/nexus

Silhouette Recipe — Distinction Proof for a Product Surface

/nexus silhouette "<surface or product>" — strip the logo and see whether anyone can still tell it is yours. The anti-genericness member of the Quality-Max design wing: sweep the category's visual prior art, pre-commit the ordinary moves everyone makes, then prove the surface is identifiable at above-chance rate in a blind test against competitor surfaces — without buying that distinctiveness with usability. hallmark runs the logo-swap test on a brand identity; silhouette runs it on the product surface itself, which is where the genericness actually shows.

Read this file before executing the silhouette Recipe.

1. Nature / When to Use / Boundaries

A tournament that maximizes craft can converge on the category average. Every team reads the same design systems, adopts the same component library, and arrives at the same competent, forgettable screen. runway will happily push that screen to all dims = 3: it is genuinely well-made, and it is also indistinguishable from four competitors. Nothing in the roster asks whether it is ours.

Silhouette's identity is four things no sibling carries together:

1. Sameness Ledger, pre-committed. A blind panel enumerates the category's ordinary design moves before seeing our surface — the same pre-commitment discipline eureka uses for its OBVIOUS test. "We're distinctive" cannot be claimed in hindsight against a ledger written after the fact.
2. Inverted-polarity visual sweep. Prior art is researched to avoid, not to reach (clone reproduces, wish reaches, silhouette/eureka avoid). The competitor set is simultaneously the sameness evidence and the blind-test distractor pool.
3. Blind Recognition Test as the oracle. Branding stripped, our surface is mixed with K competitor surfaces and a fresh panel is asked to attribute them. The pass threshold is declared before the measurement, expressed against chance (1/(K+1)). This is the only design oracle in the roster that is a measured rate rather than a rubric score.
4. Distinctiveness must be free. A hard constraint, not a preference: a11y and task friction may not regress. Weirdness that costs usability is rejected inside the loop, so the recipe cannot succeed by making the product strange. Every surviving distinctive move must also derive from a brand truth — a Hallmark Charter value, a stated positioning, a real product property — never from arbitrary novelty.

Use silhouette for

• A surface that scores well on craft and still "looks like every other AI/SaaS product."
• Post-runway or post-restyle verification that the winner is distinctive, not just competent.
• A category diagnosis before a redesign: what are the default moves here, and which ones are we making?
• Establishing a defensible visual position when the brand is settled but the product surface has drifted to convention.

Not this — route to

Scale: 14-30 agents × ≤3 cycles, 5-9× feature cost. Per-phase: P0 ≈ 1 · P1 ≈ 3 (blind default-move panel) · P1.5 ≈ 3-6 (visual prior-art sweep) · P2 ≈ 3-4 (recognition panel + sameness scoring) · P3 ≈ 3-6 per cycle (direction + implementation + re-measure + regression guards) · P4 ≈ 3 (fresh panel + derivation check) · P5 ≈ 1.

Range derivation (convention: reference/recipe-contract.md §1 element 7): floor = one surface, ACCEPT on cycle 1 → 1+3+3+3+3+... = 14; ceiling = a surface set, 3 cycles, wide competitor pool → 1+3+6+4+(6×3=18 with shrinking re-measures ≈ 14)+3+1 = 30. Confirm-before-launch always (intentional parity with runway/wish).

Model selection (Plan-and-Execute): the Phase 1 default-move panel, the Phase 4 derivation check, and threshold adjudication are judgment-heavy → plan-tier (opus / Fable 5). The recognition panels, sweep, and implementation → Sonnet 5 default. The recognition panel is engine-diverse where the hub allows (_common/CLI_COMPATIBILITY.md §4) — a single model family attributing surfaces carries one set of visual priors, which is exactly the bias the test is meant to exclude.

2. Termination Bound

Convergence loop ≤ 3 cycles (default N=3). Exit reasons use the canonical vocabulary:

On any non-ACCEPT exit the recipe reports the measured attribution rate, the chance baseline, and the surviving sameness items. A rate without its chance baseline is meaningless and is never reported alone.

The threshold is declared before the first measurement (Phase 1, ratified with the user) and may not be lowered mid-run. Lowering it after seeing the result converts the oracle into a formality; a threshold that turns out to be unreachable exits cap-reached with the number, not with a softer bar.

3. Confirm / Safety Gate

• Phase 0 framing: Confirm before launch — always (unconditional; intentional parity with runway/wish, not drift). Surfaces: the target surface(s), the category definition and competitor-pool size K, the proposed pass threshold against chance, the cost envelope, and — when the ask is really a craft-quality ask — a recommendation of runway/restyle instead.
• Phase 1 Distinction Contract: contract-level checkpoint; AUTORUN cannot skip. The category definition, K, and the threshold are the oracle. A category drawn too wide ("software") makes any surface distinctive; drawn too narrow it makes distinction impossible. The user ratifies the boundary.
• Phase 1.5 competitor set: contract-level. The user confirms the pool before it becomes both the sameness denominator and the distractor set — a pool missing the obvious competitor invalidates every downstream measurement, and that is cheap to catch here.
• Phase 3 implementation: standard Ask First tiers (brand-level token changes, 10+ files) apply unchanged. A distinctive move that changes brand tokens escalates to hallmark, never decided inside the loop.
• On a Fable 5 hub the Phase 0 gate subsumes the F8 cost gate — one confirmation, not two.

4. Resume

Checkpoint-resume (≥ 4 phases; silhouette resume): persist the Distinction Contract (category, K, threshold) at Phase 1 exit; persist the pre-committed Sameness Ledger and the competitor set with citations at Phase 1.5 exit — re-running the sweep on resume both re-spends and risks a different ledger, which would silently move the denominator; persist each cycle's attribution rate, the surface state, and the a11y/friction baselines at every Phase 3 boundary. The Phase 1 ledger is never regenerated after a candidate exists — that is the whole point of pre-commitment.

5. Phase Contract (AUTORUN chain template)

Phase 0   FRAMING ★Confirm-before-launch (always)
          Nexus[surface + category + competitor pool size K + proposed threshold + envelope
               + cheaper-path recommendation → runway (craft) / restyle (standard bar)]
   ▼
Phase 1   PRE-COMMIT ★contract-level checkpoint (AUTORUN cannot skip; plan-tier)
          BLIND default-move panel — Vision ‖ Muse ‖ Palette, given the CATEGORY ONLY and
          explicitly NOT our surface:
            "enumerate the design moves a competent team in this category makes as a matter
             of course" (layout skeleton · type stack · colour move · component idiom ·
             motion idiom · imagery idiom · copy pattern)
          → SAMENESS LEDGER, frozen and timestamped BEFORE any candidate is seen
          → DISTINCTION CONTRACT: category boundary · K · pass threshold vs chance 1/(K+1)
            · the brand truths distinctive moves must derive from (Hallmark Charter / stated
              positioning / real product property)
          → ✓user ratifies
   ▼
Phase 1.5 SWEEP ★contract-level ratification            [research to AVOID, not to reach]
          Compete ‖ Vector?[capture competitor surfaces] ‖ Frame? — per
          reference/research-grounding.md (tiered sources, cited Evidence Ledger),
          polarity INVERTED per eureka's Prior-Art Ledger
          → COMPETITOR SET (K surfaces, branding-strippable, cited)
          → ledger enrichment: which pre-committed moves are actually observed, and at what rate
          → ✓user ratifies the pool → the denominator FREEZES
   ▼
Phase 2   MEASURE — the baseline
          (a) BLIND RECOGNITION TEST: our surface + K competitor surfaces, wordmarks / product
              names / logos / brand-unique colour stripped, presented in randomized order to a
              fresh, engine-diverse panel that has NOT seen the Sameness Ledger
              → attribution rate vs chance 1/(K+1)
          (b) SAMENESS SCORE: % of our surface's load-bearing design decisions that appear
              verbatim in the pre-committed ledger
          (c) BASELINES CAPTURED: Palette[a11y] + Echo[task friction] — the constraint floor
              for Phase 3
   ▼
Phase 3   DIVERGE loop ⟲ {
          Vision[distinctive direction DERIVED from a named brand truth — never novelty for
                 its own sake; each move states the truth it expresses]
          → implement (Muse[tokens] ‖ Artisan ‖ Flow?[motion signature] ‖ Ink?[icon/illustration
            signature] ‖ Prose?[voice on the surface])
          → RE-MEASURE (a) fresh panel each cycle, (b) sameness score
          → CONSTRAINT CHECK (hard, non-negotiable): Palette[a11y ≥ baseline] +
            Echo[task friction ≤ baseline] + Radar[no-regression]
            → any distinctiveness gain purchased with friction is REJECTED and reverted;
              the move is recorded in the ledger as "distinctive but not free"
          }⟲  loop ≤ 3 cycles (default 3)
   ▼
Phase 4   ATTRIBUTION GATE (both parts; one oracle — the Distinction Contract)
          (a) FRESH-PANEL BLIND TEST — a panel that saw none of the prior rounds, no ledger,
              no history (producer ≠ verifier, Q9): attribution rate ≥ the pre-declared
              threshold
          (b) DERIVATION CHECK — plan-tier: every distinctive move traces to a named brand
              truth. A move that traces to nothing is decoration; it is stripped, and the
              surface is re-measured without it
          → both pass ∧ constraints held → ACCEPT
   ▼
Phase 5   DELIVER — Distinction Dossier + Guardian (if code shipped)

6. Measuring attribution honestly

• Chance is stated, always. With K competitors the baseline is 1/(K+1). A 40% attribution rate is strong at K=9 and worthless at K=1. Every reported rate carries its K.
• Panels are fresh per measurement. A panel that has seen our surface before is measuring memory, not distinctiveness. Phase 2, each Phase 3 cycle, and Phase 4 use different panels.
• Strip more than the logo. Wordmark, product name, brand-unique colour token, and any copy naming the product are removed. Leaving the brand colour in makes the test trivially passable and proves nothing about the design.
• K ≥ 3. Below that the test cannot separate distinctiveness from a coin flip; Phase 0 refuses a smaller pool and expands the category boundary instead.
• A single surface is a weak claim; a set is a strong one. When the scope is a product rather than a screen, sample 2-4 surfaces and report per-surface rates — a distinctive marketing page attached to a generic product interior is a known and reportable pattern.

7. Output Report — Distinction Dossier (named)

Emitted inside NEXUS_COMPLETE on top of the base ## Nexus Execution Report:

• Distinction Contract — category boundary, K, pre-declared threshold, the brand truths in scope
• Sameness Ledger — the pre-committed default moves with their timestamp, plus the observed rate of each in the competitor set, plus which ones we were making at baseline
• Competitor set — the K surfaces with citations (per reference/research-grounding.md tiers)
• Measurement trajectory — attribution rate per cycle with its chance baseline, sameness score per cycle, panel composition per measurement
• Distinctive moves — each surviving move, the brand truth it derives from, and its measured contribution; moves rejected as "distinctive but not free" listed with their friction cost
• Constraint record — a11y and task-friction baselines vs final; any reverted move
• Attribution Gate — fresh-panel verdict, derivation-check verdict, stripped decorative moves
• Exit reason (§2 vocabulary) + final rate vs threshold vs chance + surviving sameness items
• Handoff — convention-locked exits name the non-layout channels to pursue (voice → hallmark/prose, motion → flow, mechanism → eureka)

8. Failure Modes Prevented

9. Shared-Protocol References

10. Integration with runway

Silhouette's Phase 2 measurement can run inside a runway tournament as an added rubric dimension ("Distinction: attribution rate ≥ threshold at K"), turning the ceiling recipe's blind spot into one of its scored axes. Use that form when a flagship surface is being created; use the standalone recipe when an existing product needs the diagnosis — the standalone run additionally produces the Sameness Ledger, which the embedded form does not, and which is reusable across every later surface.

When both run, silhouette's threshold is declared at runway's Phase 1 rubric gate and inherits runway's cycle cap — the two never nest as separate loops (converge-recipe.md §4 flatten rule).

11. Decision Tree vs Neighbors

Is there a settled brand identity to derive distinctive moves from?
  NO  → hallmark first (silhouette cannot pass its derivation check without one)
  YES ↓
Is the question "is it well made?" or "is it recognizably ours?"
  WELL MADE → runway (flagship ceiling) / restyle (standard bar)
  OURS ↓
Is the surface being created now, or already shipped?
  CREATED NOW → runway with silhouette's Phase 2 as a rubric dimension (§10)
  ALREADY SHIPPED ↓
Is the object a look, or a mechanism?
  MECHANISM → eureka (novelty-proven invention; same pre-commit discipline)
  LOOK → silhouette — pre-commit → sweep → measure → diverge ⟲ → fresh-panel gate → prove