Hero reframe to the inversion. New ICP landings, manifesto, FAQ, quickstart, changelog.
Pivot the homepage from feature-led to outcome-led. The H1 now reads "AI that knows your users. Without knowing your users." and the page leads with the inversion: cloud AI gets smart by sending data away, kolm gets smart by leaving it there.
- New trust-split section on the homepage proves the inversion visually. What cloud AI sees vs what kolm sees, side by side.
- New compared-to table against llama.cpp, MLC.ai, ExecuTorch, CoreML. kolm sits one layer above the runtime, the way Webpack sits above V8.
- New ICP pages: /healthcare, /finance, /defense. Each with quantified pain, reference deployment, compliance posture row.
- New /manifesto ("The case for compiled AI") establishes the category in long form.
- New /quickstart ships a 7-step Hello World tutorial. Install through bundle for ios/wasm.
- New /faq answers the questions developers, security, and procurement actually ask.
- New /changelog (this page) signals active development.
Hero outcome lede, Why-now pain quanta, em-dash audit, SW cache bump.
Replaced the prior feature-led hero with the outcome-led version, added the Why-now strip with quantified pain across latency, cost, data egress, and vendor risk, and audited body copy across 6 pages to enforce the no-em-dash rule.
4-tier pricing, ICP grid, 5-card moat, /security trust chain, /docs entry, manual, privacy, terms, 404.
The locked v5 design rebuild. New pricing structure (Developer / Mobile / Pro / Enterprise), ICP grid on the homepage, five-layer moat narrative, security trust chain on a dedicated page, full manual, privacy and terms pages, custom 404.
- Pricing reset: Developer $0 / Mobile $9 / Pro $49 / Enterprise custom.
- Trust chain page: HMAC chain explained step by step, public registry anchor, third-party verifier protocol.
- Manual: full reference for the four engines (Distill, Decompose, Recall, Run).
- Locked design system:
--bg #0a0a0a,--accent #7c8cff, system fonts, square 6px CTAs.
RS-1 spec at v0.1, manifest schema and receipt schema published.
Published the RS-1 (Recipe Spec 1) JSON Schemas for the manifest and the receipt chain. MIT license. Hosted at /docs/manifest-v0.1.json and /docs/receipt-v0.1.json. Spec markdown at /docs/rs-1.md.
- Schemas are stable for v0.1.x; breaking changes will land in v0.2.
- Receipt schema includes the HMAC chain anchor field for public-registry verification.
- Manifest schema includes K-score breakdown, base hash, adapter hash, recipe hash, recall hash, eval hash.
On-device PWA at /device, browser-native SDK at /sdk.js, service worker, manifest.
The first runtime surface that runs entirely in a browser. Loads a quantized base via wllama.wasm, sqlite-vec for the recall index, and the recipe pack. Installable as a PWA. Service worker caches everything for offline use.
Live smoke battery at 128/0 against production.
The full smoke battery (128 tests covering compile, run, registry, receipts, security headers, rate limits, auth, and SDK) passes against the live Railway deployment with zero failures. Reproducible via URL=https://kolmogorov-stack-production.up.railway.app bash scripts/smoke-live.sh.
CSP rolled out, helmet ahead of static, scrypt password hashing, OAuth tokens AES-256-GCM at rest.
Round-one hardening to remove the C+ blockers. Helmet middleware now runs before express.static so HSTS and CSP cover static assets. Passwords rehash to scrypt opportunistically on login. OAuth tokens encrypted at rest with backward-compat decrypt and lazy re-encrypt on refresh.
Verified Inference live on the production tenant. K-sample → deterministic verify → labels.
The Verified Inference endpoint runs k-sample frontier inference, verifies each sample with a deterministic synthesized verifier, and returns the winning label with a receipt. First step of the Distill pipeline.
SWE-bench Lite breakout: +15.33pp strict, n=150, p<0.05.
Reproducible improvement of Opus-4.7 on SWE-bench Lite via REM hooks loop. 26 recovered vs 3 regressed. Apply-errors down 48%. Methodology and scripts at github.com/sneaky-hippo/kolmogorov-stack.
kolm reframed as the AI compiler. Four engines (Distill, Decompose, Recall, Run) under one orchestrator.
The strategic pivot to "specialization-as-a-service": train with any frontier model, deploy to any local model. Four engines under kolm compile, three surfaces (CLI, Cloud, Mobile App), one artifact format.