The FABLE target does not need to be reachable for diagnostic recovery
The delay is not caused by PID or phase lead. It comes from the FABLE authority law waiting for slow, gated b_osi evidence to satisfy an impossible upward threshold after the game returns to nominal.
a even when target_outer_ncf is above the attainable FABLE margin. Keep b_st, b_res, b_fast, and measured b_osi unchanged; add a narrowly armed restore_pending path to the authority supervisor.1/√2 is the configured software cap of the selected FABLE margin, not the mathematical maximum of the underlying chordal expression. Also, the exact gain-return time is 75.469238 s, so the measured slow-margin delay is 2.307617 s—not 2.78 s.Exact reason in FABLE terminology
b_st is stale, then unavailable
After the gain returns 2→1, the coherent/excited-bin branch b_st still contains the damaged slow window. Through 77.277 s it is approximately 0.272 while b_res = 0.707107, so b_slow = min(b_st,b_res) remains low. The three-slow-hop median b_osi therefore cannot change immediately.
The evidence gate then blocks valid_up
At 77.776855 s, b_st = NaN and b_res = 0.707107, allowing b_osi to reach the configured cap. From 78.401855 s, the evidence gate makes b_osi = NaN and valid_up = false while not g_armed. The saved no-evidence branch holds a rather than recovering it.
saved upward test: b_filt ≥ 0.80 + 0.10 = 0.90
FABLE cap: cfg.cap = 1/√2 = 0.707107
cfg.cap, and healthy evidence disappears before b_filt/above_n can complete. The only fast event in this record is trip_reason = 1 at 64.402 s; b_fast is not used by the ordinary upward path. regime_recal is also not a plant-repair signal—it only handles changed input support and can undo a recent fast trip inside trip_undo_s.What each FABLE branch must continue to do
| Signal | Existing FABLE role | Policy in the fix |
|---|---|---|
b_st | ST margin on coherent/excited bins; valid only when st_valid. | Unchanged. Preserve its calibration, rings, and measurement-only recovery role. |
b_res | Residual/quiet-bin margin: cap while disengaged, residual-growth margin while res_on. | Unchanged. It remains essential where coherent-bin evidence is weak. |
b_slow | min(b_st,b_res), including the existing all-NaN handling. | Unchanged. |
b_osi | Median of slow_hist across three distinct slow updates, capped by cfg.cap. | Never synthesize a healthy value. Show NaN / WARMING while evidence refreshes. |
b_fast | Short-memory ST margin; selected into b_osi only for trip_reason = 2. | Keep the fast CUSUM/trip path alive. After a diagnostic return, use publications only as permissive “no fast fault” checks. |
a_cont, a | Continuous and committed authority governed by trip, cooldown, evidence, filtering, deadband, and commit states. | Only this supervisor receives the additive diagnostic-return state. |
Exact record timeline and winning replay
The diagnostic gain is result.Results signal 21: 1 → 1.5 at 50.224121 s, 1.5 → 2 at 56.520508 s, then 2 → 1 at 75.469238 s.
Offline replay comparison
The replay reads the actual v7.3 MAT streams and changes only recovery logic. It is a fixed-record counterfactual, not a closed-loop plant simulation. Full metrics are in offline_replay_subagent_metrics.json.
| Candidate | Authority ≥ 0.90 after restore | Full authority | False pre-restore triggers | Verdict |
|---|---|---|---|---|
| Saved target-tracking law | Never | Never; final 0.20 | 0 | Deadlocked |
| Target clamp only | Does not complete before gate blanks | Never in saved no-evidence branch | 0 | Insufficient |
Cap-relative measured recovery, b ≥ 0.6871 | Would start before actual repair | Unsafe | 2 windows: 64.400–64.776 and 70.777–72.526 s | Reject |
| Aggressive NaN/blind recovery | 3.933 s | 3.933 s | 1: no-evidence interval 65.527–69.526 s while gain is abnormal | Reject |
| Production blind fallback, 5 s + 0.06/s | 19.683 s | 21.308 s | 0 in this replay | Safe fallback, too slow |
| Supervised 1.0 s ramp | 0.875 s | 1.000 s | 0 | Acceptable |
| Gain edge + 3 clean fast checks + 0.5 s ramp | 0.807617 s | 0.807617 s | 0 | Recommended |
| Supervised 0.5 s ramp without confirmation | 0.500 s | 0.500 s | 0 | Fastest bumpless option |
| Immediate supervised step | 0.125 s | 0.125 s | 0 | Reject: 5× authority jump |
Only changes needed
When diagnostic mode is enabled, arm
restore_pending. This is not regime_recal.b_fast publicationsNo
trip_reason; then begin a bounded ramp of a_cont/a.a = 1Existing trip and downward authority paths remain higher priority.
target_outer_ncf = 0.8 remains the requested/reported target. For authority decisions use b_ref = min(target_outer_ncf, cfg.cap - up_deadband - 0.02). With current up_deadband = 0.04, b_ref = 0.647107. This prevents an immediate re-shed after supervised restoration. The saved record used up_deadband = 0.10; that is why its upward test was 0.90.Add
diagnostic_restore_enable, diagnostic_gain and diagnostic_gain_nominal. Do not hardcode 1; the saved run used 1, while other model configurations may use another nominal value.After—and only after—the known gain-return edge, require three consecutive fast updates with no trip. Treat
b_fast as a permissive no-fast-fault check, not proof of healing: the replay found 12 pre-edge intervals that would satisfy the same fast-only condition.Preserve
b_st/b_res calibration, baselines, spectral rings, and the b_fast CUSUM. Clear only stale selected aggregation and authority transient memory needed for a clean commit—such as slow_hist, b_filt, below_n, above_n, cooldown/latch counters, and the old shed target. Keep raw b_osi = NaN/WARMING until fresh evidence exists.At the 0.125 s update cadence, use
a = min(a_init, a + 1.6·0.125). Starting from 0.2 gives four updates: 0.2 → 0.4 → 0.6 → 0.8 → 1.0. Confirmation plus ramp takes 0.807617 s in this record.trip_reason always wins.A quiet-band burst (
trip_reason = 1), fast FRF collapse (trip_reason = 2), or newly valid degradation cancels restore_pending within one update and executes the existing downward path.When
diagnostic_restore_enable = 0, the new branch is unreachable. Existing synthetic and physical-test behavior is therefore unchanged by construction.if trip_reason ~= 0 || fresh_degradation
restore_pending = false;
run_existing_drop_path();
elseif diagnostic_restore_enable && gain_returned_to_nominal
restore_pending = true;
clean_fast_count = 0;
elseif restore_pending && clean_fast_count < 3 && b_fast_published
clean_fast_count = clean_fast_count + 1;
elseif restore_pending && clean_fast_count >= 3
clear_stale_aggregation_and_authority_memory();
a_cont = min(a_init, a_cont + 1.6*dt_update);
a = commit_existing_way(a_cont);
restore_pending = a < a_init;
endWhy the diagnostic edge is additive, not a replacement
Autonomous cases still recover from measurements
Cases 7 and 8 have no diagnostic gain event. Their recovery must remain entirely b_st/prediction based; the new branch is unreachable and their traces remain unchanged.
A scalar cap is not proof of repair
b_osi reaches the cap in two pre-return windows, and the same three-clean-b_fast condition occurs 12 times before the gain edge. The known edge is what makes the confirmation safe in this diagnostic test.
heal_reason, confirm it for 2–3 steps, and abort on any existing trip. That design needs a new closed-loop replay because the required directional statistic was not logged in GDBqFpKb.Old implementation versus candidate: ten FABLE cases
The archived suite_runtime_sliced.mat baseline was replayed with an isolated wrapper. Because none of the ten cases contains the diagnostic gain-return event, restore_pending activates zero times. Authority, b_osi, and trip outputs are therefore sample-for-sample identical. Full machine-readable results: diagnostic_restore_validation.json.
| # | Canonical case | FABLE behavior being exercised | Old | Candidate difference |
|---|---|---|---|---|
| 1 | healthy broadband | b_st selected; b_res at cap; no fast trip | Pass | 0; pass |
| 2 | intensity step | b_st normalization; residual branch disengaged | Pass | 0; pass |
| 3 | broadband damage | b_fast type-2 trip, then b_st/st_binding | Pass | 0; pass |
| 4 | low-band damage | Intended b_res defense; archived run actually uses one type-2 fast trip | Pre-existing fail | 0; same fail |
| 5 | silent | Evidence gate: b_osi = NaN, authority hold | Pass | 0; pass |
| 6 | gradual softening | Slow b_st tracking without false fast trip | Pass | 0; pass |
| 7 | damage then repair | Type-2 shed followed by measurement-only b_st/prediction recovery | Pre-existing fail: 0.625 s versus 0.5 s limit | 0; same fail |
| 8 | starts damaged, improves | Pure b_st/prediction recovery; no diagnostic event | Pass | 0; pass |
| 9 | healthy regime change | regime_recal/gate prevents stale residual or fast effects | Pass | 0; pass |
| 10 | transient pulse | Bounded fast trip, residual release, b_st recovery | Pass | 0; pass |
|Δa| = 0 in every case, authority total variation is identical, and b_osi/trip streams are identical. The two failures already exist in the archived implementation and were not created or hidden by this change.b_st, b_res, and b_fast traces as NaN. Add these branch traces plus decision and g_armed; add a residual-only case that truly forces res_on; and run both canonical and production-configured suites. The archive predates current HEAD, so a fresh MATLAB-MCP baseline/candidate run remains mandatory.PID and phase-lead conclusion
The model does not retune PID or phase lead during authority adaptation. auth from Fable OSI + Auth multiplies the displacement command in Apply OSI gain before saturation/MTS. PID and pl_5k remain static.
Keep them unchanged for this fix
Changing P, I, D or phase lead cannot change the state-machine conditions that hold on NaN or require b_filt ≥ target + deadband. It would change tracking and estimator excitation in every case.
Robust redesign remains separate
Existing studies show phase-lead changes are not monotonic across damage cases. The saved-lead/full-authority envelope includes very poor worst-case margins, while the robust candidate required authority attenuation and a significant tracking tradeoff. That is not a recovery-logic patch.
Implementation validation gates
| Gate | Required result |
|---|---|
| GDB fixed-record replay | Gain return to full authority in ≤0.81 s with three fast checks; zero activation before 75.469238 s; no re-shed for 10 s unless a real trip/degradation is injected. |
| Trip during restore | Ramp cancels and downward action begins within one 0.125 s update. |
| Legacy ten-case replay | Verified: zero diagnostic activations; a, total variation, b_osi, and trip outputs unchanged in all cases. |
| Fresh current-HEAD suites | Still required through MATLAB MCP: canonical plus production configuration, with explicit b_st/b_res/b_fast/decision/g_armed outputs. |
| Runtime equivalence | Desktop and real-time implementations retain maximum absolute difference 0. |
| Bumpless command | Adjacent dc change remains inside the configured command slew/saturation limit and does not create a new dm transient. |
| Raw indicator semantics | After restoration, display WARMING/known nominal restored; never fabricate a measured 0.7071 value. |