Exp05: causal subspace bridge in GPT-2-small (exploratory)

Exp05 causal subspace bridge in GPT-2-small (exploratory)

Highlights

• Across eight registered seeds, L7H4 + L8H5 were the smallest tested set that passed the frozen
  number-specificity and transport controls in every seed (0.521–0.546 of the frozen direct effect).
• The fixed 12-row layer-8 SAE decoder span retained R_span = 0.8935 (t(7) CI
  [0.8905, 0.8966]) and beat the frozen second-largest edge of 100 matched spans in every seed.
• A fresh exploratory bridge retained R_target = 0.6786 (t(7) CI [0.6738, 0.6834]) and beat
  the maximum of 100 target-excluded matched spans in all eight seeds. The complete L8H5 hook_z@final
  clamp arm remained 0.6740 (CI [0.6696, 0.6785]).

Evidence and provenance

• Research brief
• Compact results packet
• Full writeup
• Bridge figure
• The model-backed raw bridge result remains outside Git. The checked-in packet is a hash-bound
  reaggregation of compact CSV rows and the historical provenance receipt; it does not revalidate the
  missing raw result in the current invocation.

Claim boundary

The reported values are directed logit-effect ratios, not activation-reconstruction percentages. This
exploratory bridge has no preregistered verdict. The release does not claim natural or monosemantic latent
semantics, individual-latent causality, necessity, sufficiency, mediation, a complete circuit, or
generalisation across models or tasks.

Reaggregation

The compact presentation can be regenerated without model execution:

python3 experiments/05_number_agreement_circuit/make_bridge_summary.py --reaggregate-checked-in

That command validates the frozen compact CSV hashes, recomputes the seed-level summaries, and records
that the off-Git raw result was not reopened.