Task-local exposure record

This is a maintainer summary of the implementing agent's record from 2026-10-30 (~02:48 UTC), not a replacement attestation. The unabridged record remains pinned to the reviewed commit (SHA256 70fa76cec4f90f36c1b809e5acf9aee903830f56c4beb7d865b2c8105c3c5a5e). The sealed delivery archive has SHA256 5257385a86089d787eee6adf80102872d5fd84c3e0858828bc2ccac1a99cc1ab.

Inputs and access

The configured model was claude-opus-5-5, using Claude Code 2.1.301 in a fresh process without inherited transcripts, saved project state, browser access or MCP tools. The implementer could use shell, read and edit tools within the isolated process. Its own record does not independently verify the backend model identity beyond that configuration.

The approved inputs were the behavioral specification and adapter protocol, fixture observations, capture/replay tools, a Rust scaffold, vendored dependencies and compiler/runtime executables. Executable references were GEOS 3.14.1, Shapely 2.2.0, NumPy 2.5.4 and the tile-encoder adapter; Rust was 1.99.0. The original checkout, repair source, repository history, host home and other conversations were unavailable. Outbound network access was disabled apart from the restricted model-inference transport. Package-integrity and boundary checks passed before implementation.

The implementer reported reading the approved inputs and a non-geometry CLI launcher, plus environment metadata and directory listings. It reported no access to GEOS, JTS, PostGIS or Terrio repair implementation source. Shapely and GEOS were called through public APIs, without opening their implementation or test source. The tile-encoder binary was executed, not disassembled.

Input identities

Design observations and background knowledge

Black-box probes covered encoder orientation/rotation sensitivity, pinched bays, chords, overlaps and degenerate exteriors. A behavioral hypothesis using public polygonize/union primitives agreed with executable linework repair on 400 random small cases; an output-vertex hypothesis agreed on 353 invalid cases. The independent design records how those observations determined the fill rule.

The implementer reported general knowledge of computational geometry, including planar arrangements and exact predicates, and documentation-level awareness that GEOS/PostGIS linework repair nodes boundaries and builds area. It stated that no remembered implementation source was reconstructed or translated. A fresh context and access controls do not establish a model's pretraining history; this record describes task-local inputs and observed access.

No accidental prohibited-source exposure was reported. During implementation, one source-exposed integrator message conveyed only the approved fallible API and validation priorities; it supplied no algorithm details. The interface decision remains recorded. Detailed process paths and transport configuration are historical evidence in the pinned original, not contributor setup requirements.