Changelog

v0.9.0 — definitions@2026.11 engine, 5-axis telemetry, remote & local MCP, CI gates (2026-11-11)

The engine catalogue is DEFINITIONS_VERSION 2026.11. New and broadened rules ship at weight 0 (status: review) and add no points. Craft-floor findings do not change Clean / Mild / Heavy. Leaderboard rows and calibration accuracy stay the measured 2026.10 corpus (rescored: false). No likelihood ratios, Heavy false-positive rate, or live self-scan deltas are published for 2026.11, because that re-score has not been run. Published npm packages (slop-detect-core, slop-detect, slop-detect-mcp) bump to 0.9.0 (pre-1.0, not a v1.0.0 release) and ship Apache-2.0 NOTICE attribution across all three tarballs.

• DESIGN.md lint and diff are callable. slop-detect --lint-design-md <file> and --diff-design-md <base> <head> run the engine checks with no browser. The MCP server adds local tools lint_design_md and diff_design_md. --init-tokens writes a DTCG tokens.json from the live DOM.
• Declared tokens waive matching design patterns. A scan that is given a parseable DESIGN.md keeps rawScore and reports the waived score as score. Craft is not waived.
• --a11y is optional and off by default. It needs the separate axe-core package, runs only in the local browser, and never changes a score or the fail-under gate.
• Scan documents carry schemaVersion 1. schemas/scan-result.v1.json is the draft 2020-12 contract for CLI --json, POST /api/scan, POST /api/aeo, and MCP structuredContent. A blocked or errored scan exits 2. That code means the scan could not run. --fail-on and --fail-under still exit 1 only when a page was scored and is over the gate. This schema is not a v1.0 release.
• slop-detect-mcp --local and remote Streamable HTTP /mcp. scan_page, check_aeo, check_design_system, and fix_prompt use the Playwright runner locally (--local) and do not call slop-detect.com. Hosted Streamable HTTP MCP is live at https://slop-detect.com/mcp. A Claude Code hook recipe scores SLOP_LOCAL_URL the same way.
• Site copy aggregation is callable. slop-detect --copy-dir <dir> and --site-copy <n> call scoreMultiPageCopy. The report sets rendered: false and pre2022HeavyRate: null. Fewer than 5 qualifying English pages is insufficient_pages with no site score. --site <n> renders same-origin pages when a browser is available, reports the median beside the homepage, and dedupes shared template selectors. The 50-word gate is not applied on that rendered path, because scan results do not carry a visible-word count. This does not change DEFINITIONS_VERSION.
• --brief prints a short avoid list. It names each triggered pattern and stays at or under 120 words. --fix-prompt is still the long remediation prompt. --diff-scan lists removed patterns and kit matches. fix_prompt with brief: true uses the short list. The default tool result does not.
• --format prints agent, github, sarif, junit, gitlab, or json. Review findings are warnings. A blocked scan is not a passing case. agent is selected when CLAUDECODE or CURSOR_AGENT is set and CI is unset.
• slop-detect init writes project files. Skill, MCP config, and optional hook recipes stay in the directory you pass. --scope=global is refused. Re-running does not duplicate a hook command. --no-hooks writes no hook files. Nothing is submitted to a registry.
• --emit-aeo-fixes writes starter files. robots.txt, llms.txt, ard.json, and link headers are byte-stable and written without a model. The training list is the installed registry, not a live mirror. Search crawlers do not get per-bot Allow lines.
• slop://spec/aeo lists the installed AEO checks. Likelihood ratios and the pre-2022 Heavy rate stay null.
• SARIF reports fingerprint each pattern and URL. One run per page. The category is the page URL. Remediation help comes from the existing fix text. Failed AEO checks are included. A selector already on the finding is a logical location. Source lines are not invented. A blocked page is a scan error, not a pass.
• compareScans refuses a definitions-version mismatch. That mismatch is not a score delta. --compare-to compares two saved scans offline, or one saved scan against a URL. --fail-on-regression exits 1 when the score rises above the allowed delta or a new pattern appears. A definitions change on the session baseline is a rescore, not a regression. --save-baseline writes one scored scan.
• --runs aggregates median or pessimistic scores beside the last scan. The page score and patterns stay the last run. runSummary holds the aggregate. This is not a measured CLI-versus-API identity proof. Both paths use a 1280×800 viewport. Chromium versions are recorded and are not asserted equal.
• A non-full preset reports fullScore. That field is the unfiltered design score. Copy, craft, AEO, and system are not rescored. The result names those axes when they were requested. The full preset leaves the JSON shape unchanged.
• MCP tools, prompts, and resources list the public favicon. Text tools return { text }. The SDK v2 plan is packages/mcp/SDK-V2.md. Nothing is published.
• Watch webhook payload builders exist for Slack, Discord, and JSON. A definitions change is rescored, not a regression. The sweep still sends email. Live delivery, retries, and disable-after-failure are not wired.
• Corrections inside unpublished 2026.11. Init merges each hook list and keeps existing commands. A declared color waives a rule only when it matches the observed surface or accent, and the score is recomputed with those rules excluded. Waived rows stay out of --brief, SARIF, and the fix prompt. A definitions rescore replaces the session baseline. --runs exits 2 when any attempt could not score. A saved --format file uses the same exit gate as a live scan. Mobile reflow runs at a phone width and is unknown when that pass cannot run. Off-screen text and unrelated focus rules are not craft failures. AEO uses the parsed preview points, splits combined X-Robots-Tag scopes, keeps the most restrictive max-snippet, requires server text or reports unknown, and judges the advertised markdown URL on its own response. A later valid markdown twin replaces an earlier HTML 200. A malformed alternate URL fails that check and does not abort the report. Robots * rules match without a backtracking expression. Multi-page copy strips shared headings and paragraphs with the body. List items keep their own text, and the blocks list is capped. Uncovered copy stays in document order with its quote, demo, and position flags, and copy checks read the full page text through those spans. Nested headings and paragraphs stay in those lists once. Inline spans stay in the same prose block. An inline quote keeps its flags and still counts toward the surrounding prose run. A chat demo does not, and short labels on either side of a block are not added together. Static HTML copy keeps links and image sources. Site discovery follows a redirect and stops reading after 200,000 characters. em line-height stays a ratio of that style's font size. Composite token aliases resolve, and an extended group keeps its type. The named binomial interval is Clopper-Pearson. A centered Inter hero keeps the capped typography points. The MCP package depends on the published slop-detect-core name and ships the scan schema beside the bundle. No new measured score deltas are published.

Unreleased — launch-readiness pass (2026-10-15)

Decisions: launch without live monitoring (waitlist), directory listings are verified-only with rel="ugc nofollow", tuning waits for a blind-labeled corpus.

Fixed

• Core package name in the docs. The engine is published to npm as slop-detect-core (core's prepack renames the @slop-detect/core workspace before packing; the scoped name does not exist on npm). The v0.8.0 note below that told consumers to switch to @slop-detect/core was wrong — keep depending on slop-detect-core.
• Documented core embed API works as written. assembleScanResult(raw, opts) now scores the page.evaluate payload (score/tier/grade/verdict, optional copy axis + unified score, blocked-page detection) instead of returning only pattern rows; buildPageEvaluateScript(patterns, opts) honours both arguments; getPatterns('strict') returns the filtered catalogue. The old call shapes still work. scoreAeo is documented as the async, fetching runAeoChecks(url, opts) it is.
• generateDesignMd keeps the whole observed palette (every CTA fill, heading colour, font and radius), so a page always passes against the DESIGN.md generated from it. spec/system.md now describes the comparator the engine actually runs (per-channel RGB tolerance, distinct-radii majority).
• Fix prompt AEO gaps use each check's message and severity (it read non-existent detail / required fields and printed generic advice).
• GitHub Action: a requested AEO audit (aeo: true) that cannot run now emits a ::warning instead of silently dropping the AEO row.
• Lockfile synced with the CLI's core-as-devDependency (bundled) layout.
• Apex→www redirects scan again. The scan endpoint refused any redirect to a different hostname as "private/internal" (a plain apple.com failed). Same-site redirects (apex/www/subdomain) are now accepted and recorded under the final host; genuinely private destinations and cross-site hops get distinct errors.
• npm release can no longer publish an uninstallable CLI. publish.yml packs with bun pm pack (rewrites workspace:*), asserts tarballs are workspace-free, install-smokes them, and publishes the tarballs with provenance. Manual runs are dry-runs by default.
• Detector grouping bugs (definitions@2026.10): icon_card_grid grouped cards by parent tag name across unrelated sections; numbered_steps fired on ranked lists (HN). Both now group by real parent/sibling structure.
• Ops metrics growth capped: unknown /api/* route labels fold into other instead of adding a key per bot probe; unknown API paths return a JSON 404 and unknown pages a real 404.html (was: homepage with 200). Bare /pricing, /terms, /privacy serve their documents.
• Unverified claims can't squat domains or publish links. A claim locks a domain and appears in /directory only after email verification; directory links are rel="ugc nofollow".
• Monitoring copy is honest: homepage, FAQ, README and llms.txt now say waitlist (baselines recorded, alerts at launch); the sweep workflow shows Skipped instead of a green no-op; the README watch example sends domain (it sent url and got a 400).
• Workflows: leaderboard job installs + builds core (it crashed monthly); calibration runs under pipefail (it "passed" while failing).
• Scan capacity: 10-minute KV result cache for repeat scans (refresh:true bypasses; the sweep sends it), browser tabs are closed before session reuse, and browser-capacity failures return a friendly retryable 503.
• Calibration harness: paced scans, failed scans counted in accuracy, JSON rows carry fired pattern ids, --check treats a labeled ERROR as a miss, and 10 hand-crafted reference candidates await blind labeling.
• Docs: @slop-detect/core README and the Agent Skill updated to the current API, 27 patterns, and tiers (Clean 0–9 · Mild 10–27 · Heavy ≥28).

definitions@2026.10 — detector grouping precision & corpus re-score (2026-10-20)

Published score deltas for DEFINITIONS_VERSION = "2026.10" (shipped in #217, re-scored in #222 / FIX-02) across the 24-site leaderboard corpus (apps/web/leaderboard/corpus.json) and the 23-entry calibration corpus (packages/cli/calibration/corpus.json).

Detector rule changes in 2026.10

• icon_card_grid (weight 4): now groups candidate cards by actual parentElement identity and then by width bucket within that parent, instead of keying by parent.tagName + ':' + Math.round(r.width / 20). Same-sized cards in unrelated sections across the page no longer merge into one phantom grid.
• numbered_steps (weight 3): now skips elements inside <table> or <nav> and ignores sibling groups with more than 6 numbered items, so ranked content lists (e.g. Hacker News story ranks, directory tables, nav breadcrumbs) no longer fire as marketing "1 · 2 · 3" step sequences.

Leaderboard corpus deltas (2026.09 → 2026.10)

The Before (2026.09) column comes from the previous committed apps/web/public/leaderboard.json (generated 2026-09-15T07:08:35.691Z, with per-pattern baselines recovered from its /r/<id> snapshots). The After (2026.10) column comes from a local Playwright Chromium re-score under this tree's 2026.10 engine (2026-10-20T12:08:59.687Z). Because these are live marketing URLs re-scanned a month later, deltas reflect the 2026.10 grouping fixes (icon_card_grid, numbered_steps), the centered_hero geometry/transform fix (T-17), and live-DOM changes on the target sites since 2026-09-15.

• Aggregate (leaderboard.json): 20/24 scored → 22/24 scored · mean score 15.7 → 14.9 · tier counts 7 Clean / 9 Mild / 4 Heavy → 7 Clean / 14 Mild / 1 Heavy · slopShare 65% → 68% (two previously unscored sites, vercel.com and replit.com, now scored in Mild).

Site (domain)	Category	Before (2026.09, 2026-09-15)	After (2026.10, 2026-10-20)	Δ	Pattern changes (- stopped firing, + started firing)
notion.so	bigtech	0 · Clean (A+)	0 · Clean (A+)	0	none (0 patterns)
anthropic.com	bigtech	0 · Clean (A+)	0 · Clean (A+)	0	none (0 patterns)
motherfuckingwebsite.com	classic	0 · Clean (A+)	0 · Clean (A+)	0	none (0 patterns)
craigslist.org	classic	3 · Clean (A)	0 · Clean (A+)	-3	-numbered_steps (-3)
figma.com	bigtech	3 · Clean (A)	3 · Clean (A)	0	unchanged (all_caps_labels)
berkshirehathaway.com	classic	3 · Clean (A)	3 · Clean (A)	0	unchanged (flat_type_hierarchy)
news.ycombinator.com	classic	6 · Clean (A-)	3 · Clean (A)	-3	-numbered_steps (-3); kept flat_type_hierarchy
gnu.org	classic	10 · Mild (B+)	10 · Mild (B+)	0	unchanged (gradient_backgrounds, accent_stripe)
bolt.new	ai-builder	31 · Heavy (D+)	15 · Mild (B)	-16	-slop_fonts (-8), -centered_hero (-4), -nested_cards (-4); kept glassmorphism, all_caps_labels, perma_dark_mode, bento_grid
linear.app	saas	23 · Mild (B-)	16 · Mild (B)	-7	-centered_hero (-4), -numbered_steps (-3); kept slop_fonts, perma_dark_mode, nested_cards
vercel.com	saas	unscored (HTTP 502)	17 · Mild (B)	—	newly scored: +slop_fonts (+8), +gradient_backgrounds (+4), +crushed_tracking (+5)
stripe.com	saas	19 · Mild (B)	19 · Mild (B)	0	unchanged (purple_accent, gradient_backgrounds, icon_card_grid, stat_banner)
supabase.com	saas	20 · Mild (B-)	20 · Mild (B-)	0	unchanged (slop_fonts, gradient_backgrounds, icon_card_grid, bento_grid)
posthog.com	saas	19 · Mild (B)	20 · Mild (B-)	+1	-numbered_steps (-3), +gradient_backgrounds (+4); kept purple_accent, glassmorphism, nested_cards
v0.dev	ai-builder	17 · Mild (B)	22 · Mild (B-)	+5	+hero_eyebrow_pill (+5); kept slop_fonts, centered_hero, crushed_tracking
resend.com	saas	34 · Heavy (D)	22 · Mild (B-)	-12	-icon_card_grid (-4), -numbered_steps (-3), -gradient_letter_avatars (-5); kept purple_accent, gradient_text, gradient_backgrounds, perma_dark_mode
framer.com	ai-builder	10 · Mild (B+)	23 · Mild (B-)	+13	+gradient_backgrounds (+4), +glassmorphism (+4), +gradient_letter_avatars (+5); kept gradient_text, perma_dark_mode
replit.com	ai-builder	unscored (cloudflare_challenge)	23 · Mild (B-)	—	newly scored: +gradient_backgrounds (+4), +glassmorphism (+4), +all_caps_labels (+3), +stat_banner (+3), +crushed_tracking (+5), +nested_cards (+4)
railway.app	saas	24 · Mild (C)	24 · Mild (C)	0	unchanged (slop_fonts, purple_accent, perma_dark_mode, icon_card_grid)
aura.build	ai-builder	25 · Mild (C)	25 · Mild (C)	0	unchanged (slop_fonts, glassmorphism, colored_glows, centered_hero, crushed_tracking)
cal.com	saas	31 · Heavy (D+)	27 · Mild (C)	-4	-centered_hero (-4); kept slop_fonts, gradient_backgrounds, hero_eyebrow_pill, numbered_steps, low_contrast_text
clerk.com	saas	36 · Heavy (D)	36 · Heavy (D)	0	-all_caps_labels (-3), -centered_hero (-4), +low_contrast_text (+7); kept slop_fonts, gradient_backgrounds, glassmorphism, icon_card_grid, aurora_mesh_gradient, nested_cards
lovable.dev	ai-builder	unscored (cloudflare_challenge)	unscored (cloudflare_challenge)	—	unscored (cloudflare_challenge)
openai.com	bigtech	unscored (cloudflare_challenge)	unscored (cloudflare_challenge)	—	unscored (cloudflare_challenge)

Calibration corpus tier-accuracy (2026.09 → 2026.10)

The Before (2026.09) column is the 2026-09-15 local calibration run recorded in CALIBRATION.md. The After (2026.10) column is the 2026-10-20 local calibration run (bun run calibrate -- --local) under DEFINITIONS_VERSION = "2026.10".

• Tier accuracy over the 13 confirmed-label rows (0 Heavy labels in seed):
  • Before (definitions@2026.09, 2026-09-15): 5/13 (38.5%) — confusion Clean → {Clean: 3, Mild: 6, Heavy: 0, null: 1}, Mild → {Clean: 0, Mild: 2, Heavy: 1}, Heavy → {0}.
  • After (definitions@2026.10, 2026-10-20): 7/13 (53.8%) — confusion Clean → {Clean: 4, Mild: 5, Heavy: 0, null: 1}, Mild → {Clean: 0, Mild: 3, Heavy: 0}, Heavy → {0}.

URL	Expected tier	Before (2026.09)	After (2026.10)	Δ	Pattern changes / triggered in 2026.10
news.ycombinator.com	Clean	Clean (6) ✓	Clean (3) ✓	-3	-numbered_steps (-3); fired flat_type_hierarchy
example.com	Clean	null (empty_page) ✗	Clean (0) ✓	—	IANA /s.js now renders ≥10 visible elements; 0 patterns fired
motherfuckingwebsite.com	Clean	Clean (0) ✓	Clean (0) ✓	0	0 patterns fired
berkshirehathaway.com	Clean	Clean (3) ✓	Clean (3) ✓	0	flat_type_hierarchy
www.gnu.org	Clean	Mild (10) ✗	Mild (10) ✗	0	gradient_backgrounds, accent_stripe
supabase.com	Clean	Mild (20) ✗	Mild (20) ✗	0	slop_fonts, gradient_backgrounds, icon_card_grid, bento_grid
bolt.new	Mild	Heavy (31) ✗	Mild (15) ✓	-16	-slop_fonts (-8), -centered_hero (-4), -nested_cards (-4); fired glassmorphism, all_caps_labels, perma_dark_mode, bento_grid
www.aura.build	Mild	Mild (25) ✓	Mild (25) ✓	0	slop_fonts, glassmorphism, colored_glows, centered_hero, crushed_tracking
v0.dev	Mild	Mild (17) ✓	Mild (22) ✓	+5	+hero_eyebrow_pill (+5); fired slop_fonts, centered_hero, hero_eyebrow_pill, crushed_tracking
lovable.dev	Clean	Mild (15) ✗	null (cloudflare_challenge) ✗	—	unscored (cloudflare_challenge), counted as a miss (Clean → null: 1)
stripe.com	Clean	Mild (19) ✗	Mild (19) ✗	0	purple_accent, gradient_backgrounds, icon_card_grid, stat_banner
linear.app	Clean	Mild (19) ✗	Mild (16) ✗	-3	-numbered_steps (-3); fired slop_fonts, perma_dark_mode, nested_cards
vercel.com	Clean	Mild (17) ✗	Mild (17) ✗	0	slop_fonts, gradient_backgrounds, crushed_tracking

Unlabeled candidate predictions in packages/cli/calibration/corpus.json under 2026.10 (label: null, not counted in accuracy): www.apple.com Clean (4: gradient_backgrounds), clerk.com Heavy (36: slop_fonts, gradient_backgrounds, glassmorphism, icon_card_grid, aurora_mesh_gradient, low_contrast_text, nested_cards), resend.com Mild (22: purple_accent, gradient_text, gradient_backgrounds, perma_dark_mode), cal.com Mild (27: slop_fonts, gradient_backgrounds, hero_eyebrow_pill, numbered_steps, low_contrast_text), www.raycast.com Heavy (37: slop_fonts, gradient_backgrounds, glassmorphism, centered_hero, all_caps_labels, perma_dark_mode, icon_card_grid, stat_banner, ai_sparkle_badges), www.notion.com Clean (0), www.figma.com Clean (3: all_caps_labels), cursor.com Mild (17: gradient_text, gradient_backgrounds, numbered_steps, nested_cards), replit.com Mild (23: gradient_backgrounds, glassmorphism, all_caps_labels, stat_banner, crushed_tracking, nested_cards), www.framer.com Mild (23: gradient_text, gradient_backgrounds, glassmorphism, perma_dark_mode, gradient_letter_avatars).

v0.8.0 — monorepo restructure: bun + turbo + apps/docs + spec/

A structural refactor. No detector behavior changed; tooling, layout, and the core package's npm name did.

Tooling

• bun replaces npm for install + script running (bun install, bun run build|test|typecheck|format|lint).
• turborepo orchestrates per-package build / test / typecheck via dependsOn: ["^build"] (no more handwritten chains in root scripts).
• vitest replaces node:test everywhere (core, cli, mcp, apps/web — 132 tests total, all green).
• tsup is the per-package bundler (dual ESM + CJS, .d.ts emit) for publishable libs.
• changesets for versioned releases (.changeset/ + bun run changeset → bun run version-packages → bun run release).
• CI rewritten on oven-sh/setup-bun@v2: lint (format:check), typecheck, unit tests, smoke (build + Playwright + scan HN). Deploy + publish + calibrate + leaderboard workflows likewise on bun.

Layout

• packages/web → apps/web (history preserved via git mv).
• New apps/docs — Next.js 15 + React 19 + Tailwind v4 docs site (bun run docs:dev).
• New examples/{astro-blog, nextjs-app-router} — reference framework integrations.
• New spec/ — versioned scoring spec (patterns.md, copy-axis.md, aeo.md, conformance.md, config.json). Downstream tools should read the spec, not the source.
• Per-package: composite tsconfig.json (project refs), tsup.config.ts, vitest.config.ts for every publishable lib.

Breaking

• The core workspace is renamed to @slop-detect/core inside the monorepo. Correction: this is a workspace-only name. The npm package is still slop-detect-core (core's prepack renames it before publishing) and @slop-detect/core is not published. Consumers should keep:

  +    "slop-detect-core": "^0.8.0"
  

  import { PATTERNS } from 'slop-detect-core';
  

• Root scripts now require bun (1.3+). npm install will fail because the lockfile is bun.lock. Install bun first: https://bun.sh.

Internal

• Apache-2.0 attribution for upstream AEO logic moved from inline packages/core/src/aeo.ts comments to NOTICE (matches how the Impeccable attribution is handled).
• prepack / pretest scripts removed from cli + mcp packages — turbo's dependsOn: ["^build"] handles build ordering.

v0.7.0 — the system axis & the continuity product

Strategy v2 implemented end to end ("slop is the hook, not the product").

Added

• System axis — DESIGN.md compliance: core parser/scorer (zero-dep), CLI --design-md, API designMd, MCP check_design_system. Relative, higher-is-better, named drift; structurally immune to the premium-site false-positive class.
• Drift alerts — watch { system: true }: the daily sweep checks each domain against its own DESIGN.md and emails once per drift event.
• Agency dashboard (/dashboard) — magic-link sign-in (HMAC sessions, anti-enumeration link endpoint), one view across every domain on an email.
• Client report (/report/<domain>) — print-friendly (PDF via print).
• Leaderboard (/leaderboard) + opt-in directory (/directory, dofollow backlinks) + monitor conversion section on the landing page.
• Calibration harness: golden fixtures, labeled corpus, npm run calibrate.

Changed

• Landing page: §04 monitoring conversion bridge, post-scan nudge, nav/footer to all surfaces; sub-pages unified on the forensic-instrument brand.
• All discovery surfaces (openapi.json — incl. fixing a wrong preset enum — llms.txt family, .well-known cards, agent view) synced to the full tool surface and version.

Fixed

• Detector calibration: eyebrow-pill no longer fires on CTA/nav buttons; glassmorphism requires >=2 elements (evidence: Stripe/Linear/Vercel scans).
• Monitor form could silently delist a domain on resubmit (opt-ins now additive).

definitions@2026.09 — calibration: fewer premium-site false positives

Driven by real evidence from the /leaderboard corpus scan (linear.app scored 36/Heavy, stripe.com 23/Mild — clear over-flagging of carefully-crafted sites).

• hero_eyebrow_pill: dropped the over-broad "weak" catch-all that fired on ANY rounded element above the hero — it misfired on pill-shaped CTA/nav buttons ("Sign up", "Log in"). Now requires the actual "Now in beta / Introducing …" keyword/emoji content tell.
• glassmorphism: now requires ≥2 frosted layers (was ≥1). A single frosted panel is common in otherwise-clean premium design; the tell is glass everywhere.
• Bumped DEFINITIONS_VERSION → 2026.09 (scores shift, so they stay comparable only within a version). Golden fixtures pin that real slop still trips both patterns while a clean page with a CTA pill no longer does.

NOTE: the font (Inter) and VibeCode-purple weights are unchanged — those are deliberate product opinions, not bugs, and re-weighting them is a separate labeled-corpus calibration decision (see CALIBRATION.md).

v0.6.0 — production-readiness pass

Monitoring + directory features plus a deep security/ops hardening pass toward a public launch. See PRODUCTION.md for the full launch checklist.

Added

• Monitored domains (/api/watch) — opt-in continuity layer: remember a domain and flag regressions; with consent recording + a privacy policy.
• Public directory (/directory, /api/sites) — opt-in, owner-gated catalogue of scanned sites with dofollow backlinks.
• CLI --remote mode — scan via the API with zero Playwright/Chromium install (fixes the npx slop-detect first-run cliff).
• Account-level cost guard — SCAN_DAILY_CAP + SCAN_DISABLED kill switch on the browser path; observability shim (_report.js, ERROR_WEBHOOK).
• Privacy policy (/privacy.md) + security headers (CSP, HSTS, nosniff…).
• Deploy workflow + a CI test job that actually runs the unit suite.

Fixed

• CI never ran the unit tests (lint + smoke only); now gated on npm test.
• CLI crashed on --help when Playwright was absent — Playwright is now lazy-loaded; the browser-free unit suite is 92/92 green.
• SSRF redirect bypass — scan + aeo re-validate every redirect hop / the final URL instead of trusting only the first.
• Version drift — all packages aligned to 0.6.0 (+ lockfile in sync).
• Deduped the design tier thresholds to a single source of truth.

Notes

• Regression-alert emails are not active yet (no sender wired): the trial captures consent only. Set up a provider + double-opt-in before enabling.

Web polish — impeccable-skill audit fixes + forensic-instrument typography (deployed)

Audited the landing page against the impeccable brand-design skill (shared design laws + brand register + audit rubric) and landing-copy, then fixed every finding. Web package only; still self-scans 4/100 Clean, Grade A.

Absolute / brand bans fixed

• Removed both side-stripe border-left accents (an absolute ban) on .step and .fix-cta; replaced with full borders + a registration-chip marker and an accent-tinted panel.
• Stripped every em dash from rendered copy (title, OG/Twitter meta, hero lead, body, pipeline, footer, and the JS rules-list separator). Colons/periods now.
• Replaced pure #000 (button ink) with a tinted near-black token.

Escaped the editorial-typographic reflex lane

• Killed the // comment-prefix eyebrows. Sections now carry a numbered lab-report registration system (§01 / §02 / §03) with lowercase labels.
• Real font pairing via the brand font-selection procedure (reflex picks Inter / Geist / IBM Plex Mono are exactly what this tool flags, so they're rejected): Hanken Grotesk for prose, Martian Mono for instrument readouts. Mono tracking pulled in hard (Martian ships wide) so it reads as precision, not decoration.
• Fluid clamp() type scale (>=1.25 ratio); neutrals retinted toward the brand blue hue (no pure greys); accent-soft tint token introduced.
• Catalogue (§03) rebuilt from a plain collapsed <ol> into a dense 2-column numbered grid (still driven by the live /api/patterns count, never hardcoded). Disclosure toggle is now a + / − chip matching the registration marks.

Dogfooding (the important part)

• The first redesign pass scored 10/100 Mild because two new moves tripped the tool's OWN detectors: numbered 01/02/03 marks fired numbered_steps, and wide-tracked uppercase labels fired all_caps_labels. Rather than game the detector, the fix eats its own dogfood: labels went lowercase, and the marks became §-prefixed section sigils (no longer a 1·2·3 sequence). Back to Clean.
• Sticky nav made fully opaque (no backdrop-filter blur — the engine flags glass) and section anchors got scroll-margin-top so jump-links never clip under it.

Web redesign — "editorial terminal" landing page (deployed)

Full redesign of the slop-detect.com landing page (web package only; not an npm release). The old page was a single left-drifting column that trailed off into dead space and read as an unfinished prototype.

Design

• New structure: sticky topbar, two-column hero (headline + scan form left, a static "sample scan" readout card right), an editorial two-column "fingerprint" section, a carded "how it works" pipeline, the live rules disclosure, and a real multi-column footer. The void is now filled with intentional sections + rhythm.
• Single restrained cold-blue accent (#5b9dff); monospace used for data/labels.
• Strong, non-flat type hierarchy; raised body/label contrast for AA legibility.
• Score count-up animation; result card reveal; "browse every rule" now wired.

Anti-slop by construction

• The page is scored by its own engine, so it deliberately avoids every tell it detects: no gradient text, no gradient/aurora backgrounds, no glassmorphism, no colored glows, no centered hero, no eyebrow pill, no numbered-step circles, no icon-card grid, no big-number stat banner, no crushed/wide tracking.
• Self-scan result: 4/100, Clean, Grade A (only the unavoidable perma-dark + grey-body tell fires). The landing page passes its own detector.

Safety

• The entire 370-line core script (Turnstile, /api/patterns, render, share/embed, fix-prompt modal) is preserved byte-for-byte; every element ID is retained. The redesign is HTML structure + CSS only. New gutter mirror labels stay dynamic by observing the live count (never hardcoded). Verified live + on mobile.

0.5.2 — Rate-limit fail-closed, foreign-origin rejection, dead-code sweep

Second post-review batch. Web changes deployed to slop-detect.com; CLI republished to keep npm in sync with source (core + mcp unchanged at 0.5.1).

Security (web — deployed)

• Rate-limit now fails CLOSED on the scan route (#5). Previously a missing RATE_LIMIT KV binding, or a KV read/write outage, let the expensive headless- browser scan route through unthrottled. Now: when KV is unavailable, the scan route enforces a tight in-memory per-isolate ceiling (3/window). Cheap routes (fix-prompt assemble) still fail open — they don't touch the browser.
• Foreign browser origins are now rejected (#6). The middleware comment claimed it "blocks third-party origins" but only down-bucketed them. It now returns 403 origin_not_allowed for any browser Origin outside the allowlist — unless the caller presents a valid API key. No-origin callers (CLI/curl) are unaffected.

Cleanup

• Removed 3 dead ctx helpers (isVisible, inViewport, isNeutral) that were wired into the page-eval context but consumed by no pattern. inHero is kept (the declarative rules engine uses it). Affects CLI + web scan context.

Tests

• +7 middleware tests (foreign-origin rejection, scan fail-closed under missing / broken KV, cheap-route fail-open, OPTIONS preflight). 35 total, all green.

0.5.1 — Security hardening + CI gating (post-review)

From a full end-to-end review. Resyncs npm with the source tree (published 0.5.0 core shipped 26 patterns; the tree had 27) and ships security + DX fixes.

Security

• SSRF guard on /api/scan (validateScanUrl): blocks cloud-metadata (169.254.169.254), loopback, RFC-1918, CGNAT, link-local/unique-local IPv6, IPv4-mapped loopback, and internal hostnames (localhost, *.internal, *.local); rejects non-http(s) schemes. The scan endpoint loaded arbitrary user URLs in a headless browser and returned page content — this closes the SSRF primitive.
• fix-prompt {url} bypass closed: mode-2 fix-prompt (which re-runs a scan in-process) is now gated by the middleware as a scan — same rate limit, shared bucket, and Turnstile — instead of fix-prompt's looser limits.

Added

• GET /api/patterns — serves the live pattern catalogue from core. The web UI and CLI now render the rule count + list dynamically (no more hardcoded "16 rules").
• CLI --fail-on <tier> (mild|heavy) — exits non-zero to gate CI; blocked/errored scans also fail. Replaces the manual jq-based gate in the docs.

Fixed

• CLI now normalizes bare hostnames (slop-detect example.com) and rejects non-http(s) schemes with exit 2.
• MCP server reports its real version from package.json (was hardcoded 0.2.0).
• MCP API client has a 45s fetch timeout (SLOP_DETECT_TIMEOUT_MS); timeouts surface as a clean ApiError.
• Tightened the Chromium-missing detection regex (was matching any launch failure).
• Corrected the CLI User-Agent repo URL and the CLI README tier thresholds (Clean <10, Mild ≥10, Heavy ≥28).

Tests

• +3 SSRF suites, +5 CLI subprocess tests, +5 MCP api tests (28 total, all green).

0.5.0 — Phase 5: Impeccable adoption (Tranche A)

definitions@2026.08 — the design fingerprint grows from 19 → 27 patterns.

Added

• 8 new design patterns ported from Impeccable (Apache-2.0), targeting high-prevalence AI tells we were missing:
  • cream_default_bg (w7) — warm off-white / beige default page surface (~74% of generated pages per Impeccable's launch data)
  • low_contrast_text (w7) — washed-out grey body text below WCAG AA on a light background. Deliberately narrowed to the pervasive grey-on-light signature (not generic WCAG failure) so premium sites with intentional brand color stay clean
  • crushed_tracking (w5) — display headings tracked tighter than -0.05em
  • gray_on_color (w4) — neutral mid-grey text on a chromatic background
  • oversized_hero_h1 (w4) — long headline (≥40 chars) blown up to ≥72px
  • nested_cards (w4) — card-like element inside a card-like ancestor (excludes 3D-tilted product-screenshot mockups)
  • wide_body_tracking (w3) — body copy letter-spacing above 0.05em
  • flat_type_hierarchy (w3) — ≥3 font sizes with a max/min ratio below 2×
• WCAG color primitives in createColorHelpers(): relativeLuminance, contrastRatio, channelSpread, isNeutral, effectiveBackground (ancestor background-color walker). Unit-tested against known WCAG reference pairs.
• Fix recipes + evidence formatters for all 8 new patterns.
• Second unit-test suite (color.test.js) — 7 tests on the contrast math.

Calibration

• Validated against premium reference sites (Stripe, Linear, Notion, Vercel) to confirm the new rules do not false-fire, and against a synthetic slop fixture to confirm they do fire. Contrast/gray rules required narrowing to the true AI signature after initial false positives on branded CTAs and white-on-dark text.

Attribution

• Patterns #20–27 adapted from Impeccable by Paul Bakaus (Apache-2.0). See NOTICE.

Backward compatibility

• Existing pattern IDs, weights, and the design/copy axis split are unchanged. New patterns are additive; the tier bands (Clean/Mild/Heavy) are unchanged.

0.4.0 — Phase 4: multi-axis slop (design + copy)

Added

• Copy-slop axis (#08) — 9 deterministic text patterns scored on the page's own prose at near-zero marginal cost (one in-DOM text extraction, no extra page load): buzzword_density, em_dash_overload, antithesis_construction, filler_openers, formulaic_closers, rule_of_three, whether_youre, unicode_artifacts, emoji_bullet_headers. Each with its own fix recipe.
• Multi-axis scoring — scoreCopy(), combineAxes(), extractTextContext(), AXES exported from core. Per-axis summaries (axes.{design,copy}) plus a unifiedScore / unifiedTier (max axis score + penalty when >1 axis dirty).
• CLI: --copy and --axes design,copy|all flags, with a per-axis pretty breakdown. API: { axes: ['design','copy'] } on /api/scan and /api/fix-prompt. Permalinks persist a compact copy-axis summary.
• buildFixPrompt now spans both axes when the copy axis was scored.
• First unit-test suite (node --test) covering the copy axis + aggregator; npm test wired at the root.

Backward compatibility

• Top-level score / grade / tier / patterns remain design-only by default. The copy axis is strictly opt-in; existing integrations are unchanged.

0.3.0 — Phase 3: keep the ruleset current

Added

• Declarative rule format (compileRule / compileRules / validateRule) — describe a pattern as { id, label, weight, detect: { scope, when, trigger } } and it compiles to a self-contained, injectable detector. Lowers the cost of contributing the 20th/21st pattern. Imperative patterns still supported. (#05)
• Named scoring presets — full (default), strict (weight ≥ 5), marketing (brand surface), minimal (3 dead-giveaways). Selectable via the CLI --preset flag and the API { preset } field; applyPreset exported.
• Public API key tiers — Authorization: Bearer / X-API-Key keys with free / pro / unlimited tiers (key-bucketed limits, Turnstile bypass). Fully documented in apps/web/API.md.

Definitions 2026.07

• +3 emerging patterns (max raw weight 73 → 85): bento_grid (w4), aurora_mesh_gradient (w5), ai_sparkle_badges (w3). Each with a fix recipe and evidence formatter. Verified for no false positives on Hacker News (Clean).

0.2.x — Phase 2: embed into the workflow (CI + agents)

Added

• slop-detect-mcp — Model Context Protocol server (stdio) exposing scan_page and fix_prompt tools so Cursor / Claude Code / Windsurf agents can self-audit a page before shipping. Thin wrapper over the public API, SLOP_DETECT_API configurable. (#07)
• slop-detect-action — composite GitHub Action that scans a deploy-preview URL, posts a sticky PR comment (grade · score · triggered patterns · OG card), sets a pass/fail status via fail-under (numeric or letter grade), and exposes score/grade/tier/verdict/result-url outputs. (#04)
• Example consumer workflow at .github/workflows/slop-check.example.yml.

0.2.0 — Phase 1: distribution primitives

Make the score travel. Every scan now produces something shareable and embeddable. See ROADMAP.md and apps/web/UX.md.

Added

• Letter grades (A+→F) layered on the 0–100 score, on every surface (core, CLI, web, API, badge). gradeForScore() + GRADE_BANDS exported from core.
• Verdict one-liners — deterministic, page-not-person, tier-keyed. verdictFor().
• Versioned slop definitions — every result carries definitionsVersion (2026.06) so historical scores stay comparable. DEFINITIONS_VERSION exported.
• Shareable result permalinks — POST /api/scan returns id + resultUrl; GET /r/<id> server-renders the result with OG/Twitter meta.
• OG share cards — GET /og/<id>.png renders a 1200×630 card (grade · score · tier · verdict · tells) via Browser Rendering, cached in KV.
• Embeddable badge — GET /badge/<domain>.svg returns a live, tier-colored shields-style badge for a domain's latest scan. Copy-paste Markdown/HTML/URL snippets in the UI.
• Share + embed UX — share panel (copy link, X, LinkedIn) and badge embed panel on the result view; /?url= deep-link prefill for the re-scan loop.

Infra

• New RESULTS KV namespace (results, OG cache, domain→latest, badge cache).

Definitions 2026.06

• Baseline: the original 16 patterns (Krebs 15 + Meng To 1). No pattern changes; establishes the versioning scheme going forward.

0.1.2

• fix-prompt rewrite: concrete per-page evidence + anti-regression list.
• Tightened empty-page detection for sparse-DOM sites.

0.1.1

• Detector tuning, anti-bot guard, rate limiting, branding.

0.1.0

• Initial monorepo — core + cli + web. 16-pattern engine.