docs: a rough roadmap to 1.0, and the road so far

[firejune]

A rough ROADMAP.md: goals only, no dates, no feature list. Linked from the README's documentation table.

The framing, which is the part worth arguing about

⛔ The standing rule here is that direction comes from what users hit, not from what this repository predicts — and writing a roadmap does not suspend it. So the file says that outright in its second paragraph and lists conditions that have to become true, not work booked in advance.

What 1.0 means is one thing: the input spec stops moving. Everything else about a version number is bookkeeping; that one is a promise to whoever writes a rig spec. Pre-1.0 a spec change is a minor bump and nobody pays. After 1.0 the author pays.

The measurement behind that, rather than a feeling about it: src/rig.ts and src/types.ts changed in four of the last nine releases, including additions of 185 and 175 lines. RELEASING.md already said "1.0.0 is a deliberate act, not the side effect of one commit"; this gives that sentence something to be deliberate about.

The four goals

Each has an exit condition and no date:

goal	done when
Real art has used the new surface	each recent construct has at least one cut of real art compiling through it
The spec has stopped growing on its own initiative	spec changes come from a need somebody reported, over enough releases to be a pattern rather than a lull
Every claim the repository makes about itself is derived	nothing load-bearing is hand-maintained, and the derivations are themselves asserted
The published package works on a machine that has never seen this repo	installing from the registry and running a build is checked mechanically, every release

⚠️ The first gates the second, which is why the order is what it is. Depth input, the turn ceiling, sliders and parameter-driven faces are measured only against generated fixtures — a registered cuts run comes back green and touches none of them, because the cuts predate them. A spec frozen at 1.0 that real art has never exercised is a prediction.

What is explicitly not in 1.0

Written down, because leaving it out is how a roadmap grows by inference:

• The format-agnostic core — ruled to come after the feature work, and nothing here changes that.
• Full coverage of the Spine 4.3 surface — ⭐ 1.0 is not everything works, it is the boundary is honest and stable: what rigc emits it emits correctly, and what it does not it refuses by name. Coverage is a separate axis and may keep moving after 1.0.
• Rewriting the dated research notes — a note stamped with the day it was measured is doing its job. Making it current would make it a live claim, and live claims have to be derived.

The road so far

Same register, in order rather than by date: the split out of a game project's sandbox and the pass it took to finish sanitising the code after the files had moved; the ladder, whose cleared rungs stay in place as regression gates; the gate being put under the gate — a rung's pass withdrawn on a tolerance change and restored after the gate answered what that exposed; publication with an allowlist and the guide shipped inside the package; the gallery and its films, which have been the best defect-finder the project has; the face and Spine's own slider; and currency.

🔸 It carries no counts — not tags, not controls, not days. Counts drift the moment they are written, which is the third goal's whole subject, and LADDER.md and the CHANGELOG hold the real record. Its last phase is that third goal, still open, which is the seam between the two halves of the file.

Verification: bun run selftest 551 PASS / 0 FAIL / 0 SKIP / 0 HOLE.

[firejune]

Reframed on the owner's correction

"The goal of this project is AI doing rigging and animation with no human hands. Spine is one of the tools we support, not a dependency."

The first draft said what 1.0 means procedurally and never said what the work is for, which let Spine read as the centre of it. Fixed at the top of the file, and it changes two other passages.

The distinction the file now draws

Spine is a backend. spine-core is linked for a narrower reason than allegiance to a format: everything written to disk has to have been read back by a parser rigc did not write. That is an oracle, not a dependency.

⚠️ Worth being precise, because "not a dependency" could be misread as licence to drop the link. It is not — the doctrine's structural invariant stands untouched. What the correction settles is the reason the link exists: verification, not format allegiance. Another backend does not need permission; it needs an oracle of its own, and building that is the larger half of the job. Not in 1.0's first bullet is now "a second backend", explicitly not because Spine is the point.

The 1.0 promise is unchanged, and its reason got better

"The input spec stops moving" was a packaging promise when Spine was implicitly the centre. Under a multi-backend goal it is the load-bearing one, because the rig and motion specs are the part that belongs to no backend:

A format can be swapped under a spec that holds still. It cannot be swapped under one that does not.

One thing the correction retires

An earlier framing — #285's, now closed — held that the value was an AI draft that a person then finishes by hand in the editor. That is the opposite of the stated goal, so the file says so: an editor round trip is how a result gets checked, not where it gets finished. The face angle that became a value rather than a moment on a timeline is the direction; nobody keys it.

---

⚠️ Open question the correction exposes, not answered here

All four goals are about whether the tool can be trusted — real art, spec stability, derived claims, a working install. None is about whether the mission has been demonstrated. If the end state is hands-free, the most direct 1.0 evidence would be a hands-free run on art nobody has seen, and the ladder and PILOT.md are that shape but sit in the record rather than in the conditions.

Deliberately not added, because inventing a goal is how a prediction-driven roadmap starts. Two ways to settle it:

• Leave it. Trustworthiness is what 1.0 certifies; the hands-free proof keeps running as the ladder and does not gate a version.
• Add a fifth goal. ⚠️ Then the exit condition has to be drawn first, and it is genuinely hard: where is the line of "a hand"? Writing the brief is a hand. Choosing the art is a hand. Without that line the goal cannot be measured, and an unmeasurable goal on this list would be the only one.

Owner's call.

Verification: bun run selftest 551 PASS / 0 FAIL / 0 SKIP / 0 HOLE.