The Edit Experience

Editing happens after Neuron. Once the compiler turns a live site into its own repo, that repo is meant to be totally self-contained — and editing is a separate, per-site thing on that repo, no factory in the loop. That constraint is the gift: it makes the loop fast, local, and magical — edit → the machine clears everything mechanical in milliseconds → the human faces one beautiful preview and one taste call → publish, reversibly. Your ruling is already half-written in the repo as a ratified line — "two doors, two speeds: provably-safe → auto; needs-judgment → preview → approve." The job isn't to invent it. It's to apply it consistently and build the two pieces never finished: the taste-moment and the provenance receipt.

The boundary. Self-containment is Neuron's deliverable, not an editing gap. The One Door's job isn't just "capture a faithful site" — it's "produce a fully self-contained, editable repo": the edit tooling, the guard, the CI, the reproject-and-publish lane vendored in and proven to run there. Fairfax missing its package.json and guard means Neuron's self-containment job isn't finished — an unfinished floor, not a broken product. So: Neuron guarantees a self-contained editable repo (part of finishing the door); the edit experience assumes that and lives entirely on top. It proceeds from the self-contained repo; it never builds the floor it stands on.

The spine below follows one edit — Alie changes the boarding blurb — end to end, on that self-contained repo.

  1. Intent

    Alie (or Fides, on the practice's behalf) says what should change, in plain language. The system resolves it to the exact claim(s) + anchors — the claim-decomposed genome makes every span addressable (fairfax already shatters into 6,353 claims, one per span) — and abstains if the scope isn't provable rather than guessing. Nothing edits what it can't locate.

  2. The edit becomes a superseding claim

    The change is a new asserted claim that supersedes the prior one for that span — attributed (to Alie, or to Fides), reversible, standing on a lineage back to genesis. Faithfulness for an edit is faithful to the provenance chain, not faithful-to-origin (that was capture). An edit can never masquerade as a certain observation — the no-false-green floor, applied to editing itself.

  3. The machine floor — [M], in the repo, in milliseconds

    The mechanical checks run right there in the self-contained repo, fast: byte-isolation (every other byte identical), structural well-formedness, injection-safety, completeness (all-or-none for a claim that recurs). Fail-closed, anti-vacuity. This clears everything mechanical off the human's plate — so what's left is only ever taste.

    ↘ go deeper — the [M] battery already exists
    render-validity R0–R5b: partition integrity, frozen-byte-identity, composition re-derived-from-claims (not trusting a stored hash), injection-safety, private/PII-leak — all byte-decidable, fail-closed, anti-vacuity. This is a durable floor Neuron vendors into the repo; it gets MORE valuable as models improve. Keep it.
  4. The router — the two doors

    Everything is a claim; the only question is whether it carries a taste dimension. No taste → auto. A phone number, an email, an hour — a claim that's simply right or wrong, no judgment in it. The machine proves it; it goes live. Taste-bearing → the judgment lane. A bio, a blurb, a color, a layout — the machine still clears the mechanical; the taste routes to a human. This one move resolves an inconsistency the project holds today: prose currently auto-ships on byte-safety alone (its taste-judge was silently dropped at session-120), while design still routes to a human. Under one rule — does this carry a taste dimension? — both behave the same, and prose gets its taste owner back.

  5. The taste-moment — the magical human surface (never built)

    The human faces exactly one beautiful thing. Not a diff dump, not a gate, not jargon: the before/after in place, the changed span highlighted, the machine's proof shown as an effortless green "safe" — you don't have to check the mechanics, we did — the provenance receipt, and one call: approve · tweak · reject. That's the whole ceremony. This is where "frictionless and magical" lives, and it's the piece sketched once in 2026-06 and never finished.

  6. Publish — reversible by construction

    On approve: reproject (produce.ts splices the new claim's bytes) → record the rollback anchor first → deploy → live, with the exact one-step undo in hand. Today this step is a deliberate no-op, gated on your ratification (a decision below).

The provenance receipt — the piece that makes it trustable. This is what turns "an agent edited your site" from scary into magical. For a Fides-authored edit, the receipt shows at a glance: what changed, what Fides considered (her reasoning — "kept your brand voice, matched the tone of your other service pages"), and the reversible lineage back to genesis. It's not a machine judging taste (that's undecidable — Rice's theorem); it's the evidence a human needs to render their own taste verdict fast. Without it, a human rubber-stamps blind or re-does the work. With it, they trust and move — and it scales, because the machine holds faithfulness at infinite scale, so the human's taste-surface stays tiny even across 490 sites.

The decisions I need from you

Tap the ✎ on any block, or just tell me in chat.

★ Q1 — Who owns taste?

DE staff today (Alie self-serving), the practice itself eventually, or both with different surfaces? This sets who the taste-moment is for — and today every approval path in the codebase is DE-internal; the practice is a beneficiary, never a participant.

★ Q2 — Ratify the self-contained publish lane?

Fides's approved → live is a deliberate no-op right now, explicitly gated on your ratification of the live-publish surface (a tracked WILLFUL-GAP with your name as the trigger). The edit experience isn't real until this opens. Ratify it — with the rollback-anchor-first / committed-tree-only safeguards — for the self-contained repo?

★ Q3 — Is the router rule right?

"Every claim; taste-bearing → human, right-or-wrong → auto." Is that the clean split, or do you want a finer line (e.g. some prose — a legal disclaimer, an hours change embedded in text — is really a settled claim and should auto)?

Q4 — Converge the five mechanisms to one.

Keep the spine (claim-genome, anchors, [M] gates, guard, produce.ts, the seam's trust primitives) — it's what Neuron vendors in. The five overlapping edit paths across four addressing schemes (10 of 13 il-edit ops stubbed) converge to the one clean loop, the rest retired. Confirm?

Q5 — Sequence.

Self-containment is part of finishing the door (Neuron's job) — already on the critical path now. Build the edit experience right after the door is proven-and-self-contained, or start prototyping the two never-built pieces (taste-moment + receipt) in parallel now — they're pure design and don't wait on the repo being wired?