Skip to content

ROADMAP.md

Latest commit

 

History

History
160 lines (126 loc) · 8.01 KB

ROADMAP.md

File metadata and controls

160 lines (126 loc) · 8.01 KB
Raw

slop-detect: Strategic Roadmap v2

Where this project goes next, and why. Last updated: 2026-06-05 · Current version: v0.6.0 Supersedes the v1 roadmap (preserved in git history). Grounded in the June 2026 deep-research pass, see "Research basis" at the bottom, and in our own calibration evidence (CALIBRATION.md).

1. The thesis (revised)

Slop is the hook, not the product.

Three research findings force the revision:

  1. The fixed visual fingerprint decays in ~6–18 months. Signature-style detectors reliably degrade (antivirus, AI-text detectors), and the generators are being actively engineered to suppress the default look: Google's DESIGN.md standard (Apache-2.0, ~15k★, ships lint/export tooling) injects per-brand constraints into AI builders and agents, the design-world equivalent of malware polymorphism. The thing we detect is being standardized away.

  2. Absolute slop-grading is commoditized. Impeccable (~34.7k★, free, OSS) owns the "most patterns" game; Sailop, aiwebsitedetector et al. fill the rest, all enumerating the same tells. A rule list is copyable in an afternoon. Meanwhile "slop" the meme peaked (2025 Word of the Year). The attention is a depreciating asset to harvest, not a moat to defend.

  3. Our own evidence shows the fatal failure mode. The detector scores Linear/Vercel as Heavy and Stripe as Mild, punishing a style as provenance, the exact false-positive trap that killed OpenAI's own AI-text classifier and discredited the text-detector industry (Stanford/Cell). An absolute "is this AI?" verdict cannot be made reliable; a relative question can.

So the strategy inverts:

Free, loud slop grader = acquisition channel (now, while attention peaks). Paid product = continuous design-system / brand-drift monitoring: "your non-designers and coding agents keep shipping to your site; we keep it on-brand and on-system", sold to agencies and teams as recurring continuity. Technical posture: ride DESIGN.md, don't fight it. Compliance with a site's own declared system is a relative signal that cannot decay and structurally cannot false-positive on great custom design.

2. The product ladder

acquire (free, peak-hype)          retain (paid, durable)
┌──────────────────────────┐       ┌──────────────────────────────────┐
│ slop grader (web/CLI/MCP)│  ───► │ monitored domains (built, v0.6)  │
│ leaderboard + directory  │       │ + design-system drift (NEW axis) │
│ badges + share cards     │       │ + agency multi-site dashboards   │
│ AEO axis (free funnel)   │       │ + white-label PDF reports        │
└──────────────────────────┘       │ + CI gate on system compliance   │
                                   └──────────────────────────────────┘

The continuity layer (watch/alerts/directory) shipped in v0.6.0. The missing piece, and the first deliverable of this roadmap, is the relative axis that makes monitoring worth paying for:

P1: The system axis: DESIGN.md compliance (NOW)

"Does this page honor its declared design system?" Parse a site's DESIGN.md (Google Labs spec: YAML front-matter tokens, colors, typography, rounded, spacing, components), compare against what the scanned page actually renders (fonts in use, CTA/surface colors, radii), and report drift as named, contestable signals. Ships in core (pure, zero-dep), CLI (--design-md), and the web API. This:

  • fixes the Linear/Stripe false-positive class structurally: a bespoke site checked against its own tokens scores aligned, not "slop";
  • cannot decay: the reference point is per-customer, not a global fashion;
  • aligns us with the standard that would otherwise obsolete us.

P2: Sell continuity to agencies (NEXT, 1–2 quarters)

The research is unambiguous about the indie buyer: digital agencies pay recurring for white-label reports, scheduled multi-site audits, and drift alerts ($29–$149/mo band; flat-rate beats per-site at 10–20 client sites). Build: multi-domain dashboard, branded PDF export, drift-alert emails (the pipeline shipped in v0.6), CI gate on system compliance. Open-core stays MIT (Plausible/Sentry/Semgrep model); charge for continuity, collaboration, scale. Realistic target: $1–10K MRR within 12–18 months, plus $800–4K/mo well-structured GitHub Sponsors.

P3: Versioned community ruleset ("slop definitions") (ONGOING)

Keep the free fingerprint current on the Semgrep registry model: community-contributed, versioned, months-cadence refresh; favor structural tells (badge-above-H1, stripe borders) that outlast color fashion; track DESIGN.md/builder updates as the leading decay indicator. Calibration (CALIBRATION.md, labeled corpus, honest precision/recall) is strategic, not housekeeping: signals-not-verdicts is the credibility model.

P4: Agent-loop depth (ONGOING)

The MCP server's job shifts from "scan after the fact" to "the design-system check inside the coding-agent loop before it ships", consuming the repo's DESIGN.md. Workflow lock-in is the only durable distribution.

3. What we will NOT do (research-backed guardrails)

  • No ML/vision/LLM-judge as the core engine. Per-call cost, OOD collapse (18–30% on current generators), and the false-positive credibility crisis. At most: an opt-in pairwise LLM critique layer; never the scoring authority.
  • No head-on AEO/citation-tracking business. The monetizable core is owned by funded startups (Profound ~$1B) and Semrush/Ahrefs; our crawlability/ llms.txt axis is the commoditized free slice; keep it as funnel only.
  • No "% AI" verdicts, ever. Named rules, contestable signals, page-not-person.
  • No one-time licenses. (Tailwind UI's collapse is the cautionary tale.) Recurring only.
  • No rule-count arms race with Impeccable. Differentiate on continuity + compliance + workflow, not pattern count.
  • Engine stays MIT, free, offline-capable. Forever. (Unchanged from v1.)

4. North-star metrics (revised)

Metric Why
Monitored domains (free → paid) The business
Agencies on multi-site plans The buyer the research says converts
Pages checked against a DESIGN.md Adoption of the durable axis
Definitions version freshness / community rules merged Anti-decay
Scans/week + leaderboard/directory traffic Top of funnel (harvest the hype)

5. Sequencing

Move Status
P1 system axis (core + CLI + API) ✅ Shipped
P2a Drift alerts on system axis via existing sweep ✅ Shipped (watch { system:true } → sweep checks DESIGN.md → drift email, once per event)
P2b-lite Print-friendly client report (/report/:domain) ✅ Shipped (the PDF-infra substitute)
P2b Agency multi-site dashboard ✅ Shipped (lite): magic-link sign-in → /dashboard, one view across every domain on an email. Team seats/roles later
P2c Pricing live ($29–$149/mo) With P2b (needs Stripe keys, operator task)
P3 Community ruleset + calibration corpus growth Ongoing
P4 MCP design-system mode (check_design_system tool) ✅ Shipped

Research basis (June 2026)

Deep-research pass across market, shelf-life, AEO, monetization, and technical frontier. Key citations: DESIGN.md spec + adoption (github.com/google-labs-code/design.md); Impeccable (impeccable.style); fingerprint-decay precedents (arXiv 2603.23146, SentinelOne signature-vs-behavioral); false-positive precedent (Stanford/Cell Liang & Zou; OpenAI classifier shutdown); MLLM-as-UI-judge limits (arXiv 2510.08783); AEO incumbency (Fortune: Profound $96M Series C; Semrush/Ahrefs AI visibility); indie monetization (Plausible $1M-ARR write-up; Semgrep registry; agency audit-tool pricing; Tailwind UI decline, devclass Jan 2026).