Some directions are research. They do not ship until they ship. A Lab Note records a hypothesis, an observation, and a next experiment. A Lab Note is not a public capability claim. A note that becomes a claim moves to the Architecture or Status page; it leaves Lab Notes.

A Lab Note carries:

• a hypothesis — the claim being explored;
• an observation — what has been seen so far, with whatever evidence supports it (a source path, an experiment record, a runtime trace, a citation, or an explicit statement that the observation is a hypothesis);
• a next experiment — what would change the observation, in either direction.

A Lab Note does not require a public capability claim's evidence. What it lacks, compared to a Status page row, is the qualifying evidence that would admit the work as a public capability. The observation is honest about its own limits.

• Engram-like sidecars. Hypothesis: small learned modules attached to the canonical state carry conditional memory that survives plan boundaries. Observation: the engine's engram_learning module implements the scaffolding; the sidecar's behavior under repeated admissions is not yet measured. Next experiment: a controlled fixture that exercises a sidecar across three plan identities, with the receipts compared.
• KV-cache compaction policies. Hypothesis: a learned eviction policy can match a hand-tuned one within a small loss bound, on representative workloads. Observation: the engine's kv_cache_compaction module implements the policy; the comparison run is not yet recorded. Next experiment: a side-by-side comparison on a fixed workload, with the receipts published.
• Shadow calibration. Hypothesis: shadow calibration on a held-out calibration set produces a quantization plan that preserves the chosen metric within a stated bound. Observation: the engine's shadow_calibration module produces a plan; the bound is not yet stated. Next experiment: a calibration run that produces a plan with a named bound, with the plan and the bound published.
• Learned adapter training. Hypothesis: a small adapter training loop on a fixed task produces a delta that admits to the same plan identity. Observation: the engine's adapter_training module implements the loop; the delta's plan-identity implication is not yet tested. Next experiment: a controlled adapter run, with the delta's plan identity compared.

Each note above is illustrative. The published notes are read from the data layer; the manuscript names the kind of note the layer holds.