AUTONOMY

n sensors fuse to a single engagement RECOMMENDATION. Byzantine fault tolerance requires n ≥ 3f+1 and a quorum of 2f+1 agreeing witnesses. With n=4 LIVE witnesses we tolerate f=1 Byzantine sensor before any recommendation. Toggle a sensor to Byzantine (corrupt value) or SAMPLE (abstains — does not witness). Drop to n=3 and the configuration goes INFEASIBLE. Khipu BFT unconditional = Conjecture 2 (never claimed proven).

Every PROPOSED action is projected onto the safe set by a Control-Barrier-Function quadratic program: minimise ½(u−u_nom)² s.t. the CBF constraint ∇h·(f+gu) ≥ −αh. An unsafe proposal is clamped to the boundary. The effector stays SIMULATED — this clamps a proposal, never drives anything.

Active inference computes Expected Free Energy for two policies, ACT (commit on the MAP class) and ASK (query a human). Policy posterior = softmax(−β·[G_act, G_ask]). Precision β is the VISIBLE human-oversight knob. Pick a confident vs uncertain belief and sweep β: the gate flips ACT↔ASK.

Replaces a bare "confidence X%" with a prediction set S: split-conformal qhat = Quantile(s, ⌈(n+1)(1−α)⌉/n); S = {y : 1−p(y|x) ≤ qhat}. Guarantee: true class ∈ S with ≥95% coverage (marginal, under exchangeability). LOCAL wrapper matching Dev B's predict_set/coverage_guarantee API — reconcile when RESULT_DEVB_AGENTIC.md lands.

Algebraic connectivity λ2 = 2nd-smallest eigenvalue of the graph Laplacian L=D−A. λ2→0 means the mesh is fracturing. We compute λ2 on the LIVE mesh topology and alert below threshold. Cut links to force a bottleneck and watch λ2 drop past the alert line.

After each reviewed C2 decision we store a reflection; on the next activation we prepend the accumulated reflections so the planner carries forward what a watch officer corrected. Log a reviewed decision, then fetch the preamble that would be prepended next time.