Skip to content

Commit b2f7f20

Browse files
sneaky-hippoclaude
andcommitted
fix(verticals): revert v2 over-promotion — keep conversion-wired v1 canonicals + SOTA foundation
The v2 promotion was wrong for a conversion-first overhaul: the v1 vertical pages are the CONVERSION-INSTRUMENTED canonicals (wave889-8: /account/signup?industry, /book-demo, distill·run·verify trio, receipt CID, wave889-vertical.css). The richer-but-unwired v2 content replaced them and broke wave889-8 #3/#4/#5/#12. Richer != better-converting. - defense/finance/legal restored to v1 (conversion-wired); -v2 stay redirect stubs -> base (dedup preserved, healthcare-v2 precedent). devtools (/devtools) net-new kept. - defense gains a real /self-host air-gap CTA (on-point conversion path for defense buyers); wave272 #15 now checks the real /self-host page. - Foundation for the exhaustive upgrade: docs/website-overhaul/SOTA-STANDARD.md (per-page gate) + positioning.json (canonical strings + banned terms; generators import it). Green: wave272 18/18, wave889-8-12 12/12, wave889-8-34 12/12, site 7/7, wave538 6/6, wave225 14/14. Remaining suite reds are pre-existing/environmental (CLI --doctor timeouts, ship-gate meta-tests need prod, account/models.html marker + /v1/serve/pods — folded into the account upgrade). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1 parent 8ccc675 commit b2f7f20

6 files changed

Lines changed: 640 additions & 741 deletions

File tree

Lines changed: 52 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,52 @@
1+
# kolm — SOTA Standard (the bar every surface must meet)
2+
3+
The goal is not fewer pages. It is that **every** surface that earns its place is upgraded to read like a $25–60M category-defining company: current, on-brand, jargon-free, intuitive, conversion-oriented, simple. Keep what's useful (programmatic-SEO `compile/*`, docs, verticals all stay); make all of it SOTA. This file is the standard the audit measures against and the generators must encode.
4+
5+
## 1. Positioning — one source of truth (use these strings verbatim everywhere)
6+
7+
- **Category:** The AI Compiler.
8+
- **One-liner:** Own your AI. Compile frontier-model quality into a small, private model you run anywhere — with a signed receipt for every call.
9+
- **Three pillars (always in this order):** **Own it** (distill frontier quality into a small model you run anywhere) · **Prove it** (a signed receipt for every call — the moat) · **Govern it** (SSO/SCIM/RBAC/BYOC/residency/audit).
10+
- **The moat sentence (the thing no competitor ships):** a cryptographic, offline-verifiable receipt for every call.
11+
- **Never** lead with internal nouns: "wrapper", "studio", "surface", "evidence-to-artifact". Name the outcome; the mechanism comes after.
12+
13+
A machine-readable copy lives at `docs/website-overhaul/positioning.json` — generators import it; no page hard-codes a competing tagline.
14+
15+
## 2. Brand system (standardize; one look)
16+
17+
- **Palette:** cool-slate monochrome — ink `#111` on slate `#f3f5f7`, dark `#0e1116`. ONE accent (precise cobalt) reserved for the moat (receipts/verify/primary CTA), never decorative. (Accent token rollout is the screenshot-verified pass; tokens already centralized in `ks.css` `--ks-accent*`.)
18+
- **Type:** Inter (display/UI) + JetBrains Mono (every hash, receipt, command — it reads "cryptographic"). No serif.
19+
- **Components:** one nav (Product · Proof · Docs · Pricing · Enterprise), one footer, one card, one button hierarchy, one receipt artifact, one code block. Defined in the shared CSS; pages compose, never re-skin.
20+
- **Motion:** one reveal curve, one duration scale; all gated on `prefers-reduced-motion`.
21+
22+
## 3. Copy voice (from `voice-rubric.md`, enforced)
23+
24+
Lead from the reader's situation, not the mechanism. Outcome-first. One idea per sentence. Every number bounded + linked to data. No filler ("powerful/seamless/unlock/leverage/robust/world-class"); never the word "honest". Real contact only (`dev@kolm.ai`); no fabricated traction; SOC 2 = "Type I; Type II in progress" only.
25+
26+
## 4. Per-page SOTA checklist (the gate)
27+
28+
A page is SOTA only if ALL hold:
29+
1. **On-brand:** shared tokens/components; no warm-paper remnant; mono for hashes/commands.
30+
2. **On-positioning:** uses the canonical category/one-liner/pillar language; no contradicting tagline.
31+
3. **No jargon / current:** no wrapper/studio/surface/evidence-to-artifact; current model names + claims; nothing stale.
32+
4. **Conversion-oriented:** exactly one clear primary CTA above the fold; a secondary path; proof near the claim; an obvious next step (no dead ends).
33+
5. **Intuitive IA:** canonical nav + footer; correct breadcrumb/canonical; the reader always knows where they are and where to go.
34+
6. **Simple + scannable:** one big idea per section; short sentences; skimmable headers; no wall-of-text; meta title 12–78 / description 50–220.
35+
7. **Accessible + responsive:** skip link, focus rings, alt text, heading order; clean at 390 / 820 / 1440, light + dark; no horizontal overflow.
36+
37+
## 5. Conversion model (every surface routes to value)
38+
39+
- Each persona has a friction-minimal path: dev → `/quickstart` (own a model free); team → `/pricing`/`/product`; enterprise → `/enterprise` (talk to sales) + `/proof` (verify).
40+
- Programmatic-SEO pages (`compile/*`, `compare/*`) must convert the long-tail visitor: a one-line answer to their query, then the canonical CTA into the product. Not dead SEO leaves.
41+
- Proof is one click from any claim (`/verify`, the benchmark JSON, `/proof`).
42+
43+
## 6. How this gets applied (leverage, not hand-edits)
44+
45+
- **Generators encode the standard:** `build-account-pages`, `build-seo-pages`, `build-comparison-seo`, `build-marketplace-pages`, `build-docs-*`, `build-cli-docs` import `positioning.json` + emit pages that pass §4. Fix the generator → regenerate → the whole family is SOTA.
46+
- **Shared CSS** carries the brand; pages inherit.
47+
- **Hand-authored spine** (home/product/proof/solutions/pricing/enterprise/security) already upgraded; held to the same checklist.
48+
- **Every change keeps its family's tests green** (grep `tests/` first); generators are re-run idempotently; `sw.js` bumped.
49+
50+
## 7. Done = exhaustiveness gate
51+
52+
Reconcile the page manifest: every kept page passes §4; positioning strings are consistent site-wide; no jargon token survives a global grep; every family's tests green; screenshots clean at 3 breakpoints × 2 themes. That is "all of it, upgraded."
Lines changed: 24 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,24 @@
1+
{
2+
"category": "The AI Compiler",
3+
"one_liner": "Own your AI. Compile frontier-model quality into a small, private model you run anywhere — with a signed receipt for every call.",
4+
"short_tagline": "Own your AI.",
5+
"moat_sentence": "A cryptographic, offline-verifiable receipt for every call — the one thing no competitor ships.",
6+
"pillars": [
7+
{ "key": "own", "label": "Own it", "line": "Distill frontier-model quality into a small, private model you run on your own hardware." },
8+
{ "key": "prove", "label": "Prove it", "line": "A signed receipt for every call, verifiable by anyone with the public key — offline, no account." },
9+
{ "key": "govern", "label": "Govern it", "line": "SSO/SCIM/RBAC, spend caps, model entitlements, data residency, BYOC, and audit export." }
10+
],
11+
"ctas": {
12+
"primary": { "label": "Start free", "href": "/quickstart" },
13+
"secondary_proof": { "label": "Verify a receipt", "href": "/verify" },
14+
"enterprise": { "label": "Talk to sales", "href": "/enterprise" }
15+
},
16+
"contact_email": "dev@kolm.ai",
17+
"banned_terms": ["wrapper", "studio", "surface", "evidence-to-artifact", "powerful", "seamless", "unlock", "leverage", "robust", "world-class", "next-generation", "honest", "bill goes to zero"],
18+
"claims": {
19+
"soc2": "SOC 2 Type I available; Type II in progress",
20+
"trinity500": "On a 57-prompt held-out support-clarification set, a Qwen2.5-7B council-distilled from Claude 3.5 Sonnet + GPT-4o + DeepSeek-R1-32B asks the right clarifying question 96.5% of the time — equal to gpt-4o-mini, above claude-haiku-4-5 (64.9%) and base Qwen (84.2%) — INT4 NF4 on one RTX 5090. Scope: support-clarification only.",
21+
"int4_32b": "Run a 32B reasoning model in INT4 on one consumer GPU = inference (DeepSeek-R1-32B ~17.9 GB). Never training.",
22+
"cost": "No per-token bill after you compile — you run the model yourself on hardware you control."
23+
}
24+
}

0 commit comments

Comments
 (0)