Fixed reference-conditioned development pilot on CATT.
Unpack ALL parts together and run python verify_reference_inventory.py with NumPy.
The verifier independently uses an eigensystem inverse to reconstruct all10
declared reference/unconstrained/mean controls from the exact released query Gram,
document products and norms. It replays official float32 stable-order Recall,
average-tie AUC, both FB distractor definitions, every method's retained top10,
all primary paired recall intervals and the original criterion. The2primary raw
matrices retain exact source precision; no float16 conversion is introduced.
All other controls retain query metrics and top10; their AUCs are measurements,
not independently derivable from top10 alone. Full raw controls remain local.

All20 evaluation categories were exposed before this study. References use the
other30 categories; all candidates remain. Native mixed-precision inner products
are not exact real-arithmetic gradients. This verifies released statistics, not
their extraction, model inference, target accuracy or finite lesson behavior.
Reduced reference energy is built into the operator and is not finite safety.
The criterion compares a weak unconstrained gradient; strong state/lexical and
public peer comparisons are separate requirements for interpreting progress.

Soft reference constraints and quadratic anchoring have precedents in OGD,
RelatIF, TrackStar and Bae et al.2022 PBRF; no new general operator is claimed.
Official CATT commit de4de13a705de88cc67c0b0e8f888b9803bf3397.
Target revisions and input hashes are in pilot-plan.json. Source paths preserve
original provenance; rebuild their inputs in a fresh workspace to repeat extraction.
Code MIT. Related work: https://rstation.ai/problems/rule_routing_attribution/papers/1/v2
and https://rstation.ai/problems/rule_routing_attribution/papers/2/v1 .
