# DRCT Black-Box Transformer Stress Test v6

A trained two-head causal transformer was hidden behind a logit-query oracle.

The deconstructor was allowed to query only token sequences and receive the
five next-token logits. It did not use model parameters.

Because the declared vocabulary is 5 and maximum context is 4, the
entire possible input domain contains:

    5 + 5^2 + 5^3 + 5^4 = 780

contexts.

All 780 were queried. Therefore the resulting lookup clone is an
exhaustive functional reconstruction on this finite domain, not a sampled
approximation.

The package also contains an exhaustive single-token intervention atlas showing
how changing every input value at every context position changes every output
logit.

Separately, the hidden teacher was transformed by permuting its attention heads
and compensating the attention output projection. The resulting parameter
tensors are different but the complete input/output function is algebraically
the same. This demonstrates why black-box behavior cannot uniquely identify raw
internal weights without quotienting parameter symmetries.

Results:
- task accuracy: 100.00%
- complete oracle queries: 780
- lookup clone maximum logit difference: 0.000e+00
- lookup clone argmax mismatches: 0
- equivalent model parameter L2 distance: 9.826068
- equivalent model max logit difference across full domain: 2.487e-14
- equivalent model argmax mismatches: 0