Lab: Measure Latent Reconstruction Loss

Mission

Run the local Latent Reconstruction Lab. Compare a two-bit bottleneck with a one-bit bottleneck, inspect every reconstruction, and export a bounded evidence record.

Guided start

1. Run the two-bit bottleneck.
2. Trace pattern B: input 1101 → code 01 → reconstruction 1101 → error 0.
3. Switch only the bottleneck to one bit.
4. Trace B again: 1101 → code 0 → reconstruction 1001 → error 1.
5. State what changed and what stayed fixed.

Procedure

1. Preserve all four two-bit rows and total error.
2. Predict which rows will collide with one-bit codes.
3. Run one-bit mode and preserve all rows and total error.
4. Identify each collision and exact changed bit.
5. Inspect every available decoder output.
6. Preserve your first “generated output” claim, then correct it.
7. Complete Observation, Measurement, Decision, and Limitation.
8. Export the record and confirm it names ai-for-beginners-latent-reconstruction-lab.

Success criteria

• one full encode/decode trace is reproducible;
• both modes include four exact errors and totals;
• collision is distinguished from implementation failure;
• decoder output is not called novel, realistic, or correct without evidence;
• the limitation rejects training, unfamiliar-data, and GAN-performance claims.

Use only supplied patterns. The deterministic browser Lab trains no model, uploads no data, and contacts no service.