round 1 ballots are sealed. Jurors do not see each other's ballots, a running tally, or the panel size. Break this and the panel measures conformity, and round 2 has no baseline. The chair unseals all ballots at once, after the last lands, and never summarizes round 1 to a juror still voting.

Splits are the product

Averaging jurors destroys the only thing a panel uniquely produces. A mean of five rankings is a number; the disagreement behind it is a design finding. Where the panel splits, that split is the unsettled decision — which is muriel's whole remit. Report it, name what each side optimizes for, hand it up:

Reader and implementer split on the disclosure control. Reader optimizes discoverability, wants always-visible. Implementer optimizes density, wants hover. This is a product call, not a visual one.

That returns cleanly to muriel's boundary: the calling agent owns product truth. A jury resolving this by fiat oversteps the skill's contract. Report a split when any of these six hold:

1. Two seats rank the top option differently.
2. One seat's top option is another's bottom option.
3. Separation between the top two is under 0.65 — P(top1 > top2) from the pooled pairwise tally, as defined under Ballot mechanics.
4. One seat's Ship (this seat's loss only): no contradicts a majority yes.
5. Two seats' fixes on the same target are mutually exclusive — applying one makes the other impossible.
6. Naming sittings only: the two rankings disagree. The ranking over every seated option and the ranking after form-fatal eliminations put different options first. See Rule (b): no Forger veto.

The split report names, per side: the seats, the favored option, the loss function driving it, and the product question that would resolve it. It names no winner.

Trigger 5 is how the panel acquits a correct form. Union aggregation has no acquittal by itself: every finding survives, so a form both seats attack from opposite directions collects both attacks and fails. Concrete case — a Tufte small-multiples grid. The Squinter fires HIGH, "no mass dominates," and its fix is to break the grid by scaling tiles. The Forger fires HIGH, "equal-weight tiles are counterfeit stock," and its fix is to make the tiles more distinctive per-tile. Meanwhile the form is correct: uniform tiles carrying one variable each is what small multiples are, and the equality is the encoding.

Under union-max both findings ship, the rollup reads NEEDS REVISION, and a chair forbidden a preference cannot say the form is right. Under trigger 5 the chair checks the fixes before the union, finds them mutually exclusive, and routes both findings to the split report instead of the defect union. The output becomes: two seats attack this form from opposite sides, one wants less uniformity and one wants more, and neither fix can be applied without defeating the other. That is a legible acquittal — it says the form is answering to a constraint neither seat holds — and the chair reaches it by a mechanical test, not by holding an opinion.

Route to the split, do not merge, and do not let either finding also enter the union. A finding routed to a split is recorded in the ledger with the disposition it later earns, exactly as a union finding is.

Delphi round

Optional. Run one only when the split blocks a commit that must happen this session. Unseal round 1, show every juror the full split — all ballots, all reasoning — allow revision, and record who moved. A revised ballot carries two extra lines: Moved: yes | no and Moved because: <the specific evidence in another ballot that changed it>.

Movement toward the most confident or most verbose juror is a failure mode, not consensus. If a seat moves and cannot name specific evidence in another ballot, the chair discards the revision and keeps round 1. Three or more seats moving in one round is a cascade: report round 1 as the finding and record the cascade. Convergence is not the goal — a split that survives round 2 with both sides stating their loss functions more sharply beats a unanimous round 2.

The ledger

Severity inflation is the failure mode that kills harsh-critic systems. If brutality is the reward, everything becomes CRITICAL, the scale collapses, and under any-CRITICAL → FAIL the panel fails everything and carries zero bits. Harshness is also cheap and usually unfalsifiable — "this lacks conviction" can never be wrong. The antidote is scoring each seat against what happened next.

Disposition vocabulary — three resolved values plus one open. hit: fixed and the re-render confirms it. miss: the human dismissed it with a stated reason. noise: nobody could act on it, or the chair struck it for straying into findings of fact. open: unresolved, excluded from all rates.

The transition point is the re-render, and it has an owner. Every finding is emitted open. It leaves open at exactly one moment: when the artifact is re-rendered after the fix pass — the Muriel delta. The agent owning that fix pass writes the disposition, one per finding, at that re-render: hit if the re-render shows the defect gone, miss if the human declined the fix and said why, open if the finding was neither fixed nor declined. There is one exception with a different owner and an earlier moment: the chair writes noise at strike time, during the sitting, because a struck claim needs no re-render to be resolved. Without a named moment and a named owner, everything stays open forever, open is excluded from every rate, and the trailing-20 threshold below can never trip even in principle.

Record shape. One record per finding, versioned, emitted beside the existing muriel.design trace and sharing its decisionId:

{"kind":"muriel.jury.finding","version":1,"decisionId":"…",
 "seat":"squinter","findingId":"…","rule":"…","target":"…",
 "claimedSeverity":"CRITICAL","effectiveSeverity":"HIGH",
 "disposition":"hit|miss|noise|open","dispositionReason":"…",
 "proof":"…","recordedAt":"…"}

claimedSeverity is recorded verbatim so any discount is auditable. proof is the render, test, or capture that confirms a hit. dispositionReason is required for miss and noise and must be the human's stated reason, not the chair's paraphrase.

Where it lands. A schema, not infrastructure. Do not build storage for it. It is the finding-level child of the TODO item "Effectiveness observability — make the Muriel delta measurable", and it maps onto Session Cartographer's proposed generic skill_outcome event: kind → event type, decisionId → correlation key, disposition → outcome, proof → verification reference. Persist only where an authorized event-log sink already exists, on the same terms as design-synthesis.md's trace: no telemetry dependency added merely to emit a record, no reasoning exposed.

Scoring rule, over a seat's trailing 20 resolved findings. miss does not count against a seat:

precision       = hit  / (hit + noise)
noise_rate      = noise / (hit + miss + noise)
dismissal_rate  = miss  / (hit + miss + noise)     reported, never penalised

The earlier form put miss in the precision denominator, which made a seat's standing a function of human agreement: "the human dismissed it with a stated reason" scored identically to "nobody could act on it." Those are different events. A correct-but-unwelcome finding is the first and scored as the second, so the seat most exposed was the Stranger, whose entire job is to produce the finding a team dismisses — the panel would have unseated its only check on grading the wrong thing well. noise is the actual slop signal: nobody could act on it, or it belonged to a tool. Score that; record miss and report dismissal_rate beside precision so a human can see a seat generating dismissals without the system acting on it.

The tradeoff, stated plainly: excluding miss means a seat that produces many dismissed findings pays nothing for the reviewer's time it consumes, and dismissal_rate is visible but toothless. That is the price of not letting the panel converge on agreeable findings, and it is the cheaper error.

Pre-ledger defense — one CRITICAL per seat per artifact. Nothing writes the ledger yet, so no seat has 20 resolved findings, so the discount table's first row applies to every sitting and no discount engages. Until a sink exists, the cap is the whole defense against severity inflation: a seat may claim at most one CRITICAL per artifact, and it must be the finding it ranks first. A second CRITICAL from the same seat is recorded at HIGH with claimedSeverity: CRITICAL preserved, so the ledger sees the inflation when it comes online. This is crude and deliberately so — it forces the seat to spend its one veto on its best finding, which is exactly the discipline the ledger buys once it can measure. Inline-seated seats are additionally barred from being the sole source of a CRITICAL; see The roster.

Discount rule, applied by the chair when setting effectiveSeverity:

Condition	Effect
fewer than 20 resolved	face value, provisional; CRITICAL cap in force
noise_rate > 0.30	effectiveSeverity capped at HIGH
noise_rate > 0.50	effectiveSeverity capped at MEDIUM
precision < 0.25, 20+ resolved	unseated: recorded, never aggregated

A capped seat can no longer trigger FAIL on its own; an unseated seat is reinstated only manually. This is what makes "worst critic wins" principled rather than merely loud: standing goes to the seat with the highest precision at an acceptable noise_rate, and a seat that shouts gets quieter automatically. Recall is deliberately not part of standing — the record shape has no false-negative field and the disposition enum has no value for a defect a seat failed to raise, so a recall figure would be uncomputable from what this schema stores. Findings of fact are exempt — they were never the jury's to claim.

Unanimity smell test

If the panel agrees on everything, it is one juror wearing hats. Cut to one critic and save the tokens. Check both after unsealing round 1:

• Concordance at or above 0.90, computed as defined under Ballot mechanics.
• Unique contribution — fewer than half the seats, rounded up, produced a finding or comparison judgement no other seat produced. That is 2 of 3, 2 of 4, 3 of 5, 3 of 6. The old absolute "fewer than 2 of N" was far stricter at N=3 than at N=5, so a three-seat panel could pass a test a five-seat panel failed on the same proportion of unique contributors.

If either trips, record panelCollapsed: true with the tripped condition, ship the strongest single ballot as the finding, and do not convene that seat combination again for that artifact class until a seat's lens or evidence access changes. Adding a personality is not a change.

"Strongest" is mechanical, because the chair holds no preference. Rank the ballots by: (1) count of findings no other seat produced, most first; (2) fewest claims struck as noise this sitting; (3) lowest noise_rate in the ledger, seats without 20 resolved findings ranking last. Ship the top ballot. If two ballots tie through all three, ship both unmerged and say they tied — that is still not a preference.

The chair

The chair is not a juror. It casts no ballot, breaks no tie by preference, and contributes no findings of its own. Duties, in order:

1. Fix the question. One sentence the panel can answer. If it contains a measurable term, stop and run the tool.
2. Run the deterministic checks, enter them as stipulated facts, and repair fact-level failures before seating anyone.
3. Seat the panel. Only seats whose lenses and losses differ for this artifact. Three is a real panel; five is the ceiling. The roster has six, so a full panel leaves one out: drop the Loupe for a social card no one will zoom into, or the Thumbnail for a figure that is never shown small. Do not seat the Stranger inline.
4. Assign evidence access per seat, denials included, and enforce it. Record each seat's Isolation honestly.
5. Randomize option order per seat with the shasum digest; record the permutations.
6. Enforce jurisdiction. Strike claims belonging to the compute tools; log each strike as noise against that seat, at strike time.
7. Route mutually exclusive fixes to the split before aggregating, so the union never stacks contradictory fixes on one target.
8. Unseal, aggregate by the correct rule per body. Apply the one-CRITICAL cap and the inline-seat CRITICAL restriction, then the ledger discounts. Ranking is computed over every seated option; a CRITICAL gates the commit and only a form-fatal CRITICAL eliminates. In a naming sitting, report the ranking both with and without form-fatal eliminations.
9. Run the unanimity smell test.
10. Write the finding, with the split left unresolved, and emit one ledger record per finding at disposition: open.

The chair may not resolve a split. It reports the split and names the product question. That is the finding.

Worked example

Compare mode. Three directions for a per-AOI dwell readout in a gaze-analysis app, about to be committed to the canonical panel component. Chair's question: which dwell-readout direction earns implementation? A — small-multiples strip, one mini heatmap per AOI, sorted by dwell. B — ranked horizontal bars, dwell in ms, AOI labels inline. C — annotated stimulus overlay, dwell numerals on the stimulus. Findings of fact, stipulated and not votable:

dwell-a.svg → tick labels 4.6:1 on panel fill. CRITICAL. Repaired
  pre-seating (label → cream); re-run 12.1:1.
dwell-b.svg → minimum text pair 11.2:1. Clean.
dwell-c.svg → numerals 7.4:1 over the darkest sampled stimulus region.
  CRITICAL. Repaired pre-seating (opaque plate); re-run 12.9:1.
All three at 1400×900, seed 7, exported to render_assets/dwell/.

None of that entered a ballot; two directions were repaired, not debated. Panel: all five seats, each as a subagent. The pedant read SVG source, so its denial is real and its Isolation is plain subagent; on a PNG export it would have carried subagent (denial partial — raster read) instead.

The stranger was invoked with three artifact paths and a decision id, and nothing else. It did not receive the brief above, the question being decided, or the fact that the three artifacts are alternatives to one another.

Round 1 sealed; ballots abridged to load-bearing lines:

squinter   order B,A,C   pairwise B>A, B>C, C>A   ranking B,C,A   high
  Blurred, B is one descending wedge: a single unambiguous reading order.
  A is nine equal-weight blobs, no focal point survives. C is the stimulus.
  A / "one dominant focal point" / HIGH — nine tiles at equal area and
  luminance; blur leaves no rank. Fix: scale tile area by dwell.
  Recused: typography, labeling, units

thumbnail  order C,B,A   pairwise B>C, B>A, A~C   ranking B,C,A   medium
  At 1/8 scale B keeps bar length; labels vanish but rank survives. A
  becomes noise. C's numerals are gone by 1/4 scale.
  C / "evaluate at multiple implied scales" / MEDIUM — panel ships in a
  320px rail. Fix: add a rank strip.   Recused: full-size composition

stranger   order A,C,B   pairwise C>B, C>A, B>A   ranking C,B,A   high
  (brief denied; one Read per option, no re-read) On C I can say what it is
  for — where people looked on this screen — and point at the region. On B I
  read a ranking but could not say what the rows are; "AOI_03" meant nothing.
  B / "premise" / HIGH — identifiers opaque to anyone who did not define
  the AOIs; ranking legible, referent not. Fix: naming, not charting.
  Recused: craft, hierarchy, contrast

forger     order C,A,B   pairwise C>B, C>A, A~B   ranking C,A,B   high
  (tokens denied) Forged B from a stock bar component in 3m40s, reached
  ~90%: same encoding, sort, inline labels. Could not reproduce the
  dwell-band shading. Forged C: gave up at 4m; AOI polygon registration is
  not template work.
  B / "anti-slop: domain-specific decisions" / MEDIUM — 90% forgeable in
  3m40s. Fix: make the dwell band carry more, or accept B as commodity.
  Recused: hierarchy, contrast, labeling

pedant     order A,B,C   pairwise B>C, B>A, C>A   ranking B,C,A   high
  B labels units on the axis. C prints bare numerals with no unit in frame.
  A's legend says "dwell", its caption "fixation time" — not synonyms.
  C / "label every number" / HIGH — 14 bare numerals, no unit anywhere.
  A / "terminology drift" / CRITICAL, form_fatal: no — legend vs caption in
    one figure. Fix: pick one term and use it in both.
  Recused: composition, scale survival, brand voice

One mechanic fires on the pedant's CRITICAL before aggregation. It is not form-fatal — aligning two words does not change A's encoding — so under Two bodies it gates A's commit and does not delete A from the ranking. A's verdict is FAIL and A still ranks last on its own merits, which is the ranking doing its own job rather than borrowing the defect body's. Had the pedant been scripted inline, the inline-seat restriction would have capped this at HIGH with claimedSeverity: CRITICAL preserved, because no second seat named the drift — one more reason to seat subagents.

Finding, written by the chair:

Muriel jury finding
Decision ID: gaze-dwell-readout-01
Question: which dwell-readout direction earns implementation
Panel: squinter, thumbnail, stranger, forger, pedant (all subagents)
Stipulated facts: 3 contrast runs; A and C repaired pre-seating; seed 7;
  devibe not applicable (SVG exports, no scannable component source)
Ranking: B 0.52 > C 0.41 > A 0.07 (Bradley-Terry, 15 comparisons)
Separation: P(B>C) = 3/5 = 0.60  (< 0.65 — split trigger 3)
Concordance: 20/30 = 0.67   Unique contributors: 4/5   Collapsed: no
Defects (union, max severity):
  CRITICAL A — terminology drift, legend vs caption      (pedant)
             form_fatal: no — gates A's commit, does not delete it
             from the ranking
  HIGH     A — no focal point survives blur              (squinter)
  HIGH     B — AOI identifiers opaque without the brief  (stranger)
  HIGH     C — 14 unlabeled numerals                     (pedant)
  MEDIUM   C — numerals die below 1/4 scale              (thumbnail)
  MEDIUM   B — 90% forgeable in 3m40s                    (forger)
No mutually exclusive fix pairs on one target — nothing routed to the
  split by trigger 5.
Verdict on A: FAIL (CRITICAL, not form-fatal). A ranks last on the
  comparative body; the defect body gates its commit. Neither eliminated
  it from the ranking.

Split: squinter + thumbnail rank B first; stranger + forger rank C first.
  Squinter and thumbnail optimize signal survival under degradation — B
  wins because bar length is the strongest channel and survives blur and
  downscale. Stranger and forger optimize referential grounding and
  distinctiveness — C wins because the answer is spatial and the spatial
  version cannot be forged. They are not disagreeing about craft. They
  disagree about whether this panel ranks AOIs or locates them.

Product question, unresolved: does the dwell readout serve triage (which
  AOI, ranked) or inspection (where, in place)? Product call, not visual.
  Returned to the calling agent. If triage: ship B; carry AOI names not ids
  (stranger, HIGH); strengthen or drop the dwell band (forger, MEDIUM). If
  inspection: ship C; add units (pedant, HIGH); add a downscale-surviving
  rank strip (thumbnail, MEDIUM).

No Delphi round: the commit was not blocked this session, and the split is the useful output. Six ledger records emitted, one per finding and not one per seat, all at disposition: open. They stay open until the calling agent picks a direction, applies the fixes, and re-renders; the agent that owns that fix pass writes each disposition at that re-render.

Both reported statistics are countable from the ballots above. Separation: the five B-vs-C comparisons split 3 for B, 2 for C, none tied, so P(B>C) is 3/5 = 0.60, under the 0.65 trigger. Concordance: 10 seat pairs × 3 option pairs = 30 judgements, of which 16 agree outright and 8 involve exactly one tie at half credit, so 20/30 = 0.67, well under the 0.90 collapse threshold.

Naming decisions

The protocol was written for visual artifacts. It has been run three times on a name: the unhost project name (2026-08-29), and two sittings on 2026-10-17 choosing the App Store title of the Liquid Light Apple TV app — ww-1017-k4 (SyrupSpin, NeonSyrup, PrismSyrup) and ww-1017-m8 (Protoplasm vs SyrupSpin), which Protoplasm won. Three seats transfer with a new lens and two do not. Three gaps the visual protocol never had to close came to light in those sittings; the rules for them below are normative.

Before the panel: harvest, then generate

The jury converges. It ranks the names it is handed and cannot add one, so from a thin option set the best it can return is a well-measured choice between thin options. Neither 2026-10-17 sitting had a step that widened the set, and afterwards the human asked for more candidates and for source material. Two steps come before the chair fixes the question. Both are the calling agent's work: neither is a seat, and neither belongs to the chair, which contributes nothing of its own.

1. Harvest source material into a file beside the ledger, each item cited. At minimum: the product's own text (the README, which the Pedant reads later); the medium's history and the vocabulary of the people who practised it — crew and studio names, apparatus, period terms; and the technical vocabulary of what the product actually does.
2. Generate a long list from the harvest, run the deterministic checks over all of it, and cut it to a shortlist of at most four, so the full pairwise set stays at six comparisons per seat. Hand the human the long list along with the shortlist. The human chooses which options are seated, and the long list is where that choice is made.

The Forger's template list under Rule (a) is the negative space for generation. A candidate that fills one of those templates with a domain word will be forged. Candidates with a chance of leaving residue lie outside the templates' reach from the use statement: a proper noun from the medium's history, a practitioner's term, a technical term more than one synonym away. Do not pre-screen the long list through seat lenses. The panel judges once, sealed. Screening during generation is inline seating, with the nominal denial The roster describes, and the panel's independence is spent before the panel is convened.

Findings of fact for a name

Stipulate these before seating, exactly as contrast and gamut are stipulated for a visual artifact:

• Trademark — live marks on the whole string and on each element, by class. No mark existed on SyrupSpin, but its element SYRUP was live in class 9 software. What a live mark means legally is counsel's call; that it exists is a fact. The USPTO tmsearch API returns empty documents when _source.includes is passed, and every class filter then reads zero. That is a source producing no output, which Jurisdiction forbids stipulating as clean. Omit the field and paginate with from.
• Store and registry collisions — the App Store (iTunes Search API), npm, PyPI, whichever the product ships through.
• Domains — RDAP on the likely TLDs.
• Length against each slot's limit — the App Store title allows 30 characters.
• Contrast on the name sheet — the 8:1 floor applies to the sheet as to any artifact. Both sittings measured at least 11.35:1.

A name can be settled by a fact before any ballot: the unhost sitting lost "papertrail" to a live trademark before a seat was convened.

Keep the question out of the decision ID. The ID is passed to every seat, so a question or candidate string inside it reaches every seat as a brief. Both were in the unhost ID, and its Stranger detected its own contamination and discounted a finding. ww-1017-k4 and ww-1017-m8 are neutral.

The name sheet

A naming panel reads a name sheet: one HTML card with the name set in each slot where it will appear. For the Liquid Light title those were the App Store row (icon, title, subtitle), the tvOS home-screen tile and label, the iPhone home-screen label at 11 px with ellipsis truncation, and a browser tab. Build the sheet once and vary only the name string, so the options differ in nothing else. Render it with

python -m muriel.capture file:///abs/path/A.html --tiers desktop --selector '#card' --slug A

which names the PNG by its real pixel size — a 1200×630 card at the default scale of 2 is written as A-desktop-2400x1260.png — then run muriel.squint on that PNG for the Thumbnail's companions.

Seats for a name

The Pedant — the highest-value seat for a name. Lens: the name string, plus the product's README extracted verbatim with line numbers as its sibling source, plus stipulated product facts the README omits (for Liquid Light: the Apple TV app ships with no audio input). Judges: each word of the name against the product's own vocabulary — supported, contradicted, drifting, or absent. Access: denied the sheet render, so its denial is real and its Isolation is plain subagent. In ww-1017-m8 it graded "Syrup" HIGH as a claim about the medium that README line 3 contradicts (mineral oil and dyed water), and "Spin" MEDIUM because "spin" reads as playing music on an app with no audio.

Pedants vary on a shared element, and the variation depends on the option set, so it is not noise. ww-1017-k4 graded "Syrup" LOW: it was in all three options and read as a house element. ww-1017-m8 graded it HIGH: beside a figurative rival, it read as a literal claim about the medium. Record both and do not reconcile them. The comparative ranking cannot be moved by an element every option shares, so its severity belongs to the defect union alone.

The Forger — the counterfeit is a name pool. Before opening any option, the seat builds a counterfeit pool from the use statement alone, stamps start and end with date -u +%FT%TZ, and writes it to forgery/<decisionId>/counterfeits.forgery.md. Only then does it open the options. For each, it inventories the name's decisions — construction, casing, length, each lexeme, metaphor source, register, era marker — and diffs each against the pool. F, R, Rs and the severity table are unchanged from the agent definition. The counterfeit sheet is the shared name sheet with a pool name substituted. Pool width, template list and hop distance are fixed by Rule (a) and are not the seat's to choose.

The Thumbnail — the name sheet's downscale companions. Lens: the thumb-eighth companion of the name sheet, with each slot's real size as the intended use size. Judges: truncation in each slot, word shape at the small slots, and how the name skims in a list. Only the 1/8 companion discriminates. At 16 px no word survives on any option, so every option fails there equally and there is no ranking signal in the 16 px reading; record it and recuse from ranking on it.

The Squinter — unseated. Blur keeps position and area and removes the letters. The sheets differ only in the name string, so their blurred renders differ by a few pixels of word length, with no naming signal in them. Record the seat as unseated rather than giving it a costume: a seat without a lens correlates with every other seat.

The Stranger — single-artifact only. The seat is denied the option set, so it cannot rank names. Seat it on one sheet, or on the published page once one exists, for its three questions. Its Ship (this seat's loss only) stays out of the direction ranking, as on any single-artifact ballot.

A naming panel is therefore Pedant, Forger and Thumbnail: three seats, the minimum The chair treats as a real panel. Budget three critiques per option plus one shared counterfeit pool. Expect low concordance. The unhost panel measured 0.33, with the meaning seats and the scale seat anti-correlated, and that split — meaning against legibility — was the finding.

Rule (a): a fixed Forger pool

On a name, the Forger's severity moved with the size of its pool. SyrupSpin was rated HIGH against a 30-name pool in ww-1017-k4, where "Syrup" survived as residue, and CRITICAL against a 105-name pool in ww-1017-m8, where "syrup" was scored near, one synonym hop from the pool's "Viscous". Same name, same seat, two severities, and the two pools were also built from different template lists. A severity that moves with the seat's stopping point is a measure of the stopping point. Fix all four inputs:

• Width — at least 100 names, and at least 6 from every template. That width is reached well inside the four-minute ceiling: the 105-name pool took at most 48 seconds. With a narrower pool, the templates' reach is understated, and that reach is the claim the seat exists to test.
• Templates — all sixteen, the union of both sittings' lists:

#	Template	From the sittings
1	Closed compound: substance or texture + effect or motion	OilSwirl
2	Open pair: colour or quality adjective + substance noun	Neon Oil
3	Portmanteau of two domain words	Liquilux
4	Single evocative noun or verb	Drift
5	Period, place or venue reference	Fillmore
6	Period slang, pun or lyric	Far Out
7	Apparatus or medium, literal	Overhead
8	Domain technical term	Meniscus
9	App suffix or cute-ifier (-ly, -io, -ify, -y, -r)	Swirly
10	Latinate or pseudo-classical coinage	Lumina
11	Platform or world suffix (-TV, -scape, -verse)	GlowTV
12	Echo of the category or subtitle	Liquid Dreams
13	Nearest consumer object	Lava Light
14	Genre prefix (Psych-, Trip-)	Psychlight
15	Rhyme or reduplication	Glow Flow
16	Two-word mood phrase	Mind Melt

• Hop distance — one. A lexeme is near when it is a first-page synonym of a pool lexeme, and the diff row names the hop (syrup ← viscous, pool #88). Two hops or more is residue. Without a stated distance, a wider pool and a stricter diff cannot be told apart.
• Use statement — written once per product. The chair writes it, stores it beside the ledger, and passes it verbatim to every sitting. A later sitting on the same product reuses the earlier pool unchanged and only tops it up to the minimum, so a name that reaches the panel twice is diffed against the same counterfeits both times.

Rule (b): no Forger veto

In a naming sitting the option is the string, so a Forger CRITICAL is form-fatal by construction: its only fix is a different name, and a different name is a different option. Applied as written, the form-fatal carve-out amounts to a Forger veto over direction selection — the outcome the carve-out was kept narrow to prevent — resting on the finding shown under Rule (a) to be least stable. In ww-1017-m8 the Forger rated SyrupSpin CRITICAL (R = 0 of F = 16, form_fatal: yes), and under the carve-out that ballot alone would have eliminated it.

In a naming sitting, the chair reports the ranking twice: once over every seated option, and once over the remainder after form-fatal eliminations, naming the finding behind each elimination.

• If both rankings put the same option first, say so. The elimination changed nothing.
• If they differ, that is split trigger 6 under Splits are the product: the comparative body prefers an option the defect body fails on form. Route it to the split report with the product question it raises — for a Forger CRITICAL, whether a name a generic template reaches is acceptable for this product — and name no winner.

That option's verdict is still FAIL under the CRITICAL. A name cannot be repaired without becoming another name, so a human who ships it anyway has declined the finding: record the disposition as miss, with the human's stated reason. The rule covers a form-fatal CRITICAL from any seat on a name; the Forger's is the one that is form-fatal every time.

Requiring a second seat to concur before eliminating, on the model of the inline-seat restriction, was the alternative. It fails here. No other seat judges distinctiveness, so concurrence on the same defect could never arrive and the Forger's CRITICAL would be silently capped. Under concurrence on any defect, two unrelated findings could combine into an elimination neither seat claimed. Under the dual report, both bodies' outputs stay intact and the conflict goes to the human, which is what the split report is for.

Rule (c): standing context crosses the subagent boundary

The chair's conversation stays on its side of the subagent boundary. What the harness loads into every agent at spawn does not. In Claude Code that includes each CLAUDE.md in scope and the auto-memory index, MEMORY.md, which is loaded into every subagent. The ww-1017-m8 Forger disclosed that the product's working name was in the index before it built its pool, and it flagged the pool items affected. A Stranger spawned into the same context would have been handed project context the protocol denies it.

Before seating:

1. Search the standing context — every file the harness loads at spawn — for each option string, the working name, and words from the question. A hit is a leak.
2. Keep pending decisions out of it. While a jury is pending, no candidate name or working name goes into an index line or a CLAUDE.md. A memory about the decision goes in a file the index does not list, or waits until the sitting closes.
3. Remove the leak or record it. Removing it means moving the line out of the index before seating. If it stays, every seat records Isolation: subagent (denial partial — standing context) and names the leaked string under Evidence seen. The Forger flags pool items that equal or contain the string and counts a template as reaching such an item only when the template produces it without the leak, as the ww-1017-m8 Forger did. A Stranger whose standing context names an option or the question is contaminated: reseat it cleanly or record it unseated, as after an abort.

Instruct every seat to report, before it reads any evidence, whether its own context names an option, the working name, or the question. A seat can see its standing context; the chair cannot see the seat's.

Anti-patterns

Do not:

• Convene a jury on a measurable fact. Contrast, gamut, dimensions, token adherence, and regex-detectable polish rules are settled by tools. A panel deliberating a contrast ratio can get it wrong.
• Seat N copies of one critic. Same lens, same loss, different adjectives buys correlated errors at one full critique per seat per option. If two seats do not deny each other's evidence or optimize different losses, cut one.
• Run a jury and muriel-critique on the same artifact. The generalist overlaps four of six seats; the union then counts one defect twice and it reads as corroboration.
• Claim a denial a scripted prompt cannot enforce. Evidence denial needs a subagent boundary. Inline, mark Isolation: inline (denial nominal) and accept the CRITICAL restriction; never seat the Stranger this way.
• Stack mutually exclusive fixes in the union. Two seats attacking one target from opposite directions is a split, not two defects, and routing it to the split is the only way the panel can acquit a form that is right.
• Average away a split. The mean of five rankings discards the finding. The split is the product; the ranking is the byproduct.
• Let severity inflate. A seat whose every finding is CRITICAL carries zero bits under any-CRITICAL → FAIL. Score dispositions, apply the discount, cap the shouters.
• Convene on a test render. Panels are for work about to be committed. A draft-DPI intermediate does not earn a panel.
• Seat a naming panel on whatever names came to mind. The jury cannot add an option. Harvest source material and generate a long list first, and hand the human the long list, not only the shortlist.
• Let the chair vote. A chair with a preference is a seventh juror holding procedural power, and it will resolve splits it was meant to report.
• Show round 1 ballots to a juror still voting. That measures conformity.
• Skip Order seen. Without the recorded permutation you cannot separate judgement from position bias, and the panel does not replay.
• Let a seat recuse from nothing, or run a Delphi round to manufacture agreement. A seat that judges everything correlates with every other generalist; movement nobody can source to specific evidence is deference.

Prior art

Adapt from the design-school crit and the award jury, which get the structure right and the measurement wrong. Right: a chair who owns the question rather than an opinion; a stated brief the work is judged against; recusal when a juror has a stake; reasoning voiced before the score is committed. Wrong: halo and star effects contaminate every subsequent criterion once one juror names a favorite; the loudest or most senior voice dominates because deliberation is public from the first minute; and absolute rubric scores show poor inter-rater reliability across juries and across sittings. Comparative judgement — Thurstone's law of comparative judgement via Bradley-Terry, as deployed in educational assessment — exists largely because absolute-scoring juries in education proved unreliable, and it yields the reliability statistic as a side effect of the binary decisions. The sealed first round is the fix for public deliberation; the Delphi round with a who-moved record is the fix for treating convergence as evidence.

The Loupe's lens is borrowed from a different tradition: the inspection loupe of print production and the close-up review in anatomy (MIT), an agent skill for isometric technical figures. anatomy's verification loop ends in 4× close-ups of every junction in both themes, because the faults its owner rejected were ones that passed at 1×: cables passing through ducts, bolt rows stopping dead mid-surface, a crease ending before its corner. muriel takes the lens and leaves the aesthetic. anatomy treats a bare face as unfinished work; muriel's Loupe may not fire on empty space.