atypica/creative-reasoning-trajectories

Creative Reasoning Trajectories (Preview)

Step-level trajectories of professional creative work in which the supervised target is the decision, not the keystroke. 76 steps across 4 trajectories, 45% of them judgment steps.

A preview release from a larger body of work. Read Status and honesty before using the files — this release previews the schema and the labelling target, and is explicit about what has and has not been recorded yet.

Why this format

Open computer-use corpora label what the tool did: click, write, photoshop_layer. That is the right target if you are training an agent to operate software. It is the wrong target if the failure you care about is an agent that operates software fluently and still produces work no client would accept.

In professional creative work the expensive steps are the ones with no interface event at all: deciding that a client's "时髦" means restraint rather than avant-garde, recognising that a search returned the wrong kind of result, knowing that a brand book is not on the open internet so further searching is waste. We measured a public design-trajectory corpus of 294 steps and found 4 steps (1.4%) carrying any signal of a rejected alternative or an uncertain judgment — and in those, the rejected artifact itself was not retained. This format is built to put that class of step in the centre.

Every step therefore carries two action labels:

• action_type — what the tool did (20 distinct values here)
• creative_function — what the creative process did (all 16 values exercised)

Plus step_criticality, so training can weight judgment over mechanics instead of sampling uniformly. 34 of 76 steps (45%) are decision.

What's in it

Two levels. The generic workflows are the transferable part; the project instance shows them under real commercial constraint.

File	Steps	What it captures
raw/wf_a_input_fallback_cascade.jsonl	14	Required context is missing. Working an ordered acquisition cascade, then resolving a failed compliance check.
raw/wf_b_reference_sweep.jsonl	16	Establishing a visual direction from nothing, via multi-source reference sweep and convergence.
raw/wf_c_replicate_then_vary.jsonl	16	Replicating a reference in full as a learning artifact, then substituting its variables one at a time.
raw/proj_summer_outfit_kv.jsonl	30	One concrete deliverable — a vertical key visual whose first round was rejected for misreading the brief.

Field definitions are in SCHEMA.md.

Files and formats

data/train.parquet is the canonical artifact — it is what the viewer and load_dataset read. The four raw/*.jsonl files are the human-readable source, one file per trajectory, and build_parquet.py regenerates the Parquet from them under an explicit Arrow schema.

Parquet is shipped deliberately. Many fields here are conditional and null for whole trajectories at a time, so type inference over concatenated JSONL depends on which file is read first: a column that is all-null in the first file is inferred as null type, and the first string in a later file fails to cast. An explicit schema removes the inference step. If you extend the JSONL, rerun the build script rather than pointing a loader at raw/ directly.

Three things this format records that we could not find elsewhere

Negative knowledge — paths proven not to exist. dead_end marks a route as unavailable and says why. The load-bearing example: a real brand book is essentially never on the open internet, so failing at the official-website rung means public channels are exhausted and the correct next action is to ask a human. An agent without this will search indefinitely. Existing corpora record successful paths only.

Hard versus soft constraints. constraint_hardness separates constraints that are legally or technically non-negotiable from those that are brand preference. The dataset includes a violation of Chinese advertising law's ban on absolute superlatives (绝对化用语) — a hard constraint no model can infer from pixels — alongside a soft palette disagreement that is resolved by annotating the dispute rather than complying.

Variable coupling. Substituting one design variable breaks something that depended on it. coupling_effect records four such couplings, including: replacing a Latin typeface with a Chinese one breaks letter-spacing, optical weight and line-height because Chinese glyphs are full-width and denser, forcing type scale and tracking to be reworked. This is muscle memory for a practitioner and invisible in any finished artifact.

Rejected work is retained rather than mentioned in passing: killed directions keep judgment, judgment_rationale, and the alternatives_considered that were on the table at that moment.

Reasoning provenance

Stated explicitly because it determines what the data is worth:

• thought — first-person reasoning, intended to be transcribed from the practitioner's narration while working, not reconstructed afterwards
• observation, reflection — model-generated descriptions of screen state before and after the action

Reasoning written after the fact reads plausibly and teaches the wrong thing. Any production release of this format must preserve the narration-grounded distinction above.

Tool independence

source is a generic role (asset_library, design_canvas, raster_editor, messaging, browser, email); source_app names a concrete product as metadata only. A lab substituting its own stack reads source and ignores source_app. The trajectories are deliberately not tutorials for any specific product.

Status and honesty

Read this before using the files.

• Screenshots are not captured. image holds placeholder paths. Nothing in this release contains a real screenshot.
• Step content is illustrative. It is modelled on the structure of real Chinese creative-industry practice, including one internal project retrospective, but the trajectories were authored rather than recorded, and no row corresponds to a real client engagement. Brands are fictional.
• thought is therefore not yet narration-grounded. It is written to the target style. The provenance rule above describes what a recorded release must do, not what this preview did.
• Not training-scale. Preview scale, one vertical, no held-out split. Use it to evaluate the schema and the labelling target.
• session_uuid is intentionally absent: at this scale each file is a single trajectory, so there is no session grouping to represent.

License

CC BY 4.0.