Compound Attack Simulations

Constant high scaffolding, slow accuracy drift, a forced unassisted session every eighth session, and occasional rushed approvals, all at once.

PurposeSee whether the three hardening measures (absolute floor, forced unassisted sessions, verifier gate) hold together when several attacks run at the same time.

StatusCannot fail — each script prints its success message on a condition the setup guarantees.

TestsTwo scripts: a 60-session run (drift 0.004 per session) and a 50-session “3-vector” run (drift 0.0035 per session).

Touches NCFCA codeNo. Self-contained numpy simulations.

AttributionCode AI-generated (tool not recorded). Run in the lab; results unchanged.

ReviewedBy no outside party. The reading below was done when this page was built, directly from the code and CSV.

Sixty-session run

Observed accuracy falls from 0.865 to 0.643. The running-maximum floor sits at 0.884, a noise draw in session 4, and holds the effective baseline at 0.813 from session 14 to session 60 while the rolling mean falls to 0.623.
The script asks “does the forcing function create a false peak that poisons the floor?” It never checks. Forced sessions lower accuracy by 0.15, so they cannot create a peak, and the 0.884 peak comes before the first forced session.
“MITIGATION EFFECTIVE: Absolute Floor protected against false peak poisoning” prints when the floor exceeds final accuracy by 0.15. Any decline of that size triggers it, so the message follows from the injected drift.
“True mastery” reaches 1.0 at session 33, when observed accuracy is 0.72 and falling. Mastery is a formula in the scaffold level and does not depend on accuracy.

Fifty-session “3-vector” run

The effective baseline is 0.814 from session 14 to the end while accuracy falls from 0.865 to 0.659.
The script prints “MITIGATIONS EFFECTIVE” when the baseline is above 0.78. The floor starts at 0.85, so the baseline can never fall below 0.92 × 0.85 = 0.782. The message cannot fail to print.
Mastery equals 0.02 × session plus a small term, so it reaches 1.0 at session 50 by the session counter alone. The five bad approvals move it by under 0.01 in the sessions where they occur and do not change the ending.
The verifier approval counter is never used, and no approval-rate limit is implemented, so the verifier-pressure vector is not tested.

What it shows

A running-maximum floor holds the baseline up while accuracy falls. That is what it was built to do, and it also shows its risk: the baseline stays at a value set by one early noise peak.

What it does not show

That the three measures work together. No outcome in either script depends on them: the success messages are fixed by construction and mastery is independent of accuracy.
That forced unassisted sessions help. They appear in the mastery formula as a bonus and nowhere else.
Anything about the four-channel architecture.

Download (unchanged from the lab’s run): compound_multi_vector_attack_test.py · compound_attack_results.csv · compound_3vector_attack_test.py · compound_3vector_results.csv
The scripts write to a path from the lab’s own machine (/home/workdir/artifacts/); change it to run them elsewhere. Each uses seed 42, so a rerun reproduces the CSV.