One practice site from capture to a teammate editing it — with every line marked by how it was checked, because four things I told you today turned out not to be true.
Your words, today. Everything else on this page is measured against these four.
Parked by you today: certificates, the provenance model, the pull-request shape, and fidelity proof. All real. None of them this week.
The compiler takes a practice's real website and reproduces it. The replica renders complete — navigation, services, booking, hours, footer.
RAN — I fetched the real site and the replica and compared them. Title, headline, phone, address, hours and all six navigation items are identical.
Your read, which the whole day supports: the compiler is proven; the machinery around it is not compiler-grade. Every failure below is machinery, not compiler.
A full certification run takes about four hours. Work exists to cut it to roughly forty minutes by running the expensive checks once per compiler version instead of once per site.
RAN — the roster guard refuses a thinned roster and names the missing check; passes clean on the real roster; and the escape hatch now genuinely blocks the certificate (zero mint attempts under the escape versus one on the honest path, with the recorder proven live on both arms).
REPORTED — and it breaks the speed win. The escape only guards the battery script. The door mints its certificate through a different path that never consults it. I have not re-run that finding.
Nine checks in that run are under investigation now. They were demoted so their findings no longer block anything, and several look like survivors of an architecture this compiler abandoned. Eleven one-line rulings are coming to you.
RAN — per-edit previews work and render the full site. I opened one.
RAN — the main address returns nothing. Nothing has ever deployed the site itself, only per-edit previews. A fleet is fixing that now.
RAN — your Cloudflare credentials work. All probes green, including the two that were missing this morning: media upload, and the permission that killed a run at minute sixty-four.
Work exists that makes every future site carry the edit engine and its workflow from birth, instead of being wired up one at a time. It is complete and pushed.
RAN — the branch is on the remote, 828 lines including its own proof and guard tests.
UNKNOWN — no adversary has attacked it.
This is the shortest path to objective 4, and it is sitting still.
A real teammate tried this morning. RAN — I pulled the logs. Fourteen attempts in thirty-six minutes. Eleven failed. The three that succeeded opened a pull request saying "No preview deployed: Cloudflare credentials are not configured." So even the successes gave them nothing to look at.
Four distinct walls, verbatim from the logs:
The diagnosis: the engine was built for a conversation and deployed as a form. Three of the four failures are the engine trying to talk. A submit-and-wait form cannot answer, so every question became a dead end.
In Cowork there is someone on the other end. That is not a path to build — Claude running the commands and asking when it needs to is the conversation.
RAN — a fact change appears in the rendered preview, attributed and versioned. Fourteen kinds of edit exist: text, phone, email, headline, images, colours and more.
REPORTED — undo is one command, and a round-trip check asserts it byte-for-byte. Not re-run by me.
A full site render measures 8.7 seconds. Rendering is not the bottleneck. Everything slow is the machinery around it: checking out a repo, installing dependencies, the round-trip to a build service, the deploy.
A fleet is timing every available path with a stopwatch and walking it three times as a non-engineer teammate. It reports the measured number even if it misses the minute — a verified ninety seconds with a named blocker is worth more to you than a promise of forty.
RAN — an edit rewrites the fact in place. The record still says that wording was observed on the practice's own website at capture time, with the capture timestamp and the confidence that follows from having observed it. A person typed it that morning.
RAN — nothing anywhere checks this. The validator that would runs against a kind of file the live book does not contain.
On a test book this is acceptable. It is the reason not to scale usage before it is fixed.
RAN — the search-tool one I proved directly. The shell's search command silently returns "no match" on the compiler's largest file while the string sits four thousand bytes in. Two confident "nothing calls this" verdicts today rested on it. Both were wrong. That is why this page separates what I ran from what I was told.
Credentials are clear, so it can run. The fast path is refuted; the slow path is about four hours and may hit a check that once hung for over two. But you have said the team needs editing, not a certificate — so my read is this simply drops down the list. Tell me if that is wrong.
Eleven one-line rulings are coming with what each measures in plain terms, what it costs, and what it uniquely catches. One of them looks like a check that should be blocking and quietly stopped. My expectation is that most come back with no recorded reason to ever revisit them — which is its own finding.
Each needs an adversary before it lands. That queue is the only thing between today's work and a moved trunk. I can run them in sequence, or land the one that is genuinely finished and unattacked first.
If the verification labels are noise, say so and I will drop them. If they are the most useful thing here, I will hold every future report to the same rule.
| Truth | Today | The gap — and what closes it |
|---|---|---|
| 1 · The compiler reproduces any practice's site faithfully | PROVEN | None. Guarded going forward by born-editable (queue #5). |
| 2 · The proof is real — checks fire, can go red, can't be thinned | Machinery exists; honesty layer broken in places | Six real checks switched back on (lane 2) · the door's mint path made tier-aware (queue #4's named blocker) · then the wiring census puts a NUMBER on the unknown-unknowns. |
| 3 · Failure tells the whole truth at once | Battery: yes (Robert built it). Door: stops at first failure | The keep-going door (lane 1) — every stage ends in a named state; two injected failures must both appear in the surface you read. |
| 4 · What comes out is alive — every visible thing a grabbable handle, honestly recorded | Engine proven (13s edits, byte-honest round-trip). Binding broken; provenance dishonest | The one-line binding fix if confirmed (lane 3) → re-mine every site · claims-not-overwrites (queue #9 → #10). |
| 5 · The experience never lies | Previews real; false-success states exist (tab-title edit, self-reversing undo, over-reporting coverage) | Kill each named false-success state · gate before publish (queue #8) · natural phrasing + multi-edit ride the claims model. |
| 6 · Independent of Robert's head | Register live; reports carry RAN/REPORTED labels | The wiring census (queued next: 768 orphans / 82 uncertifiable gates → a checked list) · wiring-is-part-of-done now binding on all new work. |
| # | Commitment | State |
|---|---|---|
| 1 | Complete failure list at the door (kills the "20 more runs" cycle) | Built + pushed. A failed run now lists all 25 stages with named states — the proof run showed 2 real failures + 21 honestly-skipped where before you'd have seen one. Fixed a live bug on the way (a stage bypassing its own dependency check). Under adversary; named risk: a wrong dependency edge could hide a real defect as "skipped." |
| 2 | Probe rulings: six real checks ON, four out, one deleted — wall-clock + honesty in one change | Lane running; adversary queued |
| 3 | Why visible text isn't editable | Diagnosis landed: binding, one line in the prose miner — independent check next |
| # | Commitment | State |
|---|---|---|
| 4 | Battery fires only when it should (~40-min site tier) | Built + repaired; one named blocker (the door's own mint path) before merge |
| 5 | Every new site born editable | Built, pushed, never attacked |
| 6 | Real parallelism (the sub-20 path) | Built, validated once, unmerged |
| 7 | Southtown identity fix | Refuted — rebuild queued; likely same root as #3 |
| 8 | Edit-lane gate (no PR without a passed check) | Built; ten bypasses found; fix list named |
| 9 | Editing core item 0 (the organs compose) | Graded 8 and 6, accept-with-fixes; three small fixes named |
| 10 | Honest provenance on edits — claims, not overwrites | Ruled YES by you; design done; queued behind #9 |
Show-your-evidence is law · every new gate proves it can fail · org transfer staged behind current merges · the probe rulings above.
Update your de-hub plugin 1.15.7 → 1.17.0 (unlocks the no-GitHub Cowork edit path) · decide runner hosting (edit requests queue silently when your Mac sleeps).
The wiring census — N things claim to gate, M actually fire — the number that ends "I don't know how deep" (ceiling already known: 768 orphans, 82 uncertifiable gates) · worker-endpoint auth before anything customer-facing · the stale-branch triage.
Journey:
Compile ·
Certify ·
Publish ·
Born editable ·
They ask ·
Applied ·
They see it ·
Recorded
Themes: Objectives · The pattern · The gap · Ruled today
Your calls: Run 10 · Nine checks · Merge queue · This map