Repository navigation
Releases
Releases · kitshelf/scamcheck
ScamCheck v0.1.0 — the jobs family, end to end
[v0.1.0] — 2026-11-20
First release: the jobs family, end to end.
Shipped
- The screen — the six-step workflow (redact → family → markers → label → verify → respond) in
screen/, withSPINE.md: the whole kit in one paste, so the quick start is two pastes. - The jobs rubric — 20 markers across seven categories (
screen/rubrics/markers-jobs.md), each with a not-a-marker-when calibration line and a dated "Verified as of" header; sourced on FTC / FBI-IC3 / CISA reporting and tuned against the vendored corpus (counts incorpus/TUNING.md). - Shared material — the three labels, the five families, the redaction pass, the recovery branch ("if you already paid": recall, dispute, freeze — channels only, no blame), and the local-model guidance with the honest model-floor note.
- Tests — 20 synthetic fixtures across four types (scam, near-miss, injection, behavior) with expected keys written first and dated recorded runs. Baseline: GLM-5.3, 20/20; adversarial release gate: 4/4.
- Worked examples — Priya (the catch) and Sam (the not-catch): the calibration story made public.
- User guide — the ninety-second walkthrough for a non-technical reader, with the before/after redaction sample and the limits section.
- Research corpus — 21 vendored public datasets under redistributable licenses plus a 1,051-dataset index (
corpus/): the learn-from library. Never fixtures, never runtime, never trained on. - Alpha marker packs — rentals, banking, marketplace, romance (
packs/): researched, sourced, calibration-lined; a family ships only when its gauntlet passes. - Docs — the PRD (why / what / nine CUJs) and
docs/learnings/: the grounded record of what development, the corpus, and model testing taught.
Model matrix (recorded, not asserted)
GLM-5.3 (hosted): 20/20. Qwen3.5-9B-MLX-4bit (local OMLX, thinking disabled): 4/20 — below the documented floor. Qwen3.7-Flash (hosted, thinking on): 1/3. Qwen3.8-Max-Prime (hosted, thinking on): 10/13, testing closed by user order. Full analysis: docs/learnings/03-local-model-testing.md.
Compatibility
Any hosted chat or local model (Ollama, LM Studio, llama.cpp, OMLX/MLX): paste screen/SPINE.md, then the redacted message. Fully offline; nothing leaves the machine.