Products / 03 — Explainability
A laboratory for explainable AI. Not one instrument — a host. Bring any model and any set of axes you want cognition explained in terms of; the lab projects onto that basis and tells you whether the basis is valid.
The idea
Every explanation is a projection onto a basis. "This loan was declined because of debt ratio and payment history" projects a decision onto two axes. The interesting question is never the projection — it is whether those axes were any good.
Two properties decide it. A basis must be irreducible: its axes must be independent, or your attribution smears across entangled dimensions and means nothing. And it must be exhaustive: the axes must span the space, or whatever matters falls silently outside the explanation.
Arc Lab tests both by measurement. That is what makes it a laboratory rather than a configuration screen.
The two verdicts
Both questions get answered by running the system, not by arguing about the taxonomy.
The lab embeds each axis definition and computes the full pairwise cosine matrix. The diagonal is meaningless; the off-diagonal is the finding. High off-diagonal similarity means two axes are measuring overlapping things — and any investment score split between them is arbitrary.
Verdicts are reported in σ above a per-model baseline, not against a hardcoded cosine cutoff. A fixed threshold would flip the verdict just by swapping the embedding model. Standard deviations do not.
Decompose an input into one answer per axis. Then discard the original and reconstruct it from the per-axis answers alone. Measure how much of the original survives.
What is lost in the round trip is precisely what your basis cannot see. It is a direct, falsifiable measurement of the blind spot — reported as fidelity renormalised against the same per-model baseline, so it means the same thing across models.
Why model-agnosticism is load-bearing
Early scoring used absolute cosine thresholds calibrated for one embedding model. Swapping models silently broke it. One model separated related from unrelated pairs by a wide, usable margin. Another compressed that margin to almost nothing — every score collapsed toward noise, and nothing crossed the fixed threshold at all.
The fix was not to mandate a model. It was to make scoring calibrate itself: embed a fixed unrelated corpus once per model, measure that model's own baseline mean and spread, and express every subsequent score relative to it.
The effect is that the same basis, measured with two different embedders, can land on opposite sides of a fixed cutoff while being the same basis. Expressed in σ above each model's own baseline, the verdict survives the swap. Agnosticism does not make models equal; it makes their verdicts comparable.
Shape of the output, with values omitted — the numbers depend entirely on the basis and the model you bring. The lab is designed to return "your basis is bad." That is the feature.
Basis-agnostic
Our default basis decomposes input into essence, form, frame, function, intent, relation and value. That is one configuration. The lab accepts any scheme from two to twelve axes — rename them, add them, remove them, rewrite their definitions. The mathematics is identical; only the basis the input is projected onto changes.
Essence · form · frame · function · intent · relation · value. General-purpose cognition, and the basis the lattice's six-face voxel is built around.
A hospital brings clinical-reasoning axes. A lender brings credit-risk axes. Explanations arrive in the vocabulary the regulator already audits against — not ours.
Crucially, the customer's own basis goes through the same two tests. If their taxonomy is entangled or leaky, they find out by measurement, before deployment.
Ante-hoc
SHAP, LIME and saliency maps are post-hoc: they observe a black box's inputs and outputs and fit a story that accounts for them. The story can be plausible and wrong, and there is no internal fact of the matter to check it against.
In Arc Lab the arc decomposition is the control flow. Ranked arc influence determines the tool chain the system executes and the memory it retrieves. There is no gap between the explanation and the mechanism, because they are the same object. In a regulated deployment, that distinction is the entire argument.
The lab includes an append-only prediction ledger. A prediction is registered with explicit kill criteria before the run; the outcome is recorded whether or not it was flattering; calibration is tracked across the whole history. Falsifiability is a build requirement, not a slogan.
Model-agnostic does not mean model-equal. The lab cannot manufacture signal a weak embedder never had — it can only report, honestly and on one scale, how much signal is there. We would rather say that than sell a number that flatters the setup.
Operator surfaces
| Cockpit | Product console. Input streams live through every pipeline stage, each with its own panel. The overview auto-detects failure signatures — a zero injection ratio, a collapsed set of arcs — instead of leaving them to be hand-diagnosed. |
|---|---|
| Audit panel | Standalone inspection surface that generates the exportable run document. Tunable depth (cognition tier) and width (lattice size). |
| Arc Lab panel | Define a basis, run the two validity tests, read the verdicts as heat by σ and relative round-trip fidelity. |
| Test bench | Append-only, pre-registered, falsifiable prediction ledger with kill criteria and calibration tracking. |
| Lattice viewer | Interactive 3D view of the memory substrate whose coordinates are read from the live backend. |