# Living Stewardship Prototype

**Status: early research prototype · locally runnable · no participant
results**

The baseline requires a participant to choose one of three authored positions,
then records the chosen branch and its authored result. The project calls this
required pause **ritual friction**. An optional second path loads a small model
in the browser, generates an argument against the initial position and requires
the participant to accept, overrule or revise.

A separate route opens only after the browser verifies a key locally, without
an account or holder record. The project calls this access pattern
**custodianship**.

The repository tests whether these mechanisms work. It does not demonstrate an
effect on autonomy or any other human outcome.

## Three-minute reviewer route

Serve the repository root so the demo can load its relative vendored runtime:

```sh
python3 -m http.server 4173
```

Open `http://127.0.0.1:4173/demo/`.

- **Baseline:** choose a position, inspect the authored result and replay its
  receipt. No model is loaded and the application initiates no request after
  its static module graph loads.
- **Counter-reading:** choose the counter-reading study condition, make an
  initial judgement, wait for the first model download, then accept, overrule
  or revise. Overruling requires a reason of at least 40 characters.
- **Offline key route:** use the public demonstration key in the
  [reviewer guide](docs/grant-reviewer-guide.md).

Run the complete evidence suite:

```sh
npm test
```

## Demonstrated in the repository

- no gate position is preselected;
- the baseline cannot continue without an explicit choice;
- each choice displays a different authored branch title, result and next
  prompt;
- baseline receipts reconstruct the same scenario version and authored result;
- the optional counter-reading layer is configured for a pinned 360-million-
  parameter instruct model running through WebGPU;
- unavailable WebGPU or rejected generated output produces a visibly labelled
  scripted fallback;
- the final response must be accept, overrule or revise;
- an overrule cannot close without a reason of at least 40 characters;
- the research receipt preserves the initial position, verbatim
  counter-reading, generator provenance, response, reason and final position;
- one declared origin supplies model assets through bodyless requests;
- model assets are cached by the browser after the first successful load,
  subject to browser eviction;
- the baseline and sealed-key paths send no participant data.

## Proposed, not demonstrated

The proposed 90-day work is a **36-participant exploratory study** comparing
the baseline and counter-reading conditions in counterbalanced order. The
primary measure is the overrule rate for model-generated encounters. Written
reasons are the secondary measure and would be coded independently by two
raters using a published scheme.

No participant has been recruited. No overrule-rate result, validated measure
or human benefit exists. See the [research plan](docs/90-day-research-plan.md)
and [preregistration](docs/preregistration.md).

## Evidence and limitations

| Statement | Status | Limit |
|---|---|---|
| Explicit judgement is required | Demonstrated by interface and tests | Two authored scenarios |
| The choices display different authored states | Demonstrated by branch and replay tests | Textual difference, not participant validation |
| Baseline receipts reconstruct displayed results | Demonstrated under the defined replay rule | Run ID and timestamp are provenance only |
| A key opens a session-only sealed route | Demonstrated by local hash verification | Public reviewer key; secrecy is not tested |
| The counter-reading receipt enforces response and provenance fields | Demonstrated by code and tests | Browser-model output quality is not established |
| Friction may reduce excessive deference | Working hypothesis | No participant results |
| Overrule rate may describe patterns of deference | Proposed measure | No validated interpretation |

Full limits: [Evidence and Limitations](docs/evidence-and-limitations.md).

## Model and network boundary

The model runtime is vendored in this repository rather than loaded as a hosted
script. In the counter-reading condition only, it requests pinned model assets
from `https://huggingface.co` using `GET` or `HEAD` without a request body.
Prompts, reasons, responses and receipts are not included in requests. The
runtime is configured for WebGPU and single-threaded WASM support so it does
not depend on cross-origin isolation headers.

“Curriculum-bounded” means the prompt contains only the approved scenario,
three authored positions and selected position; there is no search, retrieval,
private corpus or participant profile. It does **not** mean that pretrained
knowledge has been removed. See the
[model technical note](docs/model-layer.md).

## The Enforced Claims Register

Restraint here is enforced by a test: `tests/test-claims.mjs` fails when public
prose turns a prohibited item below into an affirmative capability or result.

<!-- CLAIMS_REGISTER_START -->

The repository prohibits claims of:

- a working Flow-State Algorithm;
- flow detection;
- cognitive-resilience measurement;
- Proof-of-Flow;
- physiological inference;
- validated autonomy benefits or participant outcomes;
- formal proof that PFLG is a new logic system;
- scenario impossibility for deterministic AI;
- an operational full Paradox Engine;
- completion of Edition 0;
- a physical edition currently for sale, in print or distributed;
- connected wearables;
- subliminal influence;
- demonstrated recursively amplified model capability.

<!-- CLAIMS_REGISTER_END -->

Qualified discussion of future research remains permitted only when visibly
labelled **proposed**, **hypothesis**, **absent** or **not demonstrated**.

## Privacy

Run data remains in browser memory unless the participant downloads a receipt.
No receipt is uploaded. The offline key is hashed locally, compared, cleared
and omitted from the receipt. No identity is requested or created.

## Project position

- **The Stewardship Model** — subtitle: *The Living Editions of Principia
  Paradoxica* — is the research and governance proposition.
- **Ludus Paradoxica** is the broader playable discipline.
- **Principia Paradoxica** is the wider publication and artistic project.
- **The Paradox Engine** is future consequence infrastructure and is not
  operational in this repository.

The Living Stewardship Prototype is a bounded experiment, not the full form of
those wider projects.

## Project files

- [Concept](docs/concept.md)
- [The Stewardship Model](docs/the-stewardship-model.md)
- [Offline Key Access](docs/custodianship.md)
- [The AI Role](docs/ai-role.md)
- [Research Question](docs/research-question.md)
- [System Boundary](docs/system-boundary.md)
- [Evidence and Limitations](docs/evidence-and-limitations.md)
- [Ethics, Privacy and Consent](docs/ethics-privacy-and-consent.md)
- [Grant Reviewer Guide](docs/grant-reviewer-guide.md)
- [90-Day Research Plan](docs/90-day-research-plan.md)
- [Preregistration](docs/preregistration.md)
- [Delivery Capacity](docs/capacity.md)
- [Local Model Technical Note](docs/model-layer.md)
- [Proposed Interaction and Sensing Roadmap](docs/zoan-fsa-wearables-roadmap.md)

## GitHub Pages

The root [project doorway](index.html) links to the demo and evidence. All
local demo paths are relative so the site works from the
`/living-stewardship-prototype/` subpath. The empty `.nojekyll` file permits
the `_reports/` directory to be served. The Pages test starts a real HTTP
server at that subpath and loads the doorway, demo assets and browser module
graph from it.

## Licence

Project code is licensed under the MIT Licence. Documentation is licensed
under CC BY 4.0. Vendored runtime files retain their included licence.

The Principia Paradoxica narrative corpus is not licensed under these terms and
is not present in this repository. See [LICENSE](LICENSE).