Interventions

Replace a sublayer or residual activation with a source (or zero); an Intervention.

site names where to patch (attn_out, mlp_out, or resid_post at a layer). source is the full-sequence source activation (1, T_src, d) to splice in; when mode is "zero" the activation is replaced with zeros and source is ignored (the crude ablation baseline). label is a short human tag used in the component name.

Replace a single attention head’s contribution by editing the o_proj input slice.

The attention output is o_proj(concat_h head_h), so patching head h means overwriting its d_head slice of the o_proj input with the source side’s slice while leaving every other head untouched. site is Site(layer, "head_out", head); source_head is the source per-head input (1, T_src, d_head) for that head.

Add a precomputed delta to a site’s activation; the path-patching primitive.

Path patching perturbs only the sender to receiver path by adding the sender’s source-minus-target residual contribution at the receiver’s input, holding every other path clean. site is the receiver read point (typically Site(layer, "resid_pre")) and delta is the (1, T, d_model) residual adjustment to add there. This is the exact mechanic v1’s PathPatcher used, expressed as an Intervention.

Add strength * unit(direction) to a site’s activation; an Intervention (section 2.6.2).

direction is the array-like (d_model,) steering direction (unit-normalized on use, so its magnitude is irrelevant and only its orientation matters). site names where to add it, in the residual stream or a sublayer output. strength is the signed displacement in unit-direction units, the dose. positions, if given, restricts the addition to those (padded) sequence columns; the default None steers every position, which is what a final-token readout needs and what the dose-response Observable uses.

Remove a direction (or a head) from an activation; an Intervention (section 2.6.2).

site names where the ablation acts. mode selects the removal:

• "directional" projects direction out of the residual (h -> h - (h.u) u); direction is required and mean_projection is ignored.
• "mean" replaces the along-direction coordinate with mean_projection instead of zeroing it (h -> h - (h.u) u + mean_projection * u); direction is required. The supplied scalar is the direction’s average projection over a reference dataset, computed upstream.
• "head" zeros the attention head named by site.head at Site(layer, "head_out", head); n_heads is required so the o_proj input can be split into per-head slices, and direction is ignored.

The direction, where used, is unit-normalized on use, so only its orientation matters.

Project a concept direction out of the reward head weight; an Intervention (section 2.6.2).

direction is the array-like (d_model,) concept direction u to remove, in the space the reward head reads (the head-input hidden state), unit-normalized on use. readout names which readout’s vector to edit ("reward" by default). strength is the coefficient alpha in w_r' = w_r - alpha (w_r . u) u: alpha = 1 removes the component entirely (the E17 edit), alpha between 0 and 1 attenuates it, alpha > 1 over-corrects past orthogonality.

Because the edit lives on the readout vector rather than on an activation, a compiled edit carries no mounts; it carries the edited vector in meta and is applied by :func:run_edited_scores.

A fitted LEACE affine eraser, persisted with its fit-data provenance (DESIGN 2.6, R8).

P is the d x d projection and mu the d fit-data mean, so the erasure is the affine map r(x) = x - P (x - mu). rank is the dimension of the erased concept subspace. fit_data_id and concept_id are the ids of the feature matrix and the concept the eraser was fit on; carrying them on the artifact is design rule R8, and it is what lets a certificate (certify.py) name the exact data an eraser was trained on. sites records the residual sites the eraser is intended for, informational metadata for the head-only and chosen-layers surfaces.

The matrices are kept in float64: the eraser’s defining properties (idempotence, exact linear erasure, minimal edit) are proven to ~1e-6 or tighter, and float32 storage would put that tolerance out of reach for no memory saving that matters at a single site.

Fit the LEACE eraser from captured features X and a concept Z (DESIGN 2.6).

X is (n, d) (features at a site). Z is the concept: a length-n vector of labels for a single binary or scalar concept, or an (n, k) matrix whose columns span a k-dimensional concept subspace (one-hot class indicators, or several concepts at once). The returned :class:Eraser removes all linear information about every direction in span(Cov(X, Z)).

Only second moments enter, so the fit is a closed form with no optimization loop: the feature covariance Sigma and the cross-covariance Sigma_XZ are formed, and :func:leace_matrix returns the projection. The fit-data and concept ids are stored on the eraser (R8); pass them from the capture provenance so the persisted eraser names its training data.

Mount a fitted :class:Eraser at one or more sites as an Intervention (DESIGN 2.6.1).

sites selects the surface: a single residual site for a targeted edit, several named residual sites for the chosen-layers surface, or every residual site the caller enumerates for the all-layers surface. Compilation resolves the eraser into one affine mount hook per site; the hooks install through the runtime’s single mounting path, exactly as patches and steers do, so an erased forward pass is measured by any Observable unchanged and the erasure fingerprint is carried into the intervened Evidence’s SubjectRef. The head-only surface is not a hook; it is :meth:Eraser.apply_to_readout applied to the reward direction.

This wraps the frozen :class:~reward_lens.interventions.base.Intervention protocol: the map is signal-independent (it is an affine function of the activation), so compile ignores the signal and the sites are architecture-resolved by the runner’s SiteMap at mount time, the same contract patch.py follows.

Certify an eraser by held-out probe recovery, as Evidence (DESIGN line 603).

The held-out features are erased, split into a probe-train and a probe-eval portion, and a fresh linear probe is trained and scored per concept column. The reported recovery_auc is the worst case across columns. If it is at most 0.5 + eps the erasure holds on data the eraser never saw, and the returned Evidence carries a :class:CalibrationRef, so gate 1 lifts it to CALIBRATED; otherwise the Evidence carries no calibration and stays EXPLORATORY. That is the mechanism by which a certificate lifts an eraser above EXPLORATORY, and by which a fake erasure is refused: the pass/fail is computed from the same held-out probe either way.

X_holdout must be disjoint from the eraser’s fit data for the certificate to mean anything; the caller is responsible for that split, and eraser.fit_data_id records what to exclude. The certificate is INVARIANT (a scalar recovery statistic, not a cross-signal geometric quantity), so no frame is required.

Compose several interventions into one that mounts them together in order.