Roadmap

The canonical roadmap is docs/ROADMAP.md in the repo — this page summarizes it. Where they differ, the repo is right.

It was authored on 2026-08-10, before any code. Phases A–C are firm; everything after is provisional by construction — each phase is shaped by what the phase before it measures. A provisional phase may be re-cut. It may not be skipped silently, and a phase that gets re-cut records why.

Definition of done

Two bars, in order. The second is the real one:

1. built — the tool ships: repo public, quality gates pass, published, CI green.
2. filled — the Director has accepted a shot armature produced and used it in a real project. (Corrected 2026-08-11: this line used to say “a real game project.” The game is one consumer — a cutscene, a film shot, or any footage counts the same.)

A test count is not the second bar. A demo reel is not the second bar. Until a shot is accepted and used in a real project, armature is built, and the roadmap says so out loud. The studio has a name for the failure this prevents: tools that get built and never filled.

As of 2026-08-13, armature is built by the first bar’s own terms — public, hard gates green (shipcheck audit exit 0), v0.1.0 released. filled remains open, and this page keeps saying so out loud.

The arc as run — recorded 2026-08-11, after the audit

The plan below was authored on day one and the first arc re-cut it silently, which the repo-wide audit names as a first-class failure: E03–E06 as run bear little resemblance to E03–E06 as planned, and identity — the planned experiment closest to the product — never ran while two unplanned proxies did. What actually ran: exporter (E01) · first contact (E02) · authored motion (E03) · the between-generation noise floor (E04) · reference-onto-schematic (E06) · one withdrawal un-run (E05). The measurements are real and kept; the framing failure is the audit’s subject.

What governs now: every credit-spending spec carries a Trajectory row — what the spend advances toward the full GLB→video scope — with the same force as its licence rows. The current line: the first performer built fresh through facet’s route (F01/E33) → the named-bone rig (E07) → the first authored performance of a real character (E08).

The arc as it stands — recorded 2026-08-13

Twelve experiments closed (E05 withdrawn un-run), and v0.1.0 — the first marked state of the record — released 2026-08-13. The phase letters below map loosely onto what actually ran: foundation, the exporter and first contact (A–C) held; control (D) became the control arc and the driven route, proven at shot level and parked for AI-animation buildout; identity (E) is measured at the Director’s eye three ways — driven (E08), unanchored (E11 wave 1), and through a held world (E12) — with reference-view mechanics now E13’s probe on a hosted identity-lock tier: dispatched, halted at zero spend on two structural premise failures, repaired by a support arc that rebuilt the reference kit with true alpha, and re-armed, all on 2026-08-13. Continuity (F) waits on the narration shelf. The posture since 2026-08-12 is a learning monorepo: experiments prove paths, no route is canon by momentum, and three routes now stand where the plan imagined one.

Later the same day: E13 ran its four generations — the repo’s first partner-credit spend, inside its pre-stated bracket — and closed with the identity verdict at the Director’s eye: the character survives the hosted tier, and the reference’s own ground steers the model-decided world. The composed route graduated from probe to measured route. E14, the LoRA scene-lever bake-off, closed behind it the same date: both T2V-trained style LoRAs bound on the camera tier’s derivative weights at zero partner credits, the character held on the grading LoRA and failed on the photo-real pair, and the free route’s lever is priced — winner named, its served-file tier question and credit obligation recorded beside it.

The arc

Phase	What it is
A — Foundation	Repo footing, the discipline, this roadmap, the research grounding, the first licence map, the exporter’s spec. No code beyond scaffold.
B — The exporter	Given a mesh and a shot spec, render per-frame control channels from a staged scene in headless Blender, plus a manifest that makes the run reproducible. Everything downstream consumes its output. Zero generation, zero credits — no credit is spent until the exporter’s output is trusted.
C — First contact	The hinge. Measure the provider’s noise floor first, then run one control probe: one character, one short shot. The gate is a control | output | reference | provenance sheet built before any metric is quoted.
D — Control	Which control signal, at what strength. Arms vary one thing each.
E — Identity	How many reference views, arranged how. A per-character adapter runs only if the reference stack leaves a gap worth the training cost.
F — Continuity	Whether the same character survives across separate generations, and where a single shot’s coherence ends.
G — Ship	The tool surface, then the quality gates, then the full treatment, then a real shot in a real project.

Phase C is where the thesis lives or dies. If a CG-rendered control sequence does not hold a character in a real video model, the honest outcome is to say so and stop — and that is written down now, in advance, so that it cannot later be renegotiated. A negative result there is a full success and saves the rest of the arc.

Two structural notes carried from the spec:

• Phase D must observe both failure directions: too weak, where structure and identity drift; and too strong, where motion goes stiff and the render’s own artifacts print through. A strength curve with only one end observed is half a result.
• Phase F’s artifact is a cut, judged as a cut — not as a set of clips that each look fine.

Drift tripwires

Named in advance so they can be caught by name. Any session may call one; the Director rules.

1. Building the tool before proving the thesis. Product polish before the hinge rules is drift.
2. Chasing the newest model. A new model is an arm in a future experiment, never a reason to restart. The arc outlives any model version.
3. A metric quietly becoming the judge. The first time a number is cited to accept an artifact the Director has not seen, this one has fired.
4. Scope leaking toward an editor. Assembly is not the job.
5. “Just for a test” non-commercial models. The licence gate has no test exemption.
6. Unbounded credits. A submission without a stated ceiling is a defect, not a shortcut.
7. Feature creep in the exporter because a channel is easy to add. It ships what the experiments need, not what Blender can do.
8. Session sprawl — a session that keeps going past its deliverable. Sessions end at the deliverable and its gates.

What is deliberately still open

These are architecture decisions that evidence should settle rather than taste:

• Which control channels the exporter emits first, and their exact format conventions.
• Which generation route the first probe uses — open weights on cloud GPU, or a partner API — gated by the licence map rather than by convenience.
• Whether the shot or the cut is the tool’s output unit.
• Whether reference views or a trained adapter is the identity mechanism.

Nothing above is a prediction. Where the roadmap states a number or a mechanism later, it carries the measurement that produced it, or it is marked provisional.

Where publishing sits

Publishing reached its first milestone on 2026-08-13: v0.1.0, a GitHub release marking a state of the record — twelve experiments, three routes, the laws, with translations landed before the tag because a tag is immutable. Nothing ships to a package registry — the reserved names remain unused, and a version here marks the record, not an installable artifact. (This section read “nothing has been published from this repo yet” until the release; the day-zero half expired and the registry half stands.) The details, including a load-bearing constraint about which workflow filename the release path authenticates, are in docs/publishing.md.