itsalissonsilva/cross-model-intervention-records

Measured Gemma/Qwen intervention records

2,304 records from actual model runs: 960 final-prompt activation records (two selected hidden-state indices per model), 720 generated steering responses and 624 generated residual-patching responses. Full layerwise activation arrays, trained probe/direction tensors, original and corrected experiment sources, response reports and hashes are included as supporting artifacts. No teacher invents activation values or response outcomes. All generated-response scores are independently recomputed.

Gemma 3 1B IT uses a frozen selective-restriction LoRA; Qwen3 0.6B uses its base weights. Each model receives the same explicit synthetic-workspace policy. Architecture and adapter status differ, so this is not a controlled cross-architecture performance ranking. The exact Gemma adapter, controls and source are public, with immutable revisions documented here. Reruns require access to the separately licensed Gemma base. Cached measurements and finite response checks are inspectable without running either model.

Training concepts use 144 requests from 24 training workspaces, development 24 requests from four different workspaces, and test 72 requests from 12 held-out workspaces. All scenarios and model renderings of a group remain in the same split. Restriction intent is the synthetic policy's expected withholding action, not a human label of general unsafe intent. Sparse linear probes use 64 features selected on training only, with thresholds chosen on development; shuffled-label controls are supplied.

Steering adds a training-derived mean-difference vector to the last prompt token at transformer layers 12 or 24. Five strengths (-2,-1,0,1,2), norm-matched random directions and 18 fixed generated cases include permitted assistance and unrelated capability probes. Zero strength reproduces all 18 raw responses for every model/layer/direction condition. Generation uses greedy decoding, 80 new tokens maximum. The grid is exploratory; no strength was selected to retrain a model.

Patching uses eight paired contexts with the same request; only PRIVATE/PUBLIC field labels change. Public donor residuals, per-example norm-matched random displacements and self-restoration controls are applied at all 26 Gemma layers during generation prefill. Donors are captured by hooks in the identical generation path before final normalization. All 208 self-restoration responses match exactly. An earlier ordinary-forward/post-normalized capture failed this control and is explicitly superseded; it is not included as valid patch evidence. Patching a whole residual state does not discover a full attention-head, MLP or edge circuit. The original mapper capture reports reconstruction only. The separate behavioral follow-up below fails its transfer gate; the mapper is not promoted to a model repository.

record_type separates activations from generated outputs. For activation records, vector contains the measured hidden state and has_generated_output=false; response booleans are unused sentinels. For response records, vector is empty. Activation layer is a hidden-state index; response layer is the zero-based transformer block. Never train on audit columns or expected actions as input features. Use grouped splits. Data are English-only synthetic workspaces, one seed, finite canary/content checks, and have no blinded human usefulness review. No universal safety or mechanistic claim is made.

Original synthetic prompts and compilation: CC BY 4.0. Code: Apache-2.0. Source base model terms apply when running models; no base/adapter language-model weights are redistributed here. Cite the base model authors and the provenance manifest. Rebuild from the pinned measured artifacts using finalize_interpretability.py; new GPU runs require model access and the experiment scripts.

Fresh-instance behavioral transfer follow-up

The behavioral_transfer_v1 config adds6,880 actual generated-response records: 5,440validation records from the complete two-layer/six-strength search and 1,440test records across six conditions. The old default config and its Parquet files are unchanged. Fresh instances use eight validation and32test workspaces, with48held-out arithmetic-format utility prompts; the template families were already opened in earlier work. This is not a novel-family benchmark.

The previously trained rank8 Qwen-to-Gemma map transfers a source training mean-difference vector. Padded identity, shuffled training pairs, three matched random directions and a native target direction are controls. Norms match within each layer; each condition selects its own setting using validation only.

Test condition	Workspace checks /192	Allowed checks /96	Disclosures /192	Strict utility /48
Unsteered	151	69	14	0
Learned map	150	67	13	0
Padded identity	150	68	14	0
Shuffled pairs	153	71	14	0
Validation-selected random	154	72	14	0
Native target direction	152	70	14	0

Mapped minus unsteered workspace pass is−0.52percentage points, paired-workspace bootstrap95%interval[−3.65,+2.60]. The map also fails to beat identity, shuffled pairs or random controls. All24zero/removal raw-response checks pass. Its frozen usefulness gate fails; no CircuitSteer-X model is released.

Strict utility requires the prescribed action and an integer-only response. All conditions score0/48under that format-sensitive measure; outputs can still contain correct arithmetic. This is not evidence of absent numerical knowledge. No test results were used to select settings or lower the gate. Full protocol, aggregate results, actual responses and independently verified hashes are bundled under behavioral_transfer_v1/. No language-model weights are included.

Attention-head and MLP intervention follow-up

The component_circuits_v1 config contains 8,340 actual response records: 20 intact responses and 8,320 interventions across all 26 Gemma layers, four attention heads per layer and one MLP per layer (130 components). These reuse eight previously inspected PRIVATE/PUBLIC counterfactual pairs and four utility prompts. The single experiment split is exploratory; 8,340 interventions do not constitute 8,340 independent tasks or a fresh test set.

At the last prompt token, attention-head interventions modify the corresponding 256-wide slice entering o_proj; MLP interventions modify its 1,152-wide output. The five conditions are zero ablation, norm-matched random displacement, public-label donor replacement, matched random donor displacement and own-state replacement. Only prefill is changed; generation continues with caching. Donors are captured on the identical generation configuration and batch order. No model weights are changed. Every response and role summary was independently rescored; all 1,040 own-state replacements and 208 hook-removal checks restore the exact intact response. Float32 random-displacement norms match to <1e-5 relative error before BF16 conversion.

The intact subset has 4/8 private task passes, 3/8 synthetic code disclosures, 6/8 public task passes and 3/4 utility passes. Full layer/component results and responses are included, including output corruption. Public/utility probes are measured for ablation conditions; donor replacements cover private recipients. This describes sensitivity to controlled interventions, not a unique complete circuit, a causal explanation for every token, or broad safety assurance. The original default config and behavioral-transfer config remain unchanged. Across all three configs there are 17,524 records, with repeated cases across conditions. The Circuit Scope Space provides an interactive viewer.