# Invisible Physics Lab: first screenshot pilot

Recorded 2026-09-07 20:31 UTC.

**Pilot advancement gate: passed.**

| Measure | Result |
|---|---:|
| Attraction/repulsion correctly identified | 20/20 |
| Median source-location error | 1.01 world units |
| Median error from always guessing (50,50) | 15.70 world units |
| Pooled forecast position RMSE | 0.80 world units |
| Pooled straight-line RMSE | 11.09 world units |
| Pooled center-source RMSE, using Astra's polarity | 5.86 world units |
| Forecast RMSE reduction against straight line | 92.8% |
| Forecast RMSE reduction against center-source baseline | 86.3% |
| Trials beating straight-line forecast | 20/20 |
| Trials beating center-source forecast | 20/20 |

## What ran

Twenty randomized synthetic scenes were generated once. Each contains one stationary attractive or repulsive source, with x and y independently sampled within 30 to 70 in a 100 by 100 world. There are ten sources of each polarity, randomly ordered. No scene was discarded or retried. Astra was not told the polarity balance.

A fresh `gpt-6-astra` agent was explicitly requested with `ultra` reasoning and no inherited conversation. The main session's Ultra/Fast setting was reported by the user. The subagent interface does not expose or independently verify a speed setting.

The solver received five anonymous four-panel PNG sheets, the known force equation and constants, coordinate and time conventions, the source bounds, and launch conditions. It was instructed to use only the supplied image viewer, without reading files, executing numerical tools, extracting pixels, or viewing hidden settings. This is blinding by tool-use protocol, not an OS-enforced access boundary. The agent's final response includes its image-only attestation.

Each scene shows three paths A, B and C for 3 seconds, with dots every 0.25 seconds. Astra estimated the source coordinates and polarity. All 20 estimates were committed before the evaluator scored them, with no answer feedback or corrective retry.

The evaluator then ran the known simulator using those estimated parameters for the fourth launch D: initial position (88,85), initial velocity (-14,-10), duration 3 seconds. D's initial conditions were included in the actual solver prompt, although the generator's initial metadata file lists only observed launches A/B/C. D's trajectory was withheld. **The forecasts are simulator rollouts from Astra's estimates, not trajectory points directly produced by Astra.**

## Scoring and checks

Source error is Euclidean position error in world units. Motion error is the square root of mean squared Euclidean position error over the 60 equally spaced post-initial samples from 0.05 through 3.00 seconds. Pooled RMSE combines all 1,200 forecast positions. The straight-line baseline uses D's initial velocity with no force. The center-source baseline uses (50,50), the same known force, and Astra's predicted polarity, isolating the contribution of source-location estimation.

The local advancement criteria were saved before solver dispatch: at least 18 correct polarities, median source error no greater than 8 units, at least 50% pooled motion-RMSE reduction versus straight line, and at least 12 wins versus the center-source baseline. These are engineering criteria for this exploratory pilot, not a formal statistical preregistration.

Independent review found the force equation, RK4 integration, time alignment, metric denominators and baseline construction correct. All stored future paths reproduce exactly; a finer integration timestep changed two checked future paths by at most 1.53e-8 units. No observed or future path clips outside the frame. The PNGs contain no metadata. SHA-256 hashes for the predictions, source dataset, plan and five test images are in `work/scores.json`.

## Limits

This is one small run on one known synthetic force family, with fixed strength and launch patterns, inside one 20-scene solver session. It tests visual parameter inference and the usefulness of those parameters for an unseen launch. It does not establish discovery of unknown physical laws, autonomous experiment selection, real-world performance, or repeatability across independent runs. A numerical fitting algorithm was not used as a competing estimator.

## Every trial

| Trial | Polarity | Source error | Forecast RMSE | Straight-line RMSE | Center-source RMSE |
|---|---|---:|---:|---:|---:|
| 01 | Correct | 1.04 | 0.67 | 11.99 | 2.44 |
| 02 | Correct | 0.83 | 0.12 | 5.45 | 4.71 |
| 03 | Correct | 1.35 | 0.55 | 11.84 | 9.47 |
| 04 | Correct | 2.63 | 0.72 | 8.74 | 4.82 |
| 05 | Correct | 0.75 | 0.24 | 9.74 | 1.44 |
| 06 | Correct | 1.06 | 0.35 | 9.00 | 5.40 |
| 07 | Correct | 1.31 | 0.47 | 9.86 | 5.29 |
| 08 | Correct | 1.68 | 0.61 | 15.45 | 10.41 |
| 09 | Correct | 1.94 | 1.57 | 12.87 | 2.43 |
| 10 | Correct | 3.25 | 0.97 | 9.75 | 4.02 |
| 11 | Correct | 0.95 | 1.63 | 12.27 | 6.63 |
| 12 | Correct | 2.33 | 2.01 | 17.09 | 9.54 |
| 13 | Correct | 0.82 | 0.15 | 8.04 | 5.40 |
| 14 | Correct | 0.98 | 0.17 | 6.22 | 4.24 |
| 15 | Correct | 0.45 | 0.19 | 13.35 | 6.03 |
| 16 | Correct | 0.87 | 0.31 | 13.18 | 6.98 |
| 17 | Correct | 0.81 | 0.16 | 7.75 | 3.27 |
| 18 | Correct | 0.43 | 0.06 | 5.64 | 4.60 |
| 19 | Correct | 1.21 | 0.47 | 9.96 | 2.93 |
| 20 | Correct | 0.86 | 0.53 | 14.38 | 6.87 |