AI-Native PDF Engine

This repository is implementing the founding plan in PLAN.md. The first executable milestone is the Phase 0 representation lab: a solver-free harness for measuring whether models can reliably consume versioned state blocks and spatial text grids.

Representation lab

Fixtures contain small-integer, hand-solved geometry and semantic element IDs. The CLI renders blinded model prompts and scores boolean spatial answers directly from fixture geometry. Conditions separate state format from grid exposure: state-only flat table, state-only element blocks, grid-only, and each state format with the grid included. This prevents the grid from masking a state-format comparison while measuring grid readability as part of the same run.

cargo run -p representation-lab -- prompt \
  fixtures/editorial-cover.json flat-table+grid
cargo run -p representation-lab -- prompt \
  fixtures/editorial-cover.json grid-only
cargo run -p representation-lab -- score \
  fixtures/editorial-cover.json \
  examples/answers/editorial-cover.correct.json
cargo test --workspace

The scorer rejects an answer sheet whose fixture ID or state version is stale. This makes the staleness protocol part of the experiment rather than an informal prompting convention.

Recording an experiment

A trial submission records the model identity, assigned prompt condition, latency, token counts, and the model's answer sheet. record scores it and appends one immutable record to JSONL. summary reports accuracy, missing-answer rate, average cost, and per-fixture rows for every condition. Its decision block applies the preregistered thresholds to flat-table+grid versus element-blocks+grid.

cargo run -p representation-lab -- record \
  fixtures/editorial-cover.json \
  examples/trials/editorial-cover.flat-table.json \
  results/state-format-v0.jsonl

cargo run -p representation-lab -- summary results/state-format-v0.jsonl

Trial IDs are caller-assigned so an orchestration script can randomize and blind conditions without the lab inventing experimental policy.

Next experiment slices

1. Add repair-batch fixtures to compare whole-draft repair with incremental verbs.
2. Add randomized, blinded trial orchestration across model providers.
3. Calibrate the spatial questions across 10–15 briefs before building the Phase 1 solver.