SFM Safe Flow Expansion: confirmed fixed recipe (CR −.100, Validity +.133, clearance +.014 @ M100 paired CIs) + pre-registered MPC-distillation follow-up

[DHLeexpress]

Recipe study (branch lineage: agent/claude-sfm-best-recipe-20260726, contained here)

Confirmed result (disjoint M100 bank ep0 330000, 700 CRN rollouts/method, temperature 1.0, no tilt/verifier/fallback; paired scenario-cluster 95% CIs vs r0):

	SR	CR	Validity	succ. clearance	succ. time
r0	.643	.353	.603	.108	8.76 s
selected	.723	.253	.736	.122	10.84 s
locked Kazuki (.3/.5)	.827	.173	.353	.168	4.17 s

ΔCR −.100 [−.157,−.043] · ΔValidity +.133 [+.114,+.153] · Δclearance +.014 [+.005,+.023] · Δtime +2.08 s [1.84,2.32].

Selected recipe (round-invariant, from exact r0 SHA 1b5179c9…): margin execution selector, original replay ∪ exact-certified deterministic recovery positives (per-row SOCP certificate audit), α=.01, 100 exposure epochs, lr 1e-4, ESS .5; selected round 1 by the predeclared M50 rule on a fresh bank (only eligible round; rounds 2–4 collapse to timeout — reported plainly, no monotonic-learning claim).

Mechanism (figures in /data3/research1/claude_sfm_best_recipe_f06e8dd/mechanism/): collisions are decided in a 2–4-step closing certifiability window at congested episode starts — flow support crashes to 1–2/16 at NVP onset while deterministic certified escapes still exist; a few steps deeper nothing certifies. Recovery data moves flow mass into that window (NVP −25%, both diagnosed collision lineages → closed-loop successes).

Integrity: predeclared disjoint banks, frozen selection rules before every read, iteration-1 null result (cost/hard) fully reported with its own 11-round M50 curve and paper plot, Kazuki locked, temperature 1.0 everywhere, 33-artifact SHA-256 manifest, DELIVERY_COMPLETE.json. Additive modules only — flags off ⇒ bitwise-identical replay; existing 43-test suite green.

MPC distillation follow-up (this branch's tip; pre-registered, COMPLETE — fail-closed at M10)

Imports the Codex privileged candidate-pool logic read-only from agent/sfm-adhoc-controller-overlay-20260726@0e0eca2 (nothing on that branch touched). Separate per-round D_MPC+ (never in D+/GP/acquisition): privileged exact-SFM look-ahead ∧ exact full-H10 SOCP, ranked by native SafeMPPI cost, distilled in dedicated post-round gradient blocks with fresh Gaussian CFM bases; per-block before/after audits; every pre/post checkpoint kept.

RESULT: the dedicated blocks do visibly change the raw policy — SOCP-positive rate at MPC contexts up in 15/16 blocks (up to .106→.482), target-recovery RMSE down in 16/16, present even when the block's CFM loss rose — but no swept dose passes the pre-registered liveness gate on the fixed M10 bank (0/32 cells; post-block round-1 SR .243–.600 vs gate .637), the no-distill control collapses identically by rounds 2–3, and the M50/M100 confirmations are therefore not triggered (MPC_STUDY_DELIVERY.json, mpc_distill/MPC_M10_SELECTION.json).

Synthesis across both studies: hard-context-only distillation (deterministic escapes or privileged MPC pool alike) trades global liveness for local certifiable support; the confirmed recipe wins by embedding certified targets in the full replay mixture (~5% mass). Composition, not target quality, is the binding constraint.

Study root: /data3/research1/claude_sfm_best_recipe_f06e8dd (all artifacts + manifests).

🤖 Generated with Claude Code

https://claude.ai/code/session_015CpH3uJa2LFTQ6m8xvXNXa

[DHLeexpress]

Follow-up study delivered on branch agent/claude-sfm-r1-stable-continuation-20260727 (descends from this PR's head): pre-registered stable-continuation search from the confirmed r1. Outcome: 0/16 round-1 candidates admissible under the hard gates (timeout binds at every dose; CR fell to .129 and Validity rose to .79 but liveness always paid), so the frozen winner degenerated to immutable r1 — declared before the final bank was read. Final M100 (ep0 410000, paired CIs): r1 vs locked Kazuki = Validity +.384 [+.354,+.412] ahead; CR +.089, clearance −.055, time +6.49 s behind → full Pareto dominance NOT achieved, reported metric-by-metric. Artifacts + 26-entry SHA manifest under /data3/research1/claude_sfm_best_recipe_f06e8dd/r1_continuation/.

[DHLeexpress]

Third follow-up delivered on agent/claude-sfm-unverified-teacher-run-20260728 (pinned tool @2a873327 used as-is + 3 additive commits): unverified-MPC-teacher continuation, pre-registered. Outcome: fail-closed at round 1 — the UNCHANGED ordinary e10 replay round from r1 already collapses liveness on the dev bank (SR .77→.44, timeout .04→.43, while CR .186→.129 and V .752→.836), the nine teacher doses (lr 1e-6..1e-5 × 1..4 ep) are epsilon on top, and the conflict probes show the teacher gradient is near-orthogonal to certified D+ (cos .03→−.07) while aligned with rejected D− (cos .38). 0/10 admissible ⇒ M50/M100 banks untouched, no Kazuki-win claim. Identical D_MPC across arms enforced by canonical content hash (torch.save bytes proven non-deterministic across equal payloads). Artifacts + manifest under /data3/research1/claude_sfm_best_recipe_f06e8dd/teacher_rounds/.

[DHLeexpress]

Fourth follow-up on agent/claude-sfm-corrected-privileged-distill-20260728 (frozen 736fa6c): corrected-replay privileged distillation study. Phase 1: the historical privileged deterministic SFM controller goes 70/70 on the current severe OOD (SR 1.000/CR 0.000, clearance .233, time 4.47 s; exact-SOCP window Validity .414, audit-only). Phase 3 fixes verified: original accumulated one-Adam-step-per-epoch replay (fail-closed W=2), exact L_T=ΣmᵢL_CFM(i) teacher weighting with a production-policy chunked-vs-direct gradient test (<1e-5), label-based baseline reader (prior r0-as-r1 read disclosed; no conclusion flips). Phase 4: the corrected teacher measurably moves r1 (drift up to 1.4e-3, RMSE .046) and the corrected ordinary one-step update is dramatically gentler than the old minibatch semantics — but the disjoint M50 confirms no arm beats r1 (best: corrected-ordinary V +.016/CR −.006, within noise; teacher M10 CR gains reverse). Gradient anti-alignment persists under the corrected objective: cos(teacher, D+)=−.07, cos(teacher, D−)=+.57 — even a perfect controller's successful executed windows point toward what the certificate rejects. Full Pareto tables, paper-contract plot, active-dodge PNG/MP4, 40+ artifact manifest under /data3/research1/claude_sfm_corrected_privileged_distill_736fa6c/.