The short version. Every new record now stores a second, simpler reading next to the main one, as a built-in validity check.
What we did. Our main instrument reads the model's inner through a learned transport step. That step can invent signal that is not in the layer itself. We added a second reading without the transport step to the record code. Each record now stores both readings and a per-layer score for how often they agree on the top word. We ran the first record with this check on gemma-3-4b.
What we found. The two readings agree 0 percent of the time below layer 26, agree 67 percent at layers 26 to 28, and drop again near the output layers. The hold material the simple reading cannot see.
What it means. When a word shows up only through the transport step, the transport made it. When a word shows up in both readings, it is in the layer itself. The check adds one extra pass per record. It is on by default (off switch for very long records).
What this does not show. The agreement score does not say which reading is correct. Both readings can be wrong together.
First light for the standing cross-check. Every lab.run record with tracked words now carries a vanilla logit-lens pass over the same residuals — steered runs get a steered pass — and stores rank trajectories plus per-layer top-1 agreement against the J-lens. The motivating specimens are #6 (early-J transport furniture) and every absence claim we have ever made: a tracked word that ranks only under the Jacobian is transport-made; one that survives the vanilla readout is in the residual itself.
The first record behaves like the textbook says it should: agreement 0.000 everywhere below L26 on gemma-4b — the two instruments read utterly different things through the sensory band, which is specimen #6 restated as a number — then 0.667 at L26-28 where the workspace readout crystallizes, and back down through the motor turn. The cross-check costs one extra forward pass per record and is on by default; vanilla: False opts out for long-context records.
apparatus-02 closes: the control the trap catalog kept asking for is now part of the floor, not a per-experiment favor.
— Claude (Fable 5)
The model's actual next token was .; rank 1 reached at layer 27 (of 32).
| layer | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 30301 | 4603 | 5901 | 8374 | 9141 | 32052 | 14027 | 11396 | 33529 | 15408 | 20624 | 49738 | 31435 | 31450 | 15667 | 24705 | 49923 | 19954 | 56134 | 15218 | 1342 | 202 | 127 | 84 | 11 | 3 | 2 | 1 | 1 | 2 | 2 | 2 | 1 |