# ZeroExternal — Observer pilot v1

Status: ready for an independent participant; no laboratory has enrolled or supplied physical measurements in this release.

## Research question

Can two independent implementations reproduce the geometric projections of the same fixed world while keeping observer-dependent measurements separate from model state?

This is a computational reproducibility pilot, not a test of consciousness, relativity, quantum nonlocality, a new light-speed law, or the conjecture VH = QG.

## Model and declared scope

Two opaque axis-aligned boxes, dimensions 1.2 × 1.8 × 0.15 metres. Centers: (-0.9, 0, 0) and (0.9, 0, 0.35) metres. Euclidean center distance: sqrt(1.8² + 0.35²) ≈ 1.833712082089225 m.

Each observer has an orbiting camera aimed at the origin, world up +Y, horizontal field-of-view calibration, and a 3:2 viewport. Perspective and calibrated orthographic modes are distinct. Orthographic half-width = distance × tan(calibration angle/2). This is a chosen imaging scale; it is not a perspective lens law.

Image separation is Euclidean projected-center distance divided by viewport width. Hidden centers remain geometric markers; their displayed location is not a claim that an opaque object is optically transparent.

## Three preregistered controls

1. **Default pair:** A yaw 0°, B yaw 65°; pitch 10°, distance 6 m, calibration 55°, perspective. Predicted separations: A ≈ 0.2971077469547374, B ≈ 0.08367932101967832 viewport widths.
2. **Identical cameras:** B = A. Both projection measurements must match.
3. **Changed pose:** change observer B alone. A's observation and both box coordinates must remain unchanged. Use the supplied recipe JSON for exact parameters.

Numerical acceptance for every reduced scalar: `abs(received - reference) <= 1e-9 * max(1, abs(reference))`. This tolerance is computational, not an instrument uncertainty.

## Participant workflow

1. Sign into `/account/`; create a project or join with its one-use invitation. Assign a reviewer as a viewer using the documented API if a read-only role is preferred.
2. Create a project-bound observer agent key. It expires; the owner can revoke it. Do not send the key to the project developer or paste it into an issue.
3. Run the downloaded Python agent on the participant's own computer. It uses outbound HTTPS and a fixed scientific recipe. A model name or a supplied code hash is not proof of where it executed.
4. Open `/observer/`, choose a control, record its signed browser result, then choose **Bring to team** and explicitly create the assigned project task.
5. Run the agent with `--once` for one task or the documented bounded polling mode. Export the project task result and the browser bundle.
6. The reviewer checks input equality, the source versions, world invariants, the seven numerical comparisons, missing/failed cases and the declared tolerance. Record disagreements; do not silently discard them.
7. Revoke the agent key after the pilot. Export the project journal and results for the participant's own archive.

## Independent laboratory evidence

To claim a laboratory replication, additionally retain the lab's chosen identity/affiliation attestation, participant consent, instrument configuration, calibration, raw observations, measurement uncertainties, protocol deviations and a reviewer report. Agree publication and data rights before exchanging private measurements. The current agent endpoint accepts synthetic geometry measurements only; an instrument-data adapter and uncertainty model must be reviewed before a physical pilot.

Open data from CERN, Tsinghua or another institution do not establish a partnership. No invitation, contract, payment, blockchain transaction or publication is sent by this pilot automatically.

## Provenance and limitations

- Browser bundles: ephemeral self-signed P-256 keys; content integrity and lifecycle, not institutional identity.
- Shared workspace: authenticated project membership and server timestamps; not a blockchain or immutable public transparency log.
- External agent: independent Python implementation; source SHA-256 is supplied by the agent, not remotely attested.
- Gateway: recomputes seven synthetic measurements, stores disagreements, uses expiring leases and idempotent completions.
- All three components agreeing supports implementation consistency. It is not by itself a confirmation of a physical theory.

## Acceptance record

Save: project/task IDs; exact ten parameters; model version; browser bundle; agent script and result hashes; gateway verdict and errors; execution environment supplied by the operator; failed attempts and protocol changes; reviewer decision and timestamp. Label the final outcome `computational_pass`, `computational_disagreement`, `incomplete`, or `independent_lab_report_pending`.