The battery

You have a reward model and a preference pair. What can you actually measure about the decision?

Eleven things, as it turns out, and they all start from the same fact. The reward is a linear readout of the final hidden state, $r = w_r^\top h + b$, so the whole opinion of the model lives in one direction $w_r$. Every instrument on this page is a different way of projecting the computation onto that direction and reading off what you find. The residual stream is a running sum, so the projection decomposes cleanly: by depth, by component, by concept, by objective. That is the battery.

One direction carries the entire score. A hidden state is a point in a high-dimensional space, but the reward only reads its shadow along $w_r$. Two completions differ in reward exactly by how far apart their shadows fall. Every instrument below measures that shadow from a different angle.

Because $w_r$ is fixed by the reward head and known exactly, none of this is probing or guessing. The projection is arithmetic. What varies between instruments is which slice of the computation they project, and, more importantly, what kind of claim the result licenses.

Three chips, read left to right

Every instrument page opens with three chips. They tell you, before you read a word of prose, what the number means and how far you can carry it.

• Tier is the kind of claim. Observational instruments read the computation as it already ran. Causal instruments intervene, change one thing, and measure what moved. Vulnerability instruments score the model against a named failure mode. The distinction is not decorative. On real models, the observational and causal instruments have ranked the same components in nearly opposite order, and knowing which tier you are holding is what keeps you from reading a correlation as a mechanism.
• Gauge is whether the number survives a change of coordinates. Invariant quantities (a depth fraction, an effect size) mean the same thing in any basis, so you can compare them across models freely. Raw only quantities are coordinate-dependent and honest about it: a raw cosine between two models’ features is a number about your axes, not about the models. Needs a frame quantities can be compared across models, but only once you express both in a shared frame, or the library refuses. The chip tells you which rung you are on so you never accidentally compare two raw numbers that were never in the same coordinates.
• Works on is the substrate the instrument attaches to: bare scores, activations plus the linear readout, an SAE, a multi-objective head. It is the capability the signal must declare. Ask an instrument for something the signal cannot give and it raises rather than returns a fabricated number.

The tier key

The eleven

The eight observational instruments read; the three causal ones intervene. Every one returns a gated Evidence object, not a bare float, and every one on this page currently reports at trust level EXPLORATORY, because no calibration provider is wired in this release. That is a statement about the epistemics, not the arithmetic: the projection is exact, but until a scorecard certifies the instrument against organisms with known ground truth, the number is exploratory by construction.

Instrument	What it answers	Tier	Gauge	Works on
Reward lens and crystallization	Where across depth does the margin form?	Observational	invariant	activations + linear readout
Component attribution	Which component wrote the score?	Observational	invariant	activations + linear readout
Patch grid	Which component causes the score?	Causal	invariant	activations + linear readout
Path effects	Which sender-to-receiver path carries it?	Causal	invariant	activations + linear readout
Concept dose-response	Does pushing a concept move the reward?	Causal	raw only	activations + linear readout
Bias battery	Which surface feature is the reward biased toward?	Observational	invariant	scores
Prompt SNR	How much signal versus noise per axis?	Observational	invariant	scores
Conflict matrix	Which reward terms fight each other?	Observational	raw only	activations + linear readout
Circuit overlap	Do two models use the same circuit?	Observational	invariant	two models, activations + linear readout
Feature-reward alignment	Which SAE features drive the reward?	Observational	raw only	an SAE + linear readout
Multi-objective geometry	How do a head’s objectives relate?	Observational	raw only	a multi-objective readout

How to run any of them

The call shape is the same for all eleven. Build a signal, choose a data view, hand both to the runner. The runner checks the capability, applies the gauge gate, and returns Evidence.

from reward_lens.signals import from_tiny
from reward_lens.data.builtin.diagnostic_v3 import load_diagnostic_v3
from reward_lens.data.schema import DataView
from reward_lens.measure import base as mb
from reward_lens.measure.battery import DirectLinearAttribution

signal = from_tiny(seed=0)                                        # real LlamaForSequenceClassification, CPU, no download
view = DataView(list(load_diagnostic_v3()["helpfulness"].items)[:5])
ev = mb.run(DirectLinearAttribution(), mb.Context(signal=signal, view=view))
print(ev.observable, "|", ev.trust, "|", ev.gauge)
# DirectLinearAttribution | EXPLORATORY | invariant

from_tiny builds a genuine two-layer reward model on CPU in under a minute, which is enough to exercise every observational instrument end to end. It is not enough to reproduce the 8B results, because a two-layer model has no late-layer structure to find. Throughout the battery, the tiny runs are there so you can see the machinery move; the Skywork and ArmoRM numbers are measured results from committed artifacts, and the call that produces them is marked as needing a GPU.

Start with the reward lens. It is the instrument that tells the others where to look.