Approved mock and UI implementation identical with not even one millimeter of deviation The discipline for making them so.
Trigger = the repeated "approved mock ≠ implementation" accidents in native/plugin (dsp) (stopping at "close" due to vibe-based porting).
Evidence = MUEDlim UI (CSS actual-value 1:1 port → 100% verify with conductor overlay/pixel-diff → notarized build sha a832976a… bit-identical)。
Source of truth = this skill. If there is a discrepancy with the section-local memory (dsp feedback_render_verify_visual_fidelity /
feedback_design_to_native_exact_pixel_transplant), then treat this skill as authoritative Assume that.
0. Scope of application (determine first — avoid bureaucratization creep)
IN (subject to this mechanism): the approved mockexists/should existwork.
- Implementation of a new UI surface
- re-skin (visual replacement of an existing UI)
- Visual redesign
OUT (out of scope; keep the existing Tier/review discipline):
- Existing implementation'sminor fix: changing one line of copy / fine-tuning padding or margin / bug fix
- Work for which no mock exists as an approved artifact or no mock is needed
The purpose is not to thinly sprinkle bureaucracy across all sections.Stricten the native/plugin where it is occurringThat is the purpose. The pattern is concentrated in native/plugin UI (dsp); web UI (React) has little applicability (§7 calibration).
1. rule 1 — pinned mock = single source of truth
- UI dispatch must alwayspin the approved mock and be performed.Prohibit UI implementation requests without a mock。
- pin は with a unique identifier such as filename / hashPerform it.Prohibit ambiguous labels such as "Plan A" or "Plan B" (They get reused across stages and collide. In fact, the first draft collided by referencing "Plan A" = a live example of this rule.)
- Example:
muedlim_ui_lufs_hero_mockA_v2_20260713.png(○) / "in the direction of Plan A" (×).
2. rule 2 — numeric spec extraction (prohibit vibe-based porting)
From the mock, extract the followingnumeric tableas a numeric table and port it as-is.Prohibit "nudging it closer by vibe"。
- Coordinates (x/y), dimensions (w/h), aspect ratio
- Colors (hex)
- Fonts (typeface, weight, size, letter spacing)
- Spacing (padding / margin / gap)
★If the mock is self-made HTML/CSS, port the actual source values 1:1(exactness is achievable without eyeballing = dsp's core insight). "There is no environment where you're doing it with the same CSS and can't use the same thing" (kimny).
3. rule 3 — diff must be attached to the completion report
In the completion report, render × mock juxtaposition + overlay / pixel-diff must be attached as a requirement.
| Type | Expectation |
|---|---|
| web(React) | pixel-perfect(Playwright screenshot match suffices) |
| native/plugin | Numerical match + pixel-diff + explicit tolerance + flag non-reproducible points(silent approximation prohibited) |
Handling of non-reproducibility flags (strict)
- Only things truly originating from the engine may be flagged: sub-pixel AA, hinting, etc.
- Missing font weight bundling / approximations of size and coordinates are fixable and therefore cannot be flagged = 100% match required. Do not use "because it's native" as an excuse for tolerance.
- ★The burden of proof for "non-reproducibility" lies with the flagging side(unless it can be shown to originate from the engine, it is treated as fixable). Example: flagging Sora without proof, assuming it was single-weight, was a failure (in reality it was resolved by unifying to the house font).
- If elements that are truly non-reproducible remain, consider webview (direct HTML/CSS rendering of the mock) as a fallback。
Conditions for render
- render は Editor portion only(crop the standalone window frame = compare under the same conditions as the mock).
4. rule 4 — double gate (block push routes without vet)
- Peer attaches diff.
- Conductor does not give an impression-based review of "clean" but actually inspects the mock comparison and vets。
- "Close" isrejected。Only "indistinguishable" passes。
- conductor vet は Verify font weight + all component sizes 100% via overlay/diff。
- → kimny push。
- Prohibit kimny push without vet(the conductor's overestimation of "clean" is a real example of this hole).
5. rule 0 — mock hygiene (prevention of upstream defects)
pinned mock は Build with shippable / house assets。
- Mixing in assets that cannot be bundled, such as system / proprietary fonts (e.g., Helvetica Neue),is prohibited. → This becomes a structural defect where impl cannot achieve a 100% match in a shippable build (today's real example → fixed by unifying to house Sora).
- Guarantee on the mock side the premise that "impl can be made 100% matching and shippable" (use house / OFL fonts, etc.).
※ MockDesign quality(errors in the mock's own grammar/direction. e.g., GR grammar error in a progress-bar) is handled separately (by grounding in a real reference
feedback_design_reference_real_convention_not_invented). Separate this from mock hygiene (asset shippability).
6. Exception table (append-only, explicitly enumerated)
Do not create implicit exceptions. List only approved exceptions here.
| # | Exception | Approval |
|---|---|---|
| 1 | version string display (not in the mock but allowed as real plugin information) | kimny 2026-07-13 |
7. calibration (handling uneven distribution — avoid blanket bureaucratization)
- pattern は Concentrate on native/plugin UI (dsp). web UI (React) is sparse in this regard (product team creators = Playwright 12/12 match the mock; the mued hero is a copy issue and not a mock-fidelity issue).
- ∴ universal principles + strict native calibration: Don't spread it thinly across all lessons; instead, make native/plugin stricter.
8. Separation of gate roles (important)
- The fidelity gate (indistinguishable) is not a substitute for the kimny aesthetic gate。
- Fidelity = "Does it match the approved mock or not" / Aesthetic = "Will kimny ultimatelygoodjudge it as such" = separate layers.
- If a design-principle change at a different layer (e.g., GR-hero → LUFS-hero) comes in during implementation, replace the mock → absorb it through reimplementation(mechanism = fidelity / doctrine = design principle separation works).
Reference
- Source-of-truth proposal:
_conductor/docs/inbox/proposal-ui-mock-fidelity-mechanism-20260713.md - Empirical record (dsp):
dsp_muedlim_ui_lufs_hero_impl_plan_20260713.md/dsp_muedlim_ui_A_size_audit_pixeldiff_20260713.md/ memoryfeedback_design_to_native_exact_pixel_transplant(iPlug2 port, 8 techniques) - Always-on version of DoD = common block in each lesson's CLAUDE.md
20-ui-mock-fidelity