# Teaching method technical rehearsal and human pilot preparation

The teaching method and classroom packet are prepared for ten real students.
Technical rehearsals on 21 October 2026 tested the exercises and diagnostic tools;
no human or AI learners participated. The human pilot remains to be conducted.

## Predictions and conditions

The original method and activity were frozen at 09:41 UTC before the first run at
09:43 UTC. The original source snapshot, SHA-256 hashes, predictions and raw results
remain in `teaching/evidence/2026-10-21-technical-rehearsal/`. Version 1.1 was frozen
at 09:47 UTC before the revised run at 09:48 UTC in the adjacent versioned directory.
Every subsequent run writes a new dated result file; `latest.json` points to the
most recent record rather than replacing the original observations.

The checks use a fresh headless Chrome context, a 1000 by 800 viewport and actual
WebGPU on an Apple Metal 3 adapter, followed by forced direct fallback. Captures
follow 60 reported frames per scene/view condition. The fixed eight-box room has
identical camera and host settings within each comparison. Pixel values describe
the presented image, not linear physical light measurements or GPU execution time.

## Observed results

| Investigation | Result | Teaching implication |
| --- | --- | --- |
| Wider roof slot in Direct | Bright pixels in the predefined bottom-40-percent region rose from 8.93% to 18.33%. Direct fallback gave essentially the same result. | A useful first prediction with an easily visible change. |
| Red wall changed to blue under starter lighting | The original neutral-region blue-minus-red mean shifted by about .36 of a displayed 8-bit channel value. | A subtle result is a poor first demonstration for an unfamiliar observer. Keep it as evidence rather than blaming a student. |
| Revised colour lesson at sun azimuth 60 and elevation 27 | In the preregistered neutral-receiver region, the channel shift was about 25.94; repeated-red variation was about .054. The red return capture resembled the earlier red captures. | Use this more legible condition for the initial colour exercise. Both light orientation and receiver region changed, so this does not isolate which change improved legibility. |
| Partial settings update | The isolated host raised `Unknown scene: undefined`. Restoring the complete settings recovered it on both backends. | Ask which required fields disappeared, as well as what the exception says. |
| Density-preserving expansion | The real resolver returned 120 at .25 units per voxel for a 20-to-30-unit span expansion from reference resolution 80. Omitting reference span returned 80 at .375. A 280 request was rejected by the 256 guard. | Detect silent spacing loss separately from whether a configuration is valid in other uses. |
| Teardown and page access | Canvases were removed, no unexpected browser errors were recorded, and the production manual and packet links resolved. | The exercises can be commissioned without changing the production renderer. |

Screenshots are under ignored `artifacts/teaching-*`. The first and revised raw
runs remain separately identifiable. Re-running on another browser/device may
change image values; the original measurements are not universal tolerances.

## Changes made after the first observation

Version 1.1 fixes both colour variants to sun 60/elevation 27 and prompts learners
to inspect whether the source wall is actually illuminated. The diagnosis prompt
asks about omitted required fields rather than treating a scene-name exception as
the whole explanation. The revised colour condition was then tested with red,
repeat-red, blue and return-red captures in that order.

Assessment preparation also exposed two design issues. Written baseline answers
cannot establish implementation or handoff, so those dimensions remain N/O until
observed. The worked 20-to-30-unit calculation is practice; the teacher supplies an
unseen numerical and camera variant for the independent transfer. These are protocol
corrections, not measured improvements in student understanding.

## Materials for the human cohort

The learner packet includes baseline questions, the room activity and an individual
observation record. `teaching/COHORT-PLAN.md` assigns ten participant IDs to five
rotating pairs and a recipient handoff route. `teaching/FACILITATOR.md` provides
setup, an adjustable first session, assessment keys and two unseen transfer variants.

Before the cohort begins, the facilitator supplies session dates, checks prior
experience and backend access, and freezes the exact materials used. Collect each
student's prediction, first attempt, help received, diagnosis and recipient outcome.
Set the delayed-transfer date before the first session. Preserve absences and
unobserved dimensions in the results.

The primary descriptive outcome is independent transfer: each observed rubric
dimension at least 2 of 3, functional checks passed, no teacher-supplied transfer
solution, and an actual recipient continuation. Report dimensions and assistance
separately. Ten students with paired work do not establish that the curriculum
outperforms another method; that needs an appropriate comparison and stronger design.

## Reproduction

```sh
make check
npm run test:teaching
```

The command verifies the frozen version 1.1 materials, runs the technical fixture
against production output and writes a fresh timestamped record. It neither creates
student submissions nor assigns learner scores. Human outcomes will be reported only
from the cohort's observed work.