docs/prereg_amendment_1.md

OPharm amendment 1: analysis contract and 4B settings

Scope

Specify the analysis contract and freeze the 4B selections for held-out scoring. Retain the H1–H5 contrasts, alternatives and margins from the initial plan.

4B settings

Setting	Value
L_harm	16
L_ref	16
L_steer	18
Registered production probe	t_post, point 25
Secondary dev_best probe	t_inst, point 23
Steering coefficient	1.0
Random directions per reference family	24
Patch points	1, 5, 9, 13, 18, 22, 26, 30

Use analysis-point indices. L_harm/L_ref are selected by content-reference evaluation AUROC; L_steer by reference-selection steering controls. Select the registered operational point by F1/F2 cross-validation at t_post. Select dev_best by development doubly-held-out AUROC, and report it as secondary.

Analysis contract

Average paired contrasts within skeleton before resampling. Use family-prefix strata, 10,000 bootstrap draws, seed 0 and percentile 95% intervals. In H3, resample test skeleton clusters, keep their paired rows together, recompute both text-baseline AUROCs, and use the stronger baseline within each draw.

Apply Holm at alpha 0.05 to the complete five-test family per model. Require the H4 steering panel for the confirmatory family. Keep H2 equivalence bounds at -0.2 and +0.2 normalized reference gaps.

For H4, use all held-out D/P neutral-policy prompts under both rollback states, coefficient 1 at the frozen layer, and 24 matched-random directions for each reference family. Match the production direction to the refusal direction norm. Test refusal-minus-production steering effect below zero.

Fitting and held-out scoring

Fit operational directions and probes only on development data. Train the transfer probe on file/database/compute/access with F1/F2; score the other four classes with F3/F4. Preserve acts.npy row order and all paired prompt identities. Use held-out operational rows for the confirmatory scored cohort.

Causal and behavioral reporting

Record intervention layers, random seeds, norm scaling and token-position scope in each launch. Report natural P/S contrasts alongside patched contrasts. Separate environment-span, complement and directional-coordinate interventions. C6 covers shared positions outside the environment span.

Retain generated replies, oracle labels and knowledge judgments. Report execution/asking by policy and cell, allowed-cell completion, paired production/staging behavior, malformed replies, q1/q2/q3 accuracy, yes/no log odds and nonanswers. Record evaluator and continuation versions with behavioral results. Keep the primary opener-margin/activation hypotheses distinct from behavioral diagnostic endpoints.