01The Idea
Most "AI site builders" hallucinate a finished site — they emit plausible pages that were never checked against a plan. SiteForge refuses that. It treats a website as the realization of a contract: a plan so precise that every requirement in it is a check the build must pass. Nothing ships that the contract cannot verify.
Contract-first
A signed plan compiles into machine-verifiable acceptance checks. The builder's only job is to satisfy them — no drift, no improvisation.
Clean-rebuild & self-healing
Structure is regenerated to match the contract rather than patched in place, so the site cannot silently diverge from its plan.
Fail-closed stages
Each stage is human-signed and consumes the previous stage's signed output. A missing signature stops the pipeline — it never guesses forward.
02Contract-First — the plan is the spec
The centre of SiteForge is a compiled build contract: a keyword-to-structure mapping where every entry is simultaneously an instruction and an acceptance test. It is a bet portfolio made concrete — each mapped page is a wager on a search opportunity — but the structure that realizes it is deterministic and checkable.
Each arrow is a signature. Upstream unknowns are surfaced and kept as unknowns rather than papered over — the contract is only as strong as the evidence signed into it, and the system is explicit about that.
03The 8 Stages
SiteForge runs a greenfield site from positioning to a filled, audited site. Every stage is human-signed and fail-closed; the badge on the right shows who owns the decision.
Discovery & positioning
Problem-first intake: business model, market, seed keywords, competitors, and constraints. Discovery-first — surface and keep the unknowns instead of jumping to a solution. Output: a signed brief.
Brand / entity foundation
Resolve the brand and author entities, the proof of expertise, and the profile plan. Every fact is supplied by the human and never inferred — this is a decision-and-dossier gate, not a content step.
Keyword & SERP evidence
Real keyword expansion (volume, difficulty, intent), the top-results shape, competitors, and cluster candidates. The build contract is fail-closed without this signed evidence corpus.
Compile the build contract
Turn the signed brief + evidence into a machine-verifiable keyword-to-structure mapping. Red-teamed and validated; every requirement becomes an acceptance check. This is the spec the build consumes.
Approve the theme
Generate a real block theme plus one rendered preview per page type, bound by hash to the exact contract. The human approves the look and signs. This is not a screenshot mock-up — it is a real, verifiable theme package.
Build the structure
Realize the signed mapping into a verified staging site — complete, themed, structurally correct, and with no prose yet. Deterministic and self-healing. Green here means it conforms to the plan.
Fill real content
Write the real article prose per page — grounded in the content brief, the author's voice, and the research evidence, with no fabricated numbers. Content only; it never touches structure, theme, or strategy.
Conformance & health audit
Verify the built site still conforms to its contract and apply deterministic repairs — drift, broken links, wrong index directives, missing schema, redirect chains. Structural and technical only; not a ranking diagnosis.
04What "Green" Means — and what it doesn't
Green means
- The site conforms to the signed contract
- Every mapped page exists with the required structure, schema, and internal links
- No drift, broken links, redirect chains, or wrong index directives
- The theme matches the approved, hash-bound design
- The rendered pages look correct, not just pass text checks
Green does not mean
- The site will rank — that is downstream and probabilistic
- The content strategy was correct — the contract is a bet, honestly labelled
- Off-page authority exists — that is a separate concern entirely
- The work is "finished" as a product — structure conformance is not the same as a living site
This distinction is the whole ethic of the system: prove what can be proven, and never dress up a structural pass as a ranking guarantee.
05sf:rival — the autonomous parity loop
Where sf builds greenfield, sf:rival sharpens an existing live site. You hand it your own site, the credentials to edit it, and a set of competitor URLs. It then runs autonomously — bench, discover, fix, verify, re-bench — until the site is internally superior to the whole field on every measured and judged dimension, and reports back. There is no per-edit confirmation.
The loop closes on a real target — internal superiority across the field — not on a fixed number of iterations. Its safety does not come from asking the human at each step; it comes from a mechanical guard on every edit.
06Autonomy & Safety — how it runs unattended
Structure-preserving verifier
Every edit passes through a mechanical verifier that checks the page's structure is intact. PASS keeps the change; FAIL reverts it and flags it. This — not a human confirmation prompt — is what makes unattended editing safe.
Snapshot → verify → restore → purge
Before any live edit the current state is snapshotted; after it, the rendered result is verified and the cache purged. A bad edit is rolled back to the snapshot automatically.
Mandatory open-ended discovery
A fixed checklist is a rut. Each round runs an open-ended pass that hunts for issues a checklist can't name — stale content, canonical mistakes, follow-link leaks, weak clustering — so the loop keeps finding real gains.
Never fabricate to win parity
The loop will not invent identity facts, authors, or trust signals just to match a rival. If closing a gap needs a real external asset, it stops and hands that decision to the human.
07Scope & the Three Human Gates
sf:rival is deliberately narrow. It competes on internal factors — the things you fully control on your own pages — and explicitly leaves the rest alone.
In scope — internal, site-wide
Schema and entity graph, navigation and footer, internal linking and clusters, canonical correctness, content freshness, link hygiene, de-cannibalization, on-page structure. The full site, not just one page.
Out of scope — always
Off-page authority: backlinks, domain rating, referring domains. Also out: "why isn't it ranking" diagnosis. sf:rival makes a site internally superior; it does not chase authority or explain the SERP.
It stops autonomously for only three classes of decision
Real external assets
Anything that requires creating a genuine off-site asset or identity fact — never fabricated to close a gap.Home title / H1 on a ranking page
The most sensitive on-page levers on a page that already ranks are proposed for a human decision, not silently changed.Genuine layout restructure
A real structural redesign (as opposed to a structure-preserving edit) is a human call, not an autonomous one.Everything else inside the internal-superiority envelope runs to done without interrupting the human.
08sf vs sf:rival — same ethic, opposite starting point
| sf (SiteForge build) | sf:rival | |
|---|---|---|
| Starting point | Greenfield — no site yet | An existing live site |
| Goal | A structurally correct site that conforms to a signed contract | A live site made internally superior to a field of rivals |
| Control model | Human-signed at every stage; fail-closed | Autonomous loop; no per-edit confirmation |
| Safety comes from | Signatures + preflight gates + hash-binding | A mechanical structure-preserving verifier + snapshots |
| Human gates | One signature per stage | Only three decision classes (assets · home title/H1 · real restructure) |
| "Green" means | Conforms to the plan — never a ranking claim | Internally superior on measured + judged dimensions |
| Off-page | A separate, later concern | Explicitly out of scope |
| Shared ethic | Prove what can be proven · never fabricate · never overclaim the outcome · verify the render, not the promise | |