research artifact phil-only software ILCA · 2026-04-22 Not distributed. Not a product. Notes from one day at the workbench.

eschaton corpse

a two-mind blind drawing experiment · v00000-003

Two large language models take turns drawing one continuous line. They never see each other’s marks. Only the subject, an assigned part, and the coordinate where the pen was left. The drawing is the residue of what they cannot share.

A continuous line in ink and terracotta, concentric rings expanding outward from a teacup to become a ship caught in a storm

Two models from different labs, asked the same question, converge. They hedge. They agree. Their signatures sand each other smooth. This instrument runs the opposite experiment. Each model makes one segment of a single continuous line, blind to every other segment. Each model gets a subject, an assigned feature, and the coordinate where the pen was left — nothing else. The drawing is what survives that not-knowing.

Over one day of iteration on April 22, 2026, the instrument moved through three shapes. A strict horizontal-band version that eliminated interference and also eliminated the collaboration. A band-free version that restored the collaboration and also sent the line wandering. A parts-planned version that kept the wandering but anchored the figure. Each shape produced visibly different drawings. This page is a field report on what the three shapes teach, with the final shape as the primary result.

The instrument is a browser. Keys to the two APIs live in a local config file. Two models alternate: Claude Opus 4.7 (Anthropic) draws in ink; Gemini 2.5 Pro (Google) draws in terracotta. Color is locked by model identity so the handoffs are visible. The drawing is a single SVG path. No pen lifts.

One model, chosen as first mover, writes the subject and splits it into N parts — one phrase per segment, ordered along the chosen direction of travel. The models then alternate, drawing one part each, each blind to the others. Each segment sees the overall subject, its one assigned part, and the entry coordinate where the previous pen left off. Nothing else crosses between them.

A segment may go anywhere on the canvas. The nominal band for its assigned part is a hint, not a wall. Where lines collide, they collide blind. The model after cannot see what the model before painted across its territory. The handoff coordinate is the only point in the system where the two “minds” touch.

The instrument changed three times.

Each change was a response to something wrong with the previous shape. The drawings below are the evidence at each step — not highlights, not cherry-picks, just the representative output at each stage.

A jellyfish drawing in seven horizontal bands, each model's work isolated in its own strip
Shape 1 · strict bands
a jellyfish with long tentacles descending through deep water
7 segments, horizontal bands, clamped strictly. The models stayed in their lanes. The drawing reads as seven stacked ribbons. Two models worked on the same feature (the bell) independently because nothing told them who owned what. Bands eliminated overdraw and with it the collaboration itself.
A diver drawing, freer composition, but with two diagonal ruler lines where one model silently failed
Shape 2 · band freedom
a deep-sea diver sinking past a shipwreck toward the ocean floor
5 segments, bands as semantic hint only, paths free to roam. The drawing regained continuity and wander. Two straight diagonals across the figure are evidence of a silent parser failure: Gemini returned no usable path twice; the fallback drew a bezier instead. Revealed a reliability problem, not an aesthetic one.
A marionette drawing, clean five-part figure with clear division of labor between ink and terracotta
Shape 3 · parts plan, freedom, locked stroke
a marionette of a sad clown slumping down from its control bar
5 segments, parts assigned by the prompt author, stroke locked by model identity, Gemini forced into strict JSON output. Control bar, head and collar, torso with buttons, limp legs, pointed shoes. Handoff drift 0.017 — the line wandered heavily inside the figure and still landed near its starting x.

A story whispered into a teacup ripples outward to become a storm at sea.

With the three shapes working, the last lever to test was orientation. The first three drawings all flowed top to bottom. Adding vertical, angled, and free modes was the question: does the handoff geometry itself shape what gets drawn, or is it merely scaffolding?

In free orientation there is no grid. Each model picks where the next one enters. Narrative replaces spatial structure. The prompt author is asked for an associative subject — something that can transform across stages rather than stack across them. Gemini proposed the teacup that becomes a storm. The models drew the following.

The eschaton corpse app in its complete state, showing the finished drawing and meta sidebar with handoff drift
Workbench · run complete · handoff drift 0.703
The transcript panel below the drawing, showing the subject, parts plan, and per-segment reasoning
Workbench · transcript view · parts + reasonings

The drift number is the honest measurement of the free mode. With no grid anchoring the handoffs, the line walked 0.703 units across the canvas — enough to change where the last ship sits relative to the first teacup rim. The drawing is not a descent; it is a circuit. Two minds built a vortex neither could see.

Four handoff geometries.

Each mode changes where the line crosses between minds and what kinds of subjects the prompt author tends to write. The instrument invites a subject appropriate to the geometry; the geometry shapes what can be drawn without strain.

Down

Horizontal bands, top to bottom. Gravity. Descending figures, slumping bodies, sinking divers.

Across

Vertical bands, left to right. Procession. Trains, rivers, phrases unfurling, skylines.

Angled

A random tilt between ±25° chosen per run. Cascade. Figures slipping, kite strings, falling diagonals.

Free

No grid. The drawer picks where the next drawer enters. Associative: dreams, memories, transformations — the teacup above.

What the day taught.

01
A hard geometric constraint eliminates the collaboration it was meant to organize. Strict bands eliminate overdraw. Strict bands also eliminate the blind-collision that is the whole reason two minds are drawing instead of one.
02
Coherence and freedom can coexist when conceptual ownership is separated from spatial constraint. Pre-planning the parts of the figure gives each model a role. Removing the geometric wall lets the line actually travel. The figure holds because the plan holds, not because the walls do.
03
Color freedom converges. Given the choice, both models reach for similar marine palettes on watery subjects. Locking each model’s stroke to a single identity color eliminates the convergence and makes the handoffs visible. The drawing reads as two hands instead of one diffuse palette.
04
Silent failures are worse than loud failures. Gemini initially returned non-JSON under this prompt about half the time, and the app fell back to a straight diagonal without flagging the error. Adding responseMimeType: "application/json" and surfacing empty candidates as retryable errors changed the output character completely.
05
The handoff geometry is not just scaffolding. Different modes produce visibly different drawings — the free teacup is not a down teacup with the grid erased. Removing the grid changes what the prompt author asks for, which changes what the models try to draw, which changes the drift. The instrument shapes the instrument-user.

What the instrument binds, what it leaves loose.

Locked

  • Stroke color (model identity)
  • Fill (none)
  • Line caps + joins (round)
  • Dash pattern (forbidden)
  • Pen lifts (stripped)
  • Start coord (handoff)
  • End coord (snapped to line)
  • Canvas bounds [0, 1]

Free, clamped

  • Stroke width [0.5, 20]
  • Opacity [0.1, 1.0]
  • Path complexity
  • Curve commands (M L C Q)

Unconstrained

  • Interpretation of the assigned part
  • Wander distance from nominal band
  • Aesthetic shape language
  • Number of self-intersections
  • How the models read each other’s absence
Research software, not a product eschaton corpse is a private workbench for looking at what blind LLM collaboration looks like under a narrow set of drawing constraints.