Adapter Discipline¶

Status: Proven by gates (Gate 0 verifies non-interference).

The patched forward is a verbatim copy of upstream Mistral attention. Three marked insertion points capture tensors; nothing else changes. Any modification to the upstream code requires Gate 0 re-verification.

This is not conservative engineering --- it is the non-interference invariant. If the observer changes the system, the measurement is meaningless.

• attn_implementation="eager" required. Non-negotiable.
• Transformers ~=4.57 pinned. Forward-replacement adapter is version-coupled.
• Model revision pinned in gate fixtures.

The three insertion points¶

The patched forward in core/adapter.py is a verbatim copy of MistralAttention.forward from transformers 4.57.x. The self reference is rebound to module via closure; no other changes to upstream code.

Insertion 1: Tier A capture (post-RoPE Q/K/V)¶

Located immediately after apply_rotary_pos_emb. The capture_fn receives live tensors --- not cloned. The callback (InstrumentManager.capture_fn) is responsible for detach().clone().float() before any computation. This is the primary capture point: post-RoPE Q/K/V at every layer, every generation step.

Insertion 3: Pre-o_proj diagnostic¶

Located after eager_attention_forward returns but before reshape and o_proj. Captures native attention output at the newest token position \( [B, 1, H, D] \) before head reshape and output projection. Only active when parity_fn is not None (Gate 1 mode). Exists solely for failure localization.

Note: insertion 3 appears before insertion 2 in the code flow. The numbering reflects development order, not execution order.

Insertion 2: Parity capture¶

Located after o_proj. Passes post-RoPE Q/K/V, native post-o_proj output, a reference to o_proj, and the pre-o_proj diagnostic tensor to the parity validation callback. Hard-fails if past_key_values is not None (cache-on parity not validated). Only active during Gate 1 testing.

Why eager-only¶

The adapter patches MistralAttention.forward directly and calls eager_attention_forward. SDPA and Flash Attention use fused CUDA kernels that do not expose the intermediate attention output tensor needed for insertion 3. The forward control flow also differs between implementations --- the patched eager forward would produce incorrect results with a different attention backend.

The adapter hard-fails at runtime if module.config._attn_implementation != "eager". This check runs inside every forward call, not just at install time, preventing silent breakage from config mutation.

Version coupling¶

The patched forward is copied from transformers >=4.57.0, <4.58.0. A runtime version guard runs at install() time and raises if the installed transformers version is outside this range. This coupling is intentional --- the forward method's internal structure (argument order, tensor shapes, existence of eager_attention_forward) is version-specific. Any upstream change requires re-copying the forward, re-marking insertions, and re-running Gate 0.

Capture tiers¶

Mistral is Tier A: the instrument sees the same tensors the model uses, intercepted inside the forward pass.