The
Certificate

The response certificate.

A certified instrument reads every output and stamps it — GROUNDED when the model knows, off the map when it’s reaching. Not a black-box score you have to trust: the model’s own state, the gauges, and the route, shown. Here is what a catch looks like.

an illustrative certificate · live reads run on honesty.tools · cold-deployed, certified ≥0.9× native

Supporting a model shouldn’t be a research project.

Porting an epistemic instrument to a new model is, today, bespoke work. The certificate turns "we support model X" into a registration pipeline: point us at a model, we fit the cross-mind translator on unlabeled concepts, deploy the rotated reading probes, and hand back a dual-null-validated certificate that the instrument runs at near-native fidelity on that model — with no labels in it.

Register. Rotate. Deploy. Certify.

A number you can underwrite.

A certificate is not a benchmark score. It is three gates — and fidelity without dead nulls is void.

A certificate that passes G1 but fails a null is void. Fidelity without a dead null is the classic interpretability mistake — a number with nothing behind it. We publish the gate values, not just the headline.

It works mechanically — with zero hand-tuning.

We ran the full pipeline on fresh strangers never used in any prior study, fitting R on unlabeled concepts alone, no per-pair tuning.

Honest detail: the strict three-gate certificate now passes on three strangers across three vendor families, zero hand-tuning each time. The 2026-06-12 runs are the strongest: a Mistral-lineage embedding fine-tune (cross-family, cross-vendor, cross-objective — peak found mechanically at depth 0.72; G2 shuffled-R 0.464 dead, G3 0.500 dead) and Microsoft’s phi-4 (peak at depth 0.90; ported 0.996 vs native 0.993; G2 0.194 — far below the 0.55 bar; the shuffled rotation carries nothing; G3 0.500). Measured concept-peak depths now span 0.21–0.90 across architectures — the one mechanical scan step handles all of it. For two very similar models (same family, close scale) G2 inflates to 0.732 — a space-similarity artifact of near-identical spaces, where G3 is the cleaner null — so the cross-family runs are the ones we headline. And the harder battery is now run: on the lexicality-controlled set — where fabricated and real names are matched on subword rarity, removing the rare-name shortcut — e5-mistral certifies cleanly (transfer 0.995, both nulls dead at 0.506 / 0.500), while phi-4 clears fidelity (0.990) but its shuffled-R null reads 0.580, just over the 0.55 bar, so by our locked criterion phi-4 is not certified on this stricter battery (n=66; plausibly small-sample — the larger battery is the clean re-test). We report the fail. Native fidelity saturates on both batteries, so the load-bearing evidence is always the dead nulls, not the transfer magnitude.

Coarse instruments. Open weights. Not a truth oracle.

The certificate covers the coarse reading instruments — the off-map / familiarity probe and the honesty/abstention reading axis. These are off-distribution and stance instruments, and they are the layer that ports cleanly up-to-rotation.

It requires residual access — open-weight, self-hosted models. It is not a truth oracle: it reads what a model knows versus when an input is off its map; on a confident, in-distribution fabrication the field reads at chance (AUROC 0.587), so that case is caught downstream by the claim-aware fact-check, not by the certified instrument. Closed APIs you can’t open are served by a separate output-logprob probe (its own certificate). The certificate states its own scope — a clean gate, not an oracle.

The map is copyable. The operation is not.

We’ll say it plainly: the shared concept geometry is reproducible from public data by a closed-form rotation — a competent team copies the underlying map in a week. The artifact is not what’s defensible. What’s hard to copy is the operation around it.

The whole bet rests on one falsifiable claim: that registering a new model is mechanical — no per-pair hand-tuning. If it weren’t, there is no flywheel; every model would be a research project. So we ran it on three strangers across three vendor families — including a contrastive embedding fine-tune and a model whose concept peak sits at depth 0.90 where every prior model peaked 0.2–0.72 — and all three cleared every gate with zero tuning, in minutes each — and on the stricter lexicality-controlled battery, one of the two fresh strangers certifies cleanly and the other posts a borderline null we report rather than hide. That is the defensible core, and it is now evidenced rather than asserted.