Research note

Where the variables actually couple

Night summary after rerunning the fixed branch: the clean benchmark split is now visible. One lane is the current harnessed detector. The other lane is the older trace simulator that is still useful for watching state motion through L1-L14.

main updated after PR #649 91/91 blocked 0/15 false positives 108/108 system checks passed

Benchmark split

Lane	File	Use	Do not use for
Harnessed detector	`scripts/benchmark/scbe_vs_industry.py`	Head-to-head protection claims, clean vs attack scoring, false-positive reporting.	Raw layer introspection.
Legacy trace simulator	`scripts/benchmark/scbe_vs_baseline.py`	Watching variable motion across L1-L14, seeing where energy spikes and decisions form.	Industry ranking or detector-vs-detector claims.
Formal invariant suite	`tests/test_full_system_verification.py`	Math axioms, tokenizer, PHDM, MMCCL, RWP, game and consistency checks.	Adversarial protection claims.

Main coupling path

L3

Weighted tongue lift

This is the first real amplification point. Small input differences stop looking like simple text metrics and become a weighted state. In the sample trace, energy jumped from about 1.01 to about 9.19 here.

L4-L5

Poincare clamp plus hyperbolic distance

The state is compressed back inside the ball and turned into a usable divergence scalar d*. This is the geometry handoff that the later layers use.

L6-L7

Breathing plus phase

This is where structured deviation starts to become visible without pure norm explosion. In the legacy sample trace, L7 was the first major spike at about 23.29 energy.

L8-L11

Coherence, spin and temporal aggregation

These layers decide whether the deviation is coherent, directional and persistent. This is where gradual attacks start to separate from static text.

L12

Harmonic wall

d* becomes governance cost. Important, but not sufficient by itself. In the sample direct-override trace, the cost was only about 2.48, which was not enough alone to hard-block.

L13

Decision fusion

This is the real interaction zone: geometry, lexical evidence, spin and session state combine here. The current fixed gate is winning because the decision is made from the constellation, not from one scalar.

What the fixed branch actually says

The session-bleed regression is gone in the harnessed benchmark lane.
The full adversarial harness is back at 91/91 detected with 0/15 clean false positives.
The old trace simulator is still useful, but now it is explicitly marked as a trace lane instead of a claim lane.
The main decision path is L3 → L7 → L12 → L13, with lexical and session evidence joining the geometric variables before the final block/allow outcome.

Next morning handoff

The next serious upgrade is not more threshold tuning. It is replacing the old front-door projection with the real tongue reader or learned tongue head, while preserving the proven L7-L13 fusion path.