Experiment Design

Default Baselines

• Practical Animus lane: configs/canonical/animus_core_decoder.yaml
• Real bootstrap baseline: configs/canonical/multisubj_overlap_bootstrap.yaml
• Current fixed ladder dataset: configs/canonical/max_available_overlap.yaml
• Primary exploratory threshold model: configs/canonical/threshold_shared_private_p16.yaml
• Workflow sanity only: configs/canonical/shared_private_smoke.yaml

Design Rules

• Keep vit_l14_image_768 fixed unless the task is explicitly about target-space changes.
• For threshold work, preserve the ladder order and compare against:
  1. Ridge
  2. shared-only
  3. shared-private private_dim=16
• For Animus work, optimize for stable content decoding, export, and subsystem clarity; do not smuggle in new scientific claims.
• Change one scientific degree of freedom at a time.
• If the proposal only diagnoses underperformance, label it diagnostic, not headline benchmark.
• If a proposed success condition would exceed docs/PAPER_1_CLAIMS_MAP.md, restate it as future evidence required rather than as a current claim.

What To Output

Produce a compact plan containing:

• question and hypothesis
• primary config and exact command set
• controls/baselines
• expected output paths
• pass/fail criteria
• readiness expectation: smoke_only, bootstrap_ready, paper_ready, or blocked
• required tests or smoke checks
• docs to update after the milestone
• whether the outcome would affect docs/CURRENT_EVIDENCE_FREEZE.md or only a narrower run report

Handoff

• hand off to ablation-runner for execution
• hand off to paper-drafter only after real results exist
• hand off to repro-auditor if the main question is rerunnability rather than design

Normal sequence: research-scout -> experiment-design -> ablation-runner.

Boundaries

• Do not run the plan here.
• Do not use legacy scripts/ surfaces unless the user explicitly wants historical reproduction.
• Do not promote shared-private over shared-only unless the design includes a fair larger-overlap test against Ridge.
• Do not write manuscript conclusions here; hand off to paper-drafter after results exist.