# Sequence Worked Trace

One project, walked end to end through the full loop - plan, compile, generate, observe, deviate, reconcile, hit the chain cap, re-anchor, break for the night, and resume. The machine half of this trace already lives in `examples/`; this is the prose half that shows how the pieces are actually used. Field schemas without a trace get re-interpreted on every run; this file is the interpretation.

The project: **seq_airport_arrival** (`examples/sequence-airport-arrival/`) - a traveler exits the terminal, enters a waiting car, and the car departs. Deviation-handling details borrow the second fixture, **sequence-observed-deviation** (`examples/sequence-observed-deviation/`), which records a real unexpected-completed-beat event.

## 0 - Plan globally

The idea ("my character lands and is driven away") is bigger than one generation, so the Sequence Gate classifies it `sequence_project` and each performed clip receives a `narrative` Director's Read lane. Before Clip 01 exists, the plan fixes: the story promise and final outcome, beats grouped into **scenes**, and per clip a `narrative_job`, a `felt_intent`, the complete canonical internal `authoring_state`, and a completed endpoint. The airport fixture's creased paper airline tag is explicitly marked `authored_choice`, never misrepresented as visible in `@Image 1`. Its thumb-catch and fold replace the refused shoulder-drop/backward-glance defaults and survive even if a different agent writes Clip 02.

Here everything happens at one location in one time envelope, so the scene map is a single scene: `scene_01`, arc position `release`, canonical anchor `@Image 1` (the traveler's identity - note the internal space in the tag; it is preserved byte-for-byte forever), `max_chain_depth: 2`, and an audio plan of curb ambience and sync SFX per clip with score unified in post. Three clips are assigned: exit terminal and approach the car (Clip 01), enter and close the door (Clip 02), the car departs (Clip 03). See `project-state.json` - `scenes`, `beats`, `clips`.

Only Clip 01 is compiled. Clips 02-03 stay provisional intent cards, because their opening states do not exist yet.

## 1 - Compile Clip 01

The contract (`clip-01-contract.json`) carries the job ("Exit terminal and approach the open rear car door"), the felt intent ("the quiet relief of arrival"), the complete planned authoring record, the reserved beats (entering the car, departure), and the locks. The compiler (see [prompt-compiler](prompt-compiler.md)) emits one natural-language prompt (`clip-01-prompt.md`) - T2V, since nothing exists yet. The visible carriers ship; canonical labels and internal JSON do not.

## 2 - Observe, and accept a deviation

The take comes back. The user attaches the clip or its final frame (locally: `python scripts/extract_last_frame.py take.mp4`), and the **agent** fills the observation record from the pixels - the Observation Fast Path in [continuation-handoff](continuation-handoff.md) - asking only what a still cannot show. The review (`clip-01-take-review.json`) records the honest result: the traveler is **two steps short of the open door**, still walking, left-to-right. The plan said "reaches the door"; the pixels disagree.

Verdict: `accept_with_deviation`. Reconciliation (see [sequence-project-state](sequence-project-state.md)) updates project canon: the deviation is recorded, the observed end state and revised authored turn live in project state, and Clip 02's current contract opens **two steps out, mid-stride**. Clip 01's status becomes `accepted_with_deviation`, but its historical contract and prompt keep the planned endpoint; they are not rewritten to imitate observed truth. Rejected takes, by contrast, change nothing: rejected footage never becomes canon or a parent source.

## 3 - Continue from what actually happened

Clip 02 (`clip-02-continuation-contract.json`) is a `seamless_continuation` inside `scene_01`: parent `clip_01`, `extension_depth: 1`, sourced from the accepted take attached as `[Video 1]`. Its `value_before` exactly matches Clip 01's revised `value_after`; this is the dramatic handoff beside the physical one. The **source-carries-state rule** applies: the video reference carries the opening state, so the prompt text carries only the delta - finish the approach, enter, close the door, the exact filmable authoring carriers, the felt-intent carrier (the door thump cutting curb noise), and the reserved-beat exclusion (no departure yet). Prose that re-describes what `[Video 1]` already shows would be budget spent turning disagreement into drift.

## 4 - The other way plans break: a beat completes early

The second fixture shows the mirror-image deviation. A rooftop courier take planned only "unlock the case" - but the accepted footage shows the case unlocked **and opened** (`take-review.json`: "case opened one clip earlier than planned"). Reconciliation removes the now-completed `beat_open_case` from the future (compare `project-state-before.json` with `project-state-after.json`: three planned clips become two, and the next clip's job becomes "Light the already visible signal device"). The rule in both directions is the same: **accepted observed state overrides planned state** - whether the take fell short or ran ahead.

## 5 - The chain cap and the scene boundary

Clip 03 chains from Clip 02's accepted footage: `extension_depth: 2` - exactly at `scene_01`'s `max_chain_depth`. That is legal, and it is also the end of the chain: a third output-sourced generation would exceed the cap, and the validator (`scripts/project_state_check.py`) fails it with instructions to open from canonical references instead. This is scheduled re-anchoring: identity decays along output-sourced chains, so the reset is routine, planned at the scene boundary - an intentional cut, opened from `@Image 1`, depth back to 0 - not an emergency repair after drift is visible. If drift appears **before** the cap, that is an immediate `reanchor_after_drift` instead.

## 6 - Break for the night, resume tomorrow

Sessions end; canon must not. The Project State Capsule (template in [sequence-project-state](sequence-project-state.md)) carries the compact project in under ~50 lines: scene map (completed scenes compressed to one line each), current scene, current actual state, open motion, completed beats, next clip job and intent, the expanded internal authoring state, its visible carriers, locks, reserved beats, extension depth, and uncertainties. A new session pastes the capsule and continues from the recorded physical and dramatic state - no hidden memory is assumed, and the machine truth (`project-state.json`, per the State Lifecycle) regenerates the capsule whenever `state_revision` bumps.

## 7 - Close out

Clip 03 ends the scene and the story: the car departs, the final outcome is met, the scene closes (its line in the capsule compresses to its outcome), and the take log archives. Total user effort per cycle after the plan: attach the take, confirm the agent's observation record, approve the intent echo, generate. The loop's guarantees did the rest: nothing replayed, nothing leaked early, nothing invented, and the story the user accepted is the story the state remembers.