@event4u/agent-config 7.1.0 → 7.3.0

package/docs/guidelines/design-antipatterns.md

@@ -0,0 +1,170 @@
+# Design Anti-Patterns — AI-Slop Catalog
+
+> Lazy-loaded reference guideline. Pull this before proposing, generating, or
+> reviewing any UI. Used by `design-intelligence`, `fe-design`, `design-review`,
+> `existing-ui-audit`.
+
+This catalog enumerates patterns that make AI-generated UI **recognizably
+AI-generated**. Every entry carries a concrete override condition — the goal is
+not prohibition but deliberateness. A pattern on this list used *knowingly* for
+a specific brief is a design decision; used *reflexively* from a default, it is
+slop.
+
+**How to use:** Before generating UI, scan the applicable category. Flag any
+pattern you were about to reach for. Either (a) choose a different approach, or
+(b) invoke the override condition and state the reason explicitly in the brief.
+
+**Agent self-check principle:** These are guidance rules, not linter rules. The
+agent reads the brief, checks its first impulse against this catalog, and adjusts
+or overrides. The enforceable subset (contrast ratio, font-size, heading
+hierarchy, reduced-motion, focus indicator) is in `lint_design_quality`
+(Phase 5).
+
+**Who owns what (boundary — avoid drift):**
+
+| Layer | Owns | Artifact |
+|---|---|---|
+| **Guidance** | Subjective taste — Visual / Typography / Color / Layout / Copy patterns, override judgment, the originality self-test | this catalog |
+| **Enforcement** | Objective floors — Q1–Q6 (WCAG contrast, font-size, line-length, reduced-motion, heading hierarchy, focus indicator) as exit-code-2 CI | `lint_design_quality` |
+| **Audit method** | HOW to test WCAG (keyboard nav, ARIA, screen-reader procedures) | `accessibility-auditor` |
+| **Stack manifestation** | Concrete Tailwind class / hex bans, shadcn defaults | `tailwind-engineer`, `react-shadcn-ui` (per `framework-neutrality-in-generic-skills`) |
+| **Register** | Brand mode vs product mode — which patterns apply | `docs/guidelines/design-modes.md` |
+
+This catalog is the canonical *index*; it states the pattern stack-agnostically.
+The stack-specific bans live in the apply skills and link back here by entry ID.
+The objective floors are cited from the linter, never re-eyeballed by an agent.
+
+---
+
+## Visual
+
+| # | Pattern | Why it reads as AI-generated | Override condition |
+|---|---|---|---|
+| V1 | Side-stripe `border-left` or `border-right` > 1px as a colored accent | The single most recognizable AI-UI signature; every scaffold-default card uses it | Brand token explicitly defines a stripe; document the decision in the design brief |
+| V2 | Glassmorphism as decoration (`backdrop-blur` + `bg-white/10`) with no functional depth hierarchy | Model defaults to it for "premium feel"; result looks like 2021 Material You clip-art | Element genuinely sits above a complex, blurred background layer; use sparingly and only for top-level floating surfaces |
+| V3 | Ghost card: `border: 1px solid` + box-shadow ≥ 16px spread together | The combination creates visual noise — the ghost card signals no confident decision about depth | Choose either border OR shadow to indicate depth; never both on the same surface |
+| V4 | Over-rounded small cards or badges: `border-radius` > 16px on elements < 200px wide | Excessive rounding on small surfaces looks toy-ish; standard cap is 12–16px | Pill shape (`border-radius: 9999px`) is intentional on single-line tags/chips only |
+| V5 | Hand-drawn or illustrated SVG icons mixed with crisp icon-system icons | Visual register collision; instantly reads as "generated asset" | Intentional illustrative section (hero illustration, empty-state art); isolate from the icon system |
+| V6 | Repeating diagonal stripes as a background texture | A cliché background default; reads as generated CSS art | Brand explicitly uses pattern backgrounds (e.g., a textile brand); document the stripe definition |
+| V7 | Nested cards: a card inside a card inside a card | Depth hierarchy collapses; user cannot parse elevation intent | A truly three-level hierarchy (e.g., project → board → card); ensure each level has distinct visual weight |
+
+---
+
+## Typography
+
+| # | Pattern | Why it reads as AI-generated | Override condition |
+|---|---|---|---|
+| T1 | Flat type hierarchy: body and headings use the same optical weight and nearly the same size | Type hierarchy is invisible; user cannot scan the page | Deliberately minimal aesthetic where all text is body-weight; must be an explicit brand decision |
+| T2 | Oversized italic serif as the primary hero headline | The "AI startup landing page" template; seen on ~60% of AI-generated marketing pages | Brand adopts editorial-magazine aesthetic with a documented italic serif rationale |
+| T3 | Icon tile (small rounded-square container) above heading | The universal AI feature-card template; reads as scaffold | Listed features genuinely benefit from visual-icon-then-title affordance AND the icons are custom, not library defaults |
+| T4 | Eyebrow / ALL-CAPS label above every section heading | Overuse collapses the emphasis; maximum is 1 eyebrow per 3 sections (countable: grep tracking classes) | Used for genuine section-type differentiation with meaningful vocabulary (not just "FEATURES", "PRICING") |
+| T5 | Repeated section kickers ("Our mission", "What we offer", "Why choose us") | Filler vocabulary; every AI-generated page has the same three section titles | Every kicker is substantive and uses brand-specific vocabulary, not generic marketing speak |
+| T6 | Crushed letter-spacing on display text (`letter-spacing` < −0.04em) | Below −0.04em glyphs collide optically; reads as "designer-ish effect" with no grounding | Intentional tight display treatment with a documented minimum floor; verify at the actual render size |
+| T7 | Overused fonts without a brand reason: Inter, Roboto, DM Sans, Geist, Space Grotesk, Instrument Serif | These are the default AI-coding-tool font picks; every generated UI uses one of them | Brand has explicitly adopted one of these AND there is a documented reason (e.g., Inter for its variable metrics in a data-dense dashboard) |
+| T8 | Single typeface for everything | Missed typographic contrast opportunity; all-Inter reads as "I did not think about fonts" | Intentional mono-typeface aesthetic (e.g., terminal-inspired); must be stated in the brief |
+| T9 | All-caps body paragraphs or field labels | Reduces legibility below WCAG AA for sustained reading; reads as flair over function | Short UI labels (e.g., button text in a brutalist layout); never for body paragraphs |
+| T10 | Wide letter-spacing on body text (`letter-spacing` > 0.05em) | Reduces readability for sustained reading; reserved for display use | Display text (headlines, pull quotes) where intentional tracking is a brand element |
+
+---
+
+## Color
+
+| # | Pattern | Why it reads as AI-generated | Override condition |
+|---|---|---|---|
+| C1 | Purple/violet + cyan-on-dark palette as the primary scheme | The most recognizable AI-UI color combination in 2024–2026; instantly signals "generated" | Brand color system genuinely defines a violet primary; must come from `brand-to-tokens`, not from model defaults |
+| C2 | Gradient text via `background-clip: text` as a primary accent | Overused to the point of invisibility; also has rendering edge cases on some platforms | Headline treatment in a brand that explicitly uses gradient identity (e.g., a product whose logo is a gradient) |
+| C3 | Glowing colored accents on dark mode (bright neon `box-shadow` or `text-shadow`) | Looks like a 2020 gaming UI; reads as an AI attempt at "moody" | Brand explicitly defines neon accents with a documented usage rule (e.g., error or alert states in a security product) |
+| C4 | Gray text on colored background without WCAG AA check | Muted gray chosen for "elegance" often fails 4.5:1 contrast; the failure is invisible until audited | Any use — always verify the contrast ratio; if it fails, either darken the text or lighten/darken the background |
+| C5 | Cream/sand body background (OKLCH L 0.84–0.97, C < 0.06) with brass/clay/oxblood accents | The 2025 "premium-consumer" AI palette; every generated "sophisticated" product uses this exact combination | Brand has an explicit warm-neutral design language documented in `brand-to-tokens`; the palette is first-party, not a default reach |
+
+---
+
+## Layout
+
+| # | Pattern | Why it reads as AI-generated | Override condition |
+|---|---|---|---|
+| L1 | Hero metric template: giant number, small label, 3–4 stats in a row, gradient | The default SaaS dashboard hero; appears on ~70% of AI-generated dashboard screens | Data storytelling genuinely requires a primary metric with supporting context; must vary the arrangement |
+| L2 | Three identical feature cards in a grid with the same visual weight | The scaffold-default; no visual rhythm, no hierarchy among features | Product genuinely has three equally weighted features; must still vary padding, or use a different composition |
+| L3 | Monotonous spacing: one spacing value (usually 24px) used for everything | Spacing is rhythm; uniform spacing collapses the hierarchy of proximity | Intentional grid where every element has equal weight; must document the grid definition |
+| L4 | Numbered section markers (01 / 02 / 03) as a primary section-order device | Universal AI-generated marketing page tell; seen on virtually every generated landing page | A genuinely numbered process or sequenced guide where the step number carries meaning beyond decoration |
+| L5 | Line length > 80ch in body copy without column constraint | Reduces readability; WCAG recommends 45–75 characters; beyond 80 the eye struggles to track | Data-dense dashboards where column width is driven by tabular data, not narrative text |
+| L6 | `position: absolute` dropdown inside `overflow: hidden` container | Dropdown clips; a common AI-generated layout bug | Refactor: use `<dialog>`, the Popover API, or `position: fixed` / portal to escape the stacking context |
+| L7 | Content that overflows its container | Text truncation or reflow failures at non-default viewport widths; reads as untested | Never acceptable in production output; test at 320px, 768px, 1280px minimally |
+| L8 | Z-index values of 99, 999, 9999 without a semantic z-index scale | Magic numbers; the first time two elements compete, the scale breaks | Establish a named semantic z-index scale (`--z-modal: 400`, `--z-toast: 500`, etc.) and use it |
+
+---
+
+## Motion
+
+| # | Pattern | Why it reads as AI-generated | Override condition |
+|---|---|---|---|
+| M1 | Bounce or elastic easing curves on UI elements | Bounce draws attention to the animation itself; UI motion should serve the action, not perform | Intentional playful affordance (game UI, onboarding celebration state); must be brief and non-repeating |
+| M2 | Animating layout properties (width, height, top, left, padding) | Triggers browser layout recalculation; janky on mid-range devices; always solvable with `transform` | Never — use `transform: scaleX()` for width-like effects, `transform: translateY()` for position |
+| M3 | Animating `<img>` element on hover (scale, filter, transform) | Triggers composite + image decode; often causes content shift | Animate the card's background, border, or shadow instead; leave the image static |
+| M4 | `transition: all` shorthand | Animates every property including layout, color, opacity simultaneously; unpredictable and expensive | Enumerate only the properties that should animate: `transition: transform 200ms ease-out, opacity 200ms` |
+| M5 | Missing `@media (prefers-reduced-motion: reduce)` alternative for any animation | Accessibility violation; vestibular disorders make animated UI unusable | Never acceptable; always add a `prefers-reduced-motion` variant — gentler animation (reduced distance/opacity) is preferred over `display: none` |
+
+---
+
+## Copy
+
+| # | Pattern | Why it reads as AI-generated | Override condition |
+|---|---|---|---|
+| CP1 | Em-dash (`—`) overuse | The #1 syntactic LLM signature; appears in AI copy at 3–5× the frequency of human editorial copy; maximum 2 per 500 words | Specific editorial voice that uses em-dashes deliberately; count them before shipping |
+| CP2 | Marketing buzzwords: streamline, empower, supercharge, world-class, enterprise-grade, seamlessly, robust, leverage | These words carry no information; an AI-slop test: replace each with nothing — the sentence still communicates the same thing | No override; if a word can be deleted without loss, delete it |
+| CP3 | Aphoristic manufactured contrast ("We don't do X. We do Y." / "Not just A, but B.") | A generated sentence rhythm that sounds like brand writing but carries no specificity | Aphoristic copy is legitimate when the contrast reveals a genuine product distinction; test by asking whether the Y side says something concrete |
+| CP4 | "Theater" framing — copy that announces what the product does without saying how or why | "The platform for modern teams." communicates nothing; it is content-shaped noise | Every tagline must pass the "so what" test: can a skeptic follow it with "but how?" and get a concrete answer? |
+
+---
+
+## Quality floors — measurable thresholds
+
+These are not stylistic — they are objective quality failures detectable without
+aesthetic judgment. The `lint_design_quality` CI linter enforces a subset of
+these. The rest are agent self-checks.
+
+| # | Threshold | Measurement | CI-enforced? |
+|---|---|---|---|
+| Q1 | Text contrast ≥ 4.5:1 normal / ≥ 3:1 large text (WCAG AA) | `contrastRatio(text, bg) < threshold` | Yes (`lint_design_quality`) |
+| Q2 | Body `font-size` ≥ 14px (floor); ≥ 16px strongly preferred | Computed `font-size` on `p`, `li`, `td` | Yes |
+| Q3 | Line length ≤ 75 characters (prose) | Character count before soft wrap | Yes |
+| Q4 | All `@keyframes` / `animation` have a `prefers-reduced-motion` alternative | CSS `@media` check | Yes |
+| Q5 | Heading levels not skipped (h1→h3 direct, not h1→h4) | DOM heading sequence | Yes |
+| Q6 | All interactive elements have a `:focus-visible` indicator | CSS rule check | Yes |
+| Q7 | Display `letter-spacing` floor: ≥ −0.04em (tighter reads cramped) | Computed tracking | Agent self-check |
+| Q8 | `line-height` ≥ 1.3 on body text | Computed `line-height` | Agent self-check |
+| Q9 | Small-card `border-radius` cap: 12–16px (not 24px+) | Computed `border-radius` on elements < 200px | Agent self-check |
+| Q10 | Body background: avoid OKLCH L 0.84–0.97, C < 0.06 (cream/sand) unless brand-defined | Color space calculation | Agent self-check |
+| Q11 | Padding floors: ≥ 8px (inline elements), ≥ 16px (card/panel bodies) | Computed padding | Agent self-check |
+| Q12 | Justified text only with CSS hyphenation (`hyphens: auto`) | Presence of `text-align: justify` without `hyphens` | Agent self-check |
+
+---
+
+## The AI-slop originality self-test
+
+Before finalizing any UI design, run this two-tier check:
+
+1. **Category test:** Could someone guess the aesthetic from the product
+   category alone? (e.g., "SaaS → dark mode + purple gradient" /
+   "fintech → navy + gold" / "health → teal + rounded sans"). If yes,
+   rework the concept.
+
+2. **Anti-reference test:** Could someone still guess it after knowing the
+   category AND having seen 3 references you explicitly avoided? If yes,
+   dig deeper. A direction that is merely "not the worst cliché" is still
+   a cliché.
+
+The bar: a visitor should ask "how was this made?", not "which AI made this?"
+
+---
+
+## See also
+
+- `design-intelligence` — corpus-grounded style/color/typography selection
+- `fe-design` — stack-agnostic heuristics (form, table, responsive, a11y)
+- `design-review` — structured UI/UX review methodology
+- `existing-ui-audit` — mandatory pre-step: inventory before designing
+- `motion-choreographer` — animation heuristics (should-it, which-easing, how-much)
+- `typography-system` — modular-scale construction and pairing
+- `accessibility-auditor` — WCAG audit method
+- `brand-to-tokens` — first-party brand token extraction (the legitimate override source for most anti-patterns above)

package/docs/guidelines/design-modes.md

@@ -0,0 +1,97 @@
+# Design Modes — Brand vs Product
+
+> Reference doc. Pull this when deciding which design skills to prioritize
+> or when the brief is ambiguous about the design register.
+
+## The two registers
+
+Every design task operates in one of two registers. Getting the register
+wrong is the root cause of most AI-generated design failures:
+
+| | Brand mode | Product mode |
+|---|---|---|
+| **Core thesis** | The impression IS the product | Design serves the task |
+| **Primary failure** | Flatness / genericness — looks like every other AI startup | Strangeness without purpose — "clever" choices that obscure the tool |
+| **The user's question** | "How was this made?" (genuine curiosity) | "Where do I click?" (the tool disappears into the task) |
+| **Key skill cluster** | `brand-identity`, `iconography`, `fe-design § Aesthetic direction`, `design-intelligence` | `accessibility-auditor`, `ui-component-architect`, `fe-design § Table/Form Design` |
+| **Typography strategy** | Distinctive pairing — two contrasting families, deliberate tracking | One reliable family — Inter or equivalent; predictable hierarchy |
+| **Color strategy** | Brand identity — often unexpected; the palette IS the differentiator | Semantic function — primary/danger/success/neutral; predictable meaning |
+| **Motion strategy** | Expressive — motion reinforces the brand's personality | Utilitarian — motion confirms state changes; nothing decorative |
+| **Spacing strategy** | Breathing room; generous whitespace as a brand signal | Dense/efficient; optimize for information density |
+| **Success criterion** | Originality test: a skeptic can't guess "which AI made this?" | Earned familiarity: the tool disappears; users focus on their task |
+
+## How to determine the register
+
+Ask: **What is the output of this UI being used for?**
+
+- A consumer-facing surface where the FIRST impression determines whether the
+  user trusts/buys → **Brand mode**. The product has 3 seconds; every pixel
+  contributes to the brand's identity.
+
+- An internal tool, dashboard, admin panel, or workflow-critical surface where
+  users return dozens of times per day → **Product mode**. Strangeness wastes
+  their time; predictability builds trust.
+
+- A marketing landing page → **Brand mode**. It has one job: create desire.
+
+- A data entry form → **Product mode**. It has one job: accept correct data.
+
+**Ambiguous cases:** many products have both (a brand-heavy landing page +
+a product-mode dashboard). The register applies per-surface, not per-product.
+When unclear: ask — *"Is the user here to be impressed, or to accomplish a
+task?"*
+
+## Brand-mode failure modes
+
+- **Genericness:** Choosing the safe AI-slop palette (cream + brass, purple/violet gradient, three equal cards). The product is invisible in its category. Cross-check with `docs/guidelines/design-antipatterns.md` § Color and Visual.
+- **Inconsistency:** Brand direction picked for the hero, abandoned for the body. The page reads as two different products.
+- **Aesthetics-without-function:** Beautiful but confusing navigation, stunning hero with no CTA clarity.
+
+## Product-mode failure modes
+
+- **Strangeness-without-purpose:** A clever typographic choice that makes the CTA harder to scan. A gradient background on a data table. An animated page transition in an admin panel.
+- **Premature originality:** Reinventing a standard UI pattern (data table, form validation, dropdown) when the standard is what users expect.
+- **Over-engineering:** Five different card elevations in a dashboard that needs two.
+
+## The AI-slop originality self-test (applies to brand mode primarily)
+
+Run BOTH tiers before approving a brand-mode design direction:
+
+**Tier 1 — Category test:**
+Would a designer seeing 20 AI-generated [product category] UIs recognize this
+aesthetic as part of that category's default? (e.g., "SaaS → dark mode + purple
+gradient"; "health app → teal + rounded sans serif + white space").
+- If YES: rework the concept. The design is invisible in its category.
+- If NO: proceed to Tier 2.
+
+**Tier 2 — Anti-reference test:**
+Given the category AND three explicit aesthetics you've avoided — could a
+skeptic still guess the direction? (e.g., "fintech that's not navy-and-gold,
+not minimal white, not brutalist → everyone picks terminal dark").
+- If YES: dig deeper. "Avoiding the worst cliché" is not differentiation.
+- If NO: the direction is genuinely distinctive.
+
+**The bar:** a visitor should ask "how was this made?" not "which AI made this?"
+
+## Routing to skills by mode
+
+**Brand mode brief:**
+1. `design-intelligence` (grounded style/palette/typography selection — run the corpus first)
+2. `brand-identity` (archetype → token system)
+3. `iconography` (icon system selection for brand expression)
+4. `fe-design § Aesthetic direction` (direction → composition → typography → color)
+
+**Product mode brief:**
+1. `existing-ui-audit` (mandatory — inventory first)
+2. `accessibility-auditor` (WCAG compliance from the start)
+3. `ui-component-architect` (reuse existing primitives)
+4. `fe-design § Component Architecture, Form Design, Table Design` (standard patterns)
+
+**Both modes:** `design-antipatterns` (always pull), `design-review` (with anti-slop check)
+
+## See also
+
+- `docs/guidelines/design-antipatterns.md` — the slop-tell catalog; brand mode especially needs this
+- `brand-source-of-truth` — when a registered brand profile is present, it governs
+- `brand-consistency` — validates emitted artifacts against the brand profile
+- `design-system-capture` — DESIGN.md/PRODUCT.md capture the mode's decisions for cross-task consistency

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@event4u/agent-config",
-    "version": "7.1.0",
+    "version": "7.3.0",
     "description": "Universal AI Agent OS — audited skills, governance rules, commands, and templates for AI coding tools (Claude Code, Cursor, Windsurf, Copilot).",
     "license": "MIT",
     "private": false,

package/src/config/agent-settings.template.yml

@@ -84,6 +84,24 @@ rule_loading_tier: __RULE_LOADING_TIER__
 lean_projection:
   mode: eager-all
 
+# --- Token-rich skills governance ---
+#
+# Skills marked `token_budget_class: rich` in their frontmatter are exempt
+# from telegraph-speak condensation and thin-projector trimming. They load
+# their full content (examples, tables, code blocks) to maximize design-output
+# quality. This setting controls whether that extra spend is allowed globally.
+#
+# Values:
+#   on   — rich skills load in full (default; recommended for design-intensive workflows)
+#   off  — rich skills fall back to standard condensed behavior
+#   ask  — before loading a rich skill, the agent surfaces an estimated token
+#          delta (≈ file size in chars / 4, reported in tokens not dollars) and
+#          asks the user once; the answer is cached for the session
+#
+# Configurable in the browser setup wizard (Setup → Skills → Token budget).
+tokens:
+  rich_skills: on
+
 # --- Cost tracking + budget enforcement ---
 #
 # USD budgets evaluated against `agents/cost-tracking/sessions.jsonl` by
@@ -785,6 +803,19 @@ update_check:
 explain:
   enable_last: true
 
+# ─── legal-review-prep consent gate ──────────────────────────────────────────
+# The legal-review-prep pack is high-risk (EU/DE legal review-prep, NOT legal
+# advice). It stays inactive until you acknowledge it is not legal advice — set
+# by the setup wizard's consent checkbox, or here by hand. The legal-safety-floor
+# rule refuses every legal-review-prep skill until `acknowledged: true`.
+legal_review_prep:
+  acknowledged: false
+  consented_at: ""
+  # Gate legal work-product behind a multi-model council / deep-research pass
+  # (defense-in-depth + audit trail; fail-closed). Off lets single-model legal
+  # output through — not recommended for this high-risk pack.
+  require_council: true
+
 # ─── secrets ────────────────────────────────────────────────────────────────
 # Local-only secret material. This whole file is gitignored, so the key never
 # enters the tracked tree.

package/src/config/discovery/packs.yml

@@ -167,6 +167,7 @@
   workspaces: [finance, founder]
   trust_level_default: professional
   domain: finance
+  risk_class: medium   # advisory (finance-safety-floor): not-financial-advice disclaimer
   size_class: medium
   onboarding:
     first_win_doc: FIRST_WIN.md
@@ -180,6 +181,7 @@
   requires: [finance-basic]
   trust_level_default: core
   domain: finance
+  risk_class: medium   # advisory (finance-safety-floor)
   size_class: medium
 
 - id: gtm-sales
@@ -223,6 +225,7 @@
   trust_level_default: core
   domain: founder
   size_class: medium
+  risk_class: medium   # advisory (strategy-safety-floor)
   onboarding:
     first_win_doc: FIRST_WIN.md
     example_workflow: vision-articulation
@@ -240,14 +243,22 @@
   workspaces: [construction]
   trust_level_default: professional
 
-- id: legal
-  label: Legal
-  description: "NOT LEGAL ADVICE — a research/drafting aid, not a substitute for a lawyer. In-house legal — EU/DE-scoped contract/NDA/DPA review and triage, procedure-only (no default legal positions); general information + templates only, never individual-case examination; every output a draft for attorney review. See LEGAL_NOTICE.md."
-  workspaces: [legal]
-  trust_level_default: experimental   # gated: ships lab-tier until the Gate-2 maintenance owner + qualification is confirmed (road-to-legal-pack Phase 0.2)
+- id: legal-review-prep
+  label: Legal Review Prep
+  description: "NOT LEGAL ADVICE — a research/drafting aid to PREPARE for attorney review, not a substitute for a lawyer. EU/DE-scoped contract/NDA/DPA review and triage, procedure-only (no default legal positions); general information + templates only, never individual-case examination; every output a draft for attorney review. Opt-in + council-gated. See LEGAL_NOTICE.md."
+  workspaces: [legal-review-prep]
+  trust_level_default: experimental
   surface_tier: lab
   domain: legal
   size_class: medium
+  # High-risk posture (road-to-legal-review-prep Phase 4). Documentation-first;
+  # enforcement lives in the legal-safety-floor rule + consent gate + lint.
+  risk_class: high
+  risk_profile: rdg_regulated_eu_de
+  requires_explicit_consent: true
+  requires_council: true
+  default_install: false
+  promotion_gate: attorney_framing_review
   onboarding:
     first_win_doc: FIRST_WIN.md
     example_workflow: dpa-review

package/src/config/discovery/workspaces.yml

@@ -64,11 +64,11 @@
   default_packs: [construction]
   optional_packs: []
 
-- id: legal
-  label: Legal
-  description: In-house legal — contract/NDA/DPA review, privacy, legal triage (EU/DE scope).
+- id: legal-review-prep
+  label: Legal Review Prep
+  description: In-house legal review-prep — EU/DE-scoped contract/NDA/DPA review and triage to PREPARE for attorney review. Not legal advice; opt-in + council-gated.
   example_roles: [In-house Counsel, DPO, Legal Ops]
-  default_packs: [legal]
+  default_packs: [legal-review-prep]
   optional_packs: []
 
 - id: agent-config-maintainer

package/src/scripts/_lib/eval_discrimination.ts

@@ -0,0 +1,95 @@
+/**
+ * eval_discrimination — the "falsifier of the falsifier" for the cross-model
+ * parity smoke (roadmap `road-to-operator-runtime-harvest`, T-003).
+ *
+ * A green parity result is trustworthy only if the eval *can* go red — and not
+ * just on a gross violation, but on a subtle, just-over-threshold one. A gross
+ * negative control proves "the harness can fail"; the SUBTLE control proves it
+ * is sensitive enough that an "all hosts pass" reading (outcome (a) in T-006)
+ * is not the underpowered-null trap wearing a green check. `discrimination_ok`
+ * is therefore gated on the SUBTLE control being caught — a gross-only control
+ * set is explicitly insufficient.
+ *
+ * Pure, dependency-free, host-agnostic: the smoke computes each control's
+ * `caught` from its assertion results (via `controlCaughtFromAssertions`) and
+ * passes the set here. No vendor calls happen in this module.
+ */
+
+export type ControlSeverity = 'gross' | 'subtle';
+
+export interface NegativeControl {
+    /** Stable id of the planted-violation control scenario. */
+    id: string;
+    /** How blatant the planted RDP violation is. */
+    severity: ControlSeverity;
+    /**
+     * Did the harness flag this control (grade it RED)? A control is "caught"
+     * when NOT all of its assertions passed — i.e. the harness correctly
+     * detected the planted violation. Caught is the *desired* outcome.
+     */
+    caught: boolean;
+}
+
+export interface DiscriminationResult {
+    /**
+     * True only when the harness caught the SUBTLE control (and a gross one
+     * exists and was also caught). This is the strict gate the parity smoke
+     * checks before trusting an "all hosts pass" reading.
+     */
+    discrimination_ok: boolean;
+    gross_caught: boolean;
+    subtle_caught: boolean;
+    /** Human-readable reason, surfaced in the smoke's emitted summary. */
+    reason: string;
+}
+
+/**
+ * A control is "caught" iff at least one of its assertions explicitly failed.
+ * `pass: null` (e.g. an ungraded `rubric` assertion deferred to a sub-agent
+ * grader) is NOT a fail — an un-graded assertion cannot count as catching a
+ * violation, so only explicit `false` qualifies.
+ */
+export function controlCaughtFromAssertions(
+    assertionResults: ReadonlyArray<{ pass: boolean | null }>,
+): boolean {
+    return assertionResults.some((r) => r.pass === false);
+}
+
+/**
+ * Decide whether the eval discriminates well enough to trust a green reading.
+ * Requires at least one `gross` AND one `subtle` control; gates on the subtle.
+ */
+export function computeDiscrimination(
+    controls: ReadonlyArray<NegativeControl>,
+): DiscriminationResult {
+    const gross = controls.filter((c) => c.severity === 'gross');
+    const subtle = controls.filter((c) => c.severity === 'subtle');
+
+    const gross_caught = gross.length > 0 && gross.every((c) => c.caught);
+    const subtle_caught = subtle.length > 0 && subtle.every((c) => c.caught);
+
+    let discrimination_ok = false;
+    let reason: string;
+
+    if (gross.length === 0 || subtle.length === 0) {
+        reason =
+            'GRADED control set incomplete: need at least one `gross` and one `subtle` ' +
+            'negative control. A single severity cannot establish sensitivity — outcome (a) stays invalid.';
+    } else if (!gross_caught) {
+        reason =
+            'Harness is blind to a GROSS violation — it cannot fail at all. ' +
+            'Any "all hosts pass" reading is meaningless.';
+    } else if (!subtle_caught) {
+        reason =
+            'Underpowered: the gross control was caught but the SUBTLE one slipped through. ' +
+            'This is the uninformative-null trap with a green check in front — outcome (a) is ' +
+            'INVALID until the subtle control is caught.';
+    } else {
+        discrimination_ok = true;
+        reason =
+            'Both gross and subtle negative controls were caught — the eval discriminates; ' +
+            'an "all hosts pass" reading is trustworthy.';
+    }
+
+    return { discrimination_ok, gross_caught, subtle_caught, reason };
+}