# Counter-Reading Slice Build Receipt

**Date:** 25 July 2026
**Status:** implementation verified
**Scope:** `living-stewardship-prototype` only

This receipt supplements rather than overwrites the historical Day One build
receipt.

## Requested slice

The build adds:

- a WebGPU-only browser model path using a pinned 360-million-parameter
  instruction model;
- a visibly labelled scripted fallback;
- explicit accept, overrule and revise responses;
- a hard minimum-length reason gate for overrule;
- a research receipt preserving the full loop and generator provenance;
- one declared remote model origin with bodyless asset requests;
- a fixed 36-participant counterbalanced research plan;
- preregistration, capacity and model technical notes;
- a GitHub Pages root doorway, empty `.nojekyll` and relative local paths.

## Existing-test treatment

No existing test was weakened, skipped or deleted.

One existing claims test was amended to include the new root `index.html` in
the public-file scan. The original 18 tests remain present. New test files
cover the research loop, network boundary and repository-subpath routing.

## Dependency and static-hosting boundary

The browser runtime is installed as a pinned development dependency, then its
browser bundle, WebGPU WASM support files and licence are copied into
`vendor/model-runtime/`. The demo imports those files through relative paths.
It does not load a hosted script.

The ONNX runtime is configured with:

- WebGPU device selection;
- one WASM thread;
- no proxy worker;
- relative WASM paths.

This avoids reliance on cross-origin isolation headers unavailable on ordinary
static hosting.

## Automated evidence

`npm test` returned **30 passed, 0 failed** in the working tree.

The same **30 tests passed without `node_modules/`** in a clean copied tree at
`/private/tmp/lsp-final-verification.p7KpM4`. This proves that the committed
vendored browser runtime, rather than an undeclared local installation, is
present in the artefact. The temporary path is verification evidence, not a
runtime dependency or public path.

The original 18 baseline tests remain and pass.

## Browser runtime evidence

The repository was served from its production-shaped subpath:

`/living-stewardship-prototype/demo/`

A supported browser completed:

1. deterministic baseline choice and authored result;
2. model asset load and local WebGPU generation;
3. provenance display:
   `Generated locally · SmolLM2-360M-Instruct · 6849e9f · WebGPU`;
4. rejection of a 39-character overrule reason;
5. acceptance of a reason above the 40-character minimum;
6. a final `lsp-research-receipt.v1` containing the verbatim generated text,
   initial position, complete generator provenance, overrule reason and final
   position.

Initial model output that echoed prompt material was rejected during build
review. The prompt now uses the model's chat format, supplies only the scenario
and authored positions, and rejects prompt echo or a direct contradiction of
the fixed “witness cannot answer” fact.

The unsupported-path fallback remains proved by deterministic tests rather
than misrepresented as a generated browser result.

The runtime emitted performance notices that some shape-related operations
were assigned to CPU while the model session used WebGPU. These notices did not
trigger fallback or change the recorded `webgpu` generator provenance.

## Vendored runtime integrity

| File | SHA-256 |
|---|---|
| `transformers.min.js` | `aa5002b70e789798da263f5f99c62bd3e8fcd0c119258a493c40c180648365fa` |
| `ort-wasm-simd-threaded.jsep.mjs` | `08fb86ec433c78bfb032c5d84a68b8e8e5a8d81268fa39e24314179a5767a5b9` |
| `ort-wasm-simd-threaded.jsep.wasm` | `c46655e8a94afc45338d4cb2b840475f88e5012d524509916e505079c00bfa39` |
| `transformers-LICENSE` | `cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30` |

## Claims boundary

The build does not claim:

- validated autonomy or participant outcomes;
- flow detection, cognitive measurement or physiological inference;
- model-output quality;
- a validated interpretation of overrule rate;
- connected wearables.

## Human and publication state

No participant data exists. Planned roles are not presented as appointed
people. Nothing has been pushed or deployed by this build.

## Completion marker

COUNTER_READING_SLICE_VERIFIED