docs/prereg_amendment_2.md

OPharm amendment 2: replication and 27B behavioral study

Replication scope

Retain the H1–H5 statistics, alternatives, margins, train/test partitions and five-test Holm family. Select each model using its own reference/development data. Keep the registered and dev_best points distinct.

9B and dense-comparator settings

Model	L_harm	L_ref	L_steer	Registered probe	Secondary dev_best
Qwen3.5-9B	14	14	23	t_post, 31	t_inst, 24
Qwen3-4B-Instruct-2507	28	19	26	t_post, 21	t_inst, 11

Use coefficient 1.0 and 24 random directions per reference family. Patch points are 1, 5, 9, 13, 18, 22, 26, 30 for the 9B, and 1, 6, 10, 15, 20, 25, 29, 34 for the comparator.

Additional mechanistic replications

Apply the same content-reference selection and operational-probe rules to Gemma-4-E4B and Llama-3.1-8B. Freeze their selected layer/probe values and intervention settings separately before their held-out analyses. Preserve architecture-specific tokenization, opener conventions and activation-handling checks. Do not transport another model's selected layer values.

27B behavioral scope

Pin Qwen/Qwen3.5-27B at revision fc05daec18b0a78c049392ed2e771dde82bdf654. Keep this behavioral study separate from activation-based H1–H5 replications.

Use the development split for the main C/N benchmark: all 40 templates, 216 skeletons and 3,456 prompts. For the harmless requester trigger, use the selected 108 development skeletons, 40 templates and 1,728 prompts, including both rollback states and both trigger variants.

For assistance, use the frozen held-out cohort defined by definitely covered consequence marks. Include DP, DS and BS; include BP only for templates whose benign twin is unambiguously outside access/outage coverage. Operational assistance arms are A0, A1, A2 and Iask, with A2u and A3 using their specified definitions. Conjunction arms are A0, A1 and A2, with A2u and A3 using the corresponding conjunction definitions. Collect tools-present operational q1/q2/q3/qapp and conjunction qreq/qqueue/qapp.

Use H100, bf16, loop solver, non-thinking greedy decoding, generation batch 32 and 256 new tokens. Decode knowledge judgments for four tokens at batch 32. Continue capped direct replies by 512 tokens at batch 8, preserving source-result hashes and remaining-capped status. Record actual launch arguments, model revision, hardware, instance hashes and cohort counts.

27B endpoints and predictions

Report main-set asking/execution by cell and rule, environment/target selectivity, coverage, and paired rule differences. Report trigger-present/absent behavior and paired correctness, with BS as the primary trigger interpretation.

For assistance, report asking where required and where not required, allowed-cell execution, operational paired success and conjunction paired success. Use the original-interface Iask ceiling; classify ability to ask only when its ask rate is at least 80% in each operational cell.

Use 10,000-resample paired intervals: skeleton-family strata for primary main/trigger intervals, template intervals for sensitivity, and template-clustered intervals for assistance. Read assistance changes against a +/-10-point margin.

Predictions:

• Main DS and BS ask rates are at least 30 percentage points lower under neutral than confirmation policy.
• BS trigger-present asks are at least 50%; trigger-absent asks are at most 20%.
• Initial 27B assistance is for estimation; there is no directional A3 prediction for 27B.

Keep A2-self conditional on the original A2 trigger: success improvement of at least 20 points with intervals excluding zero in at least two models. Do not substitute A2u into that trigger.

Freezing and reporting

Record behavioral study settings separately from mechanistic layer/probe settings. Preserve the initial study contracts and report all predefined arms within each model's scope. Document any later change before collecting the data it governs, with its reason, affected endpoints and version hashes.