Thirty-Seed Wave-Causal Uncertainty Study

Question

Once complex and phase-blind selectors actually choose different interventions, does phase-aware (or balanced) acquisition beat its controls on uncertainty-sensitive outcomes?

Context

Confirmatory follow-up to Five-Node Intervention Discovery and the primary empirical result of Wave-Causal Circuit Search. The pilot showed phase cancellation without a clean selector contrast; this study freezes a harder contract motivated by Complex Phase Representation and documented in Thirty-Seed Wave-Causal Uncertainty Study.

Durable prior ideas under test:

• A manipulation check is necessary before a mechanistic comparison: if selectors pick the same targets, “phase value” is confounded, not tested (Five-Node Exact Intervention Pilot).
• Saturated MAP rank hides small effects; primary outcomes here are log true-graph mass, expected SHD, and blind transfer loss.
• Vectors are typed encodings of a declared causal vocabulary, not learned embeddings — exact arrays and hashes decide claims (Thirty-Seed Wave-Causal Uncertainty Study).

Findings

• Manipulation check passed. Complex and phase-blind selectors chose different interventions in 20 of 30 paired seeds (floor: 18) across 9 distinct ordered target pairs (floor: 3). The phase comparison was eligible.
• Eligible is not positive. All four eligible superiority claims — balanced vs observational, complex vs observational, complex vs phase-blind, phase-blind vs observational — were killed at frozen preregistered gates.
• Some expected-SHD effects pointed in a positive direction but sat far below the fixed 5% magnitude gate; co-primary outcomes did not pass. Sign ≠ verdict.
• Two further contrasts (balanced vs random; raw random policy) were preregistered as diagnostic-only and carry no confirmatory weight either way.

Method

Thirty paired confirmatory seeds; frozen preregistration with per-claim kill gates; canonical hash-verified artifact wccs.thirty_seed_uncertainty_study.v1 with a runtime sidecar. Search uses verifier-induced proposal mass, not a calibrated posterior. Targets: natural-log true-graph proposal mass, expected directed SHD, blind causal-transfer loss. The published notebook only reads the artifact — it never imports or restarts the confirmatory runner.

Limits

• Restricted to a fully observed, linear, acyclic five-node synthetic family — no latent confounding, measurement noise, nonlinear dynamics, feedback, or real intervention cost.
• Preregistration was internal and commit-anchored, not external. Sensitivities (allocation, beta-by-sparsity, blind-prefix, cancellation-source) cannot overturn failed primaries.
• Passing manipulation does not imply a treatment effect; complex phase awareness showed no preregistered superiority here.
• Do not retune this failed study. Any larger graph, learned encoder, or physical-circuit successor needs a new frozen contract (Recursive Discovery and Evolution).

Artifact

Executed snapshot of the committed study artifact; hash and source commit pinned above. Canonical JSON and runtime sidecar live under artifacts/ in wave-causal-circuit-search, beside notebooks/thirty_seed_uncertainty_study.py. The notebook route leads with a reading brief and claim scorecard over the frozen JSON.