# Teaching FMVisual through guided project work This method prepares ten learners to continue FMVisual with decreasing teacher intervention. The shared project is a modest extension of the atrium that preserves visual quality and gives the visitor a believable boundary. Completion requires independent explanation, implementation, diagnosis and handoff, not only a finished image. Version 1.1 is a protocol to evaluate, rather than an established claim of teaching effectiveness. ## Learning outcomes Each learner should be able to explain a visible result using the relevant representation, predict a change before trying it, make a bounded contribution, verify it with appropriate evidence and enable another person to continue it. In particular, learners must distinguish visible triangles from lighting voxels, direct shading from indirect diffuse transfer, temporal stability from lighting correctness, and frame intervals from GPU timings. They must identify FMVisual's original work, Three.js implementations and planned features accurately. ## The teaching cycle Use the same cycle for each small investigation: 1. **Observe** a fixed view and describe a specific effect in ordinary language. 2. **Predict** what one change will do and identify the representation it affects. 3. **Change** one variable in an isolated exercise or reviewed contribution. 4. **Measure** with the same camera and settings, recording contradictory evidence too. 5. **Explain** the result, resource cost and limits without the teacher supplying the answer. 6. **Hand off** the work and observe where another learner needs help. Keep the original prediction after discovering a mistake. A plausible explanation written after seeing the result is not evidence that the learner predicted it. ## Six milestones ### Observe the same room Give everyone the same camera and light settings. Ask why the scene feels deep, how the shaded ceiling receives light and what is visible beyond the last doorway. Record individual answers before discussion. Let learners describe what they see before introducing terminology. Use the Methods Library to connect their words to parallax, indirect illumination and composed boundaries. ### Build a room individually Everyone completes the eight-box example and makes one controlled alteration to the roof or wall colour. Each learner must start the scene, compare direct and combined output, reset the camera and release resources. An unsupported GPU path is recorded as direct fallback, not scored as failed understanding of GI. The teacher demonstrates one investigation aloud, then leaves the keyboard with the learner. Pair discussion begins only after an individual first attempt exists. ### Explain the system Trace a frame from scene factory to displayed image. Each learner explains one ownership boundary to a partner and answers a question about an unfamiliar change. Use the manual's world, settings and backend contracts. Memorizing function names is insufficient if the learner cannot predict which resources or history must change. ### Extend in rotating pairs Arrange five pairs. Initial responsibilities are architecture, materials, lighting, navigation and integration. Label pairs A through E and responsibilities 0 through 4. In round r, pair i takes responsibility `(i + r) mod 5`. After five rounds every pair has performed every role. Within each session, exchange driver and observer roles halfway through; both learners make individual predictions and explanations. | Responsibility | Contribution | Evidence | | --- | --- | --- | | Architecture | A modest place with layered views and intentional edges | Fixed-camera views, bounds and route inspection | | Materials | Correctly scaled stone or fabric detail | Close and distant comparisons plus owned-resource disposal | | Lighting | Coherent sunlight and shaded-space illumination | Direct/combined comparison, voxel spacing and shadow coverage | | Navigation | A comfortable journey through reachable space | Actual input, full rail clearance and boundary cases | | Integration | Compatible contracts and reproducible results | Build, fallback, resource lifecycle and settings records | Do not assign one learner permanent ownership of a specialty. Group output can conceal that one person understands the system while another only follows instructions. ### Integrate under explicit budgets Preserve the original atrium fixture. Integrate one region at a time and record changes in bounds, voxel spacing, shadow coverage, frame intervals and resource ownership. Geometry visible on the horizon must not inflate nearby transport coverage without a deliberate policy. A further substantial wing needs lighting zones; current navigation does not provide collision or containment automatically. ### Transfer without the teacher Give each learner a new but comparably sized requirement. They use the documentation and available tools, but receive no teacher solution during the first attempt. Then another pair runs and modifies the contribution from its handoff instructions. The original authors observe rather than repair the recipient's work. Record questions, failures and time until the recipient can act independently. For human learners, repeat a comparable transfer after several days to examine retention. The exact delay and task must be set before the first session. ## Assistance and teacher restraint Record assistance when it happens, including its exact content: | Level | Teacher action | Interpretation | | --- | --- | --- | | 0 | No teacher help; normal reference material and tools are allowed | Independent attempt | | 1 | Ask the learner to inspect a representation or state a prediction | Guided attention | | 2 | Identify a contract, module or diagnostic output | Targeted hint | | 3 | Supply an approach, algorithm or code | Assisted completion | Ask what the learner expected, which observation disagrees, and what the smallest isolated test could be. Do not silently fix learner code and count the result as independent success. Stop a harmful or destructive action if necessary, record the intervention and let the learner resume from a safe state. ## Assessment before and after instruction Use the same rubric for observed initial work and a different transfer task. The written baseline questions observe explanation, prediction and cost reasoning; implementation, diagnosis and handoff remain unobserved until an actual attempt. Mark these N/O rather than zero and do not compare totals with different observed dimensions. Record individual scores before combining pair results. Each dimension is scored 0 through 3: 0 means absent or materially incorrect; 1 means partial or dependent on a solution; 2 means correct with supporting evidence; 3 means correct, independent and transferable to a new case. Level 3 teacher help limits the corresponding independence score to 1. | Dimension | Evidence needed for at least 2 | | --- | --- | | Explanation | Correct representation and ownership, with current versus planned capability distinguished | | Prediction | Written before execution, specific enough to be contradicted | | Implementation | A bounded working change satisfying the actual world and host contracts | | Diagnosis | A reproduced failure, a justified correction and a verification result | | Cost and limits | Correct spatial/resource reasoning and honest backend/measurement labels | | Handoff | A recipient can run, explain and make a follow-up change from the supplied instructions | The facilitator supplies an unfamiliar transfer variant after practice; repeating the worked 20-to-30-unit example does not qualify as a new task. A learner meets the pilot independence criterion only when every dimension scores at least 2, the transfer has no level 3 assistance, and functional checks pass. Report each dimension as well as the total out of 18. A strong image must not compensate for an incorrect resource-lifecycle explanation. The initial ten-person cohort is a descriptive pilot, not a causal comparison. Document prior experience, missing submissions, failures and assistance. Report individual changes and pair outcomes; do not treat two members of one pair as independent evidence of a collaboration effect. Practice, task difficulty and teacher scoring can explain changes. Use counterbalanced tasks or a suitable comparison cohort before claiming this method outperforms an alternative. ## Applying the protocol to AI learners An AI pilot evaluates whether agents can follow the packet, perform tasks, use feedback and hand work to another agent. It does not establish that human students learn, that novices are represented, or that model parameters changed. Start agents without the surrounding conversation. State their allowed references and output scopes, keep predictions before execution, preserve assistance messages, and prevent access to another participant's answers until handoff. If isolation is instructional rather than enforced, record that limitation. Use the same declared model and tool conditions across participants unless variation is itself the question. Record baseline knowledge rather than pretending the agents are beginners. A high baseline can expose a ceiling effect: successful work may demonstrate prior capability or instruction following, with little evidence of learning gain. Repeated calls to one model are not a representative sample of ten people. ## Evidence and revision The learner packet is in `teaching/`. Freeze the protocol and exercise version before the run. Keep dated raw predictions, submissions, check outputs, assistance, handoff observations and rubric assessments. The pilot report must distinguish technical rehearsal, AI-agent results and human observations. Revise only after recording the first result. Preserve the original task and the reason for each revision. Re-test the changed instruction on a new attempt so that an edited prompt is not credited retroactively for a result it did not produce. The first technical rehearsal may establish that the exercise runs and its checks distinguish a correct submission from a known defect. Only observed learner attempts can establish whether the instructions support independent continuation.