EdgeStack research lab

Cites H-E153.v1 (ADOPT_CANDIDATE, sortino +0.1448/+0.2460, dd_delta train +0.0155 valid -0.0415) and the three H-E149 REJECT_DD_CONSTRAINT verdicts at train dd +0.0241: train drawdown, not Sortino, is what now decides this family's proposals, and the Sortino surface is saturated at valid ~+0.246 across five distinct adopted configs. H-E016|max_drawdown calibration is tight (bias train -0.0028, valid +0.0020, mae 0.0053/0.0122), so the predictions above are bias-corrected downward on the train side by ~0.003. Provenance note as required by protocol rule 5: the objective switch to max_drawdown is data-motivated - it follows directly from the observed dd_delta trajectory in the recent verdicts, not from inspecting drawdown diagnostics of an already-run variant of this configuration, which has never been tested under any objective. Family FDR is 0.160 against a 0.25 ceiling with 299 adoptions over 350 trials; this registration deliberately asks a question whose modal answer is negative rather than adding another Sortino-surface point to an already heavily mined grid. midpoint of the dose; if train dd is sleeve-driven this is where it should cross into positive territory

Cites H-E153.v1 (ADOPT_CANDIDATE, sortino +0.1448/+0.2460, dd_delta train +0.0155 valid -0.0415) and the three H-E149 REJECT_DD_CONSTRAINT verdicts at train dd +0.0241: train drawdown, not Sortino, is what now decides this family's proposals, and the Sortino surface is saturated at valid ~+0.246 across five distinct adopted configs. H-E016|max_drawdown calibration is tight (bias train -0.0028, valid +0.0020, mae 0.0053/0.0122), so the predictions above are bias-corrected downward on the train side by ~0.003. Provenance note as required by protocol rule 5: the objective switch to max_drawdown is data-motivated - it follows directly from the observed dd_delta trajectory in the recent verdicts, not from inspecting drawdown diagnostics of an already-run variant of this configuration, which has never been tested under any objective. Family FDR is 0.160 against a 0.25 ceiling with 299 adoptions over 350 trials; this registration deliberately asks a question whose modal answer is negative rather than adding another Sortino-surface point to an already heavily mined grid. mild tightening from the adopted 0.8; expect valid dd gain to persist and train to stay roughly flat vs baseline, i.e. below the 2pp floor

Parent H-E153.v1 (cap 1.5, vol_scale_hi 0.8) is the family's best adjudicated point: +0.1448 train / +0.2460 valid, dd_delta +0.0155 train, ADOPT_CANDIDATE at iteration 347. What it taught is that the marginal Sortino now comes from letting total sleeve exposure run, while the vol scaler pays the drawdown bill - H-E154.v1 removed the scaler at that cap and dd_delta jumped to +0.0309, an immediate REJECT_DD_CONSTRAINT. That fixes the open question: how much further can the cap go before the scaler stops covering it. I predict SATURATION, not gain: the cap binds on progressively fewer sessions, so all three doses land within noise (<=0.005) of the parent on both windows, and the DD cost grows faster than the Sortino, with the 3.0 dose most likely at or through the +0.02 ceiling. Predictions are corrected by this family's measured calibration bias (-0.007 train / -0.0067 valid over n=327), i.e. shaded down from the parent's realized numbers. A flat curve is the useful answer here: it would retire max_total_sleeve as a lever and say plainly that the last 150 trials have been re-measuring one compound effect - relevant with family_trials at 350, 299 adoptions and the null-calibrated FDR estimate already at 0.16 against a 0.25 ceiling. edgestack_live_c001.py is integrity-ok (sha ee13e3df5ab21dbb) and the parent verdict is reproducible=true. first step past the tested ceiling; expect the cap to bind on only a handful of extra sessions

H-E153.v1 is the family's best adjudicated config (+0.1448 train / +0.2460 valid, ADOPT_CANDIDATE, iteration 347) but it carries ten simultaneously-set flags and I have never established which of them matter. The digest gives direct evidence that the config is over-specified: H-E150.v3, H-E152.v1 and H-E152.v2 return byte-identical deltas (+0.1232/+0.2468) while differing in sleeve_trail_pct, and H-E150.v1 (max_total_sleeve 1.05) matches H-E150.v3 (1.5) on train, so both the sleeve trail and the sleeve cap above ~1.05 are inert at this point in the space. Meanwhile the family's null calibration is tightening - fdr_estimate 0.159 against fdr_q 0.25 with 297 adoptions in 347 trials - so each further dose step on an inert knob costs FDR headroom and buys nothing. Leave-one-out is the instrument the rule audits used and the reason I instrumented this file; applying it to my own winner should either confirm the components or let me shrink the config, which is the honest way to lower the overfitting risk that the rising trial count implies. Predictions are set below the naive 'no change' level and corrected by the family's measured sortino bias (-0.0073 train / -0.0069 valid, i.e. I have been mildly over-optimistic), with the tier_mode+vol_med_lo removal predicted as the largest loss since tier_mode appears in essentially every adopting config in the family. drop tier_mode together with its threshold vol_med_lo=2.2 (the threshold is meaningless without the tiering). This is the pair I expect to be load-bearing - the biggest predicted loss of the four.

H-E153.v1 is the family's best adjudicated config (+0.1448 train / +0.2460 valid, ADOPT_CANDIDATE, iteration 347) but it carries ten simultaneously-set flags and I have never established which of them matter. The digest gives direct evidence that the config is over-specified: H-E150.v3, H-E152.v1 and H-E152.v2 return byte-identical deltas (+0.1232/+0.2468) while differing in sleeve_trail_pct, and H-E150.v1 (max_total_sleeve 1.05) matches H-E150.v3 (1.5) on train, so both the sleeve trail and the sleeve cap above ~1.05 are inert at this point in the space. Meanwhile the family's null calibration is tightening - fdr_estimate 0.159 against fdr_q 0.25 with 297 adoptions in 347 trials - so each further dose step on an inert knob costs FDR headroom and buys nothing. Leave-one-out is the instrument the rule audits used and the reason I instrumented this file; applying it to my own winner should either confirm the components or let me shrink the config, which is the honest way to lower the overfitting risk that the rising trial count implies. Predictions are set below the naive 'no change' level and corrected by the family's measured sortino bias (-0.0073 train / -0.0069 valid, i.e. I have been mildly over-optimistic), with the tier_mode+vol_med_lo removal predicted as the largest loss since tier_mode appears in essentially every adopting config in the family. drop no_restack. With max_total_sleeve at 1.5 the cap is not binding, so the restack ban is the only thing left limiting same-name sleeve stacking; if it is inert here the flag should be retired from the config.

H-E153.v1 is the family's best adjudicated config (+0.1448 train / +0.2460 valid, ADOPT_CANDIDATE, iteration 347) but it carries ten simultaneously-set flags and I have never established which of them matter. The digest gives direct evidence that the config is over-specified: H-E150.v3, H-E152.v1 and H-E152.v2 return byte-identical deltas (+0.1232/+0.2468) while differing in sleeve_trail_pct, and H-E150.v1 (max_total_sleeve 1.05) matches H-E150.v3 (1.5) on train, so both the sleeve trail and the sleeve cap above ~1.05 are inert at this point in the space. Meanwhile the family's null calibration is tightening - fdr_estimate 0.159 against fdr_q 0.25 with 297 adoptions in 347 trials - so each further dose step on an inert knob costs FDR headroom and buys nothing. Leave-one-out is the instrument the rule audits used and the reason I instrumented this file; applying it to my own winner should either confirm the components or let me shrink the config, which is the honest way to lower the overfitting risk that the rising trial count implies. Predictions are set below the naive 'no change' level and corrected by the family's measured sortino bias (-0.0073 train / -0.0069 valid, i.e. I have been mildly over-optimistic), with the tier_mode+vol_med_lo removal predicted as the largest loss since tier_mode appears in essentially every adopting config in the family. drop sleeve_batch (back to default). Adopted at H-E096.v2 and carried ever since without being re-tested at this operating point; expect a small real loss.

H-E153.v1 is the family's best adjudicated config (+0.1448 train / +0.2460 valid, ADOPT_CANDIDATE, iteration 347) but it carries ten simultaneously-set flags and I have never established which of them matter. The digest gives direct evidence that the config is over-specified: H-E150.v3, H-E152.v1 and H-E152.v2 return byte-identical deltas (+0.1232/+0.2468) while differing in sleeve_trail_pct, and H-E150.v1 (max_total_sleeve 1.05) matches H-E150.v3 (1.5) on train, so both the sleeve trail and the sleeve cap above ~1.05 are inert at this point in the space. Meanwhile the family's null calibration is tightening - fdr_estimate 0.159 against fdr_q 0.25 with 297 adoptions in 347 trials - so each further dose step on an inert knob costs FDR headroom and buys nothing. Leave-one-out is the instrument the rule audits used and the reason I instrumented this file; applying it to my own winner should either confirm the components or let me shrink the config, which is the honest way to lower the overfitting risk that the rising trial count implies. Predictions are set below the naive 'no change' level and corrected by the family's measured sortino bias (-0.0073 train / -0.0069 valid, i.e. I have been mildly over-optimistic), with the tier_mode+vol_med_lo removal predicted as the largest loss since tier_mode appears in essentially every adopting config in the family. drop the vol-scaling pair (vol_scale_mode/vol_scale_hi). The 0.7->0.8 step was worth +0.022 train and -0.001 valid, so scaling looks train-side only; removing it entirely should cost ~0.04 train and ~0.035 valid if it is real infrastructure, and nothing if it is not.

Honest framing first: this family stands at 345 trials with 295 adoptions and an FDR estimate of 0.159 against the 0.25 ceiling (H-E152.v2 null_calibration), so the marginal value here is not another adoption — it is a clean answer about where the drawdown constraint actually binds. The evidence is unusually sharp for that. Across the recent block, train dd_delta takes exactly two values: +0.0241 whenever vol_scale_hi is 0.9 (H-E146.v3, H-E149.v1/v2/v3, all REJECT_DD_CONSTRAINT despite train +0.155..+0.165 and valid +0.176..+0.225 — the largest Sortino deltas in the block), and +0.0069 whenever it is 0.7 (H-E150.v1/v3, H-E152.v1/v2, all ADOPT). Every one of those was run at max_total_sleeve <= 1.05. H-E150.v3 established that raising the cap to 1.5 costs nothing on dd at vol_scale_hi 0.7, and H-E152.v1/v2 returned identical metrics (0.1232/0.2468) at trail 10.0 and 6.0, so the trail is not binding there and the cap is the only live change. The cross of the two axes is untested. Two outcomes are both informative: if dd at sleeve 1.5 tracks the +0.0241 signature, the drawdown is a property of the vol scaler alone and the 0.9 dose is permanently closed under this objective, which retires an axis the family keeps re-probing; if it lands under 0.02, the sleeve cap is a genuine dd absorber and the best Sortino configuration in the record becomes reachable. I have corrected all three predictions down by the family's measured calibration bias (-0.0073 train, -0.0070 valid over n=322). I predict v1 ADOPTs and v2/v3 are blocked on drawdown — a bold call, since v2's Sortino would otherwise be the family's best. the dose that was blocked at sleeve 0.9 (H-E146.v3, dd +0.0241). I expect it blocked again at sleeve 1.5 — the dd is coming from the vol scaler, not the cap.

Honest framing first: this family stands at 345 trials with 295 adoptions and an FDR estimate of 0.159 against the 0.25 ceiling (H-E152.v2 null_calibration), so the marginal value here is not another adoption — it is a clean answer about where the drawdown constraint actually binds. The evidence is unusually sharp for that. Across the recent block, train dd_delta takes exactly two values: +0.0241 whenever vol_scale_hi is 0.9 (H-E146.v3, H-E149.v1/v2/v3, all REJECT_DD_CONSTRAINT despite train +0.155..+0.165 and valid +0.176..+0.225 — the largest Sortino deltas in the block), and +0.0069 whenever it is 0.7 (H-E150.v1/v3, H-E152.v1/v2, all ADOPT). Every one of those was run at max_total_sleeve <= 1.05. H-E150.v3 established that raising the cap to 1.5 costs nothing on dd at vol_scale_hi 0.7, and H-E152.v1/v2 returned identical metrics (0.1232/0.2468) at trail 10.0 and 6.0, so the trail is not binding there and the cap is the only live change. The cross of the two axes is untested. Two outcomes are both informative: if dd at sleeve 1.5 tracks the +0.0241 signature, the drawdown is a property of the vol scaler alone and the 0.9 dose is permanently closed under this objective, which retires an axis the family keeps re-probing; if it lands under 0.02, the sleeve cap is a genuine dd absorber and the best Sortino configuration in the record becomes reachable. I have corrected all three predictions down by the family's measured calibration bias (-0.0073 train, -0.0070 valid over n=322). I predict v1 ADOPTs and v2/v3 are blocked on drawdown — a bold call, since v2's Sortino would otherwise be the family's best. midpoint dose. If the sleeve cap absorbs dd, this lands near 0.012-0.015 train dd and ADOPTs with a higher train delta than H-E150.v3.

H-E016 is at 343 trials with fdr_estimate 0.160 against a 0.25 ceiling, so an arm that merely re-ADOPTs near the frontier buys little; what is worth spending trials on is a dimension whose measured dose ordering was established on a base that has since moved. sleeve_trail_pct was doses at 3/6/10 only under max_total_sleeve 0.9 and vol_scale_hi 0.8 (H-E147.v1 +0.1072/+0.1707 dd +0.0191; H-E147.v3 +0.1448/+0.2287 dd +0.0155). The frontier has since moved to a loose cap (H-E150.v1 at 1.05 and H-E150.v3 at 1.5 both ADOPT, train identical at +0.1232, valid +0.2389 and +0.2468 - the cap is non-binding in train and rarely binding in valid) and to the gentler vol_scale_hi 0.7 (H-E146.v1, train dd_delta +0.0069 vs +0.0241 at 0.9). Those two changes leave more sleeve capital outstanding for longer, which is precisely the state a per-sleeve adverse-move exit is supposed to govern, so the earlier ordering should be re-measured rather than assumed. Predictions carry the family's measured bias correction (bias_train -0.0073, bias_valid -0.0071 over n=320: I have been mildly optimistic, so each arm is shaded down from the naive parent-plus-effect estimate). The tight arm is the informative one: I am predicting a large valid loss, and a flat result there would falsify the trail-dose story and reassign the H-E147 spread to the cap. tight dose. On the 0.9/0.8 base this was the worst arm (valid +0.1707); a bold call that the monotone ordering holds and tight trailing keeps cutting sleeves that recover - if valid instead HOLDS near 0.24 here, the earlier curve was a max_total_sleeve interaction, not a trail effect.

H-E016 is at 343 trials with fdr_estimate 0.160 against a 0.25 ceiling, so an arm that merely re-ADOPTs near the frontier buys little; what is worth spending trials on is a dimension whose measured dose ordering was established on a base that has since moved. sleeve_trail_pct was doses at 3/6/10 only under max_total_sleeve 0.9 and vol_scale_hi 0.8 (H-E147.v1 +0.1072/+0.1707 dd +0.0191; H-E147.v3 +0.1448/+0.2287 dd +0.0155). The frontier has since moved to a loose cap (H-E150.v1 at 1.05 and H-E150.v3 at 1.5 both ADOPT, train identical at +0.1232, valid +0.2389 and +0.2468 - the cap is non-binding in train and rarely binding in valid) and to the gentler vol_scale_hi 0.7 (H-E146.v1, train dd_delta +0.0069 vs +0.0241 at 0.9). Those two changes leave more sleeve capital outstanding for longer, which is precisely the state a per-sleeve adverse-move exit is supposed to govern, so the earlier ordering should be re-measured rather than assumed. Predictions carry the family's measured bias correction (bias_train -0.0073, bias_valid -0.0071 over n=320: I have been mildly optimistic, so each arm is shaded down from the naive parent-plus-effect estimate). The tight arm is the informative one: I am predicting a large valid loss, and a flat result there would falsify the trail-dose story and reassign the H-E147 spread to the cap. mid dose, the untested middle of the earlier curve; predicted to sit between the 10.0 and 3.0 arms, slightly below the parent on both windows.

H-E016 is at 343 trials with fdr_estimate 0.160 against a 0.25 ceiling, so an arm that merely re-ADOPTs near the frontier buys little; what is worth spending trials on is a dimension whose measured dose ordering was established on a base that has since moved. sleeve_trail_pct was doses at 3/6/10 only under max_total_sleeve 0.9 and vol_scale_hi 0.8 (H-E147.v1 +0.1072/+0.1707 dd +0.0191; H-E147.v3 +0.1448/+0.2287 dd +0.0155). The frontier has since moved to a loose cap (H-E150.v1 at 1.05 and H-E150.v3 at 1.5 both ADOPT, train identical at +0.1232, valid +0.2389 and +0.2468 - the cap is non-binding in train and rarely binding in valid) and to the gentler vol_scale_hi 0.7 (H-E146.v1, train dd_delta +0.0069 vs +0.0241 at 0.9). Those two changes leave more sleeve capital outstanding for longer, which is precisely the state a per-sleeve adverse-move exit is supposed to govern, so the earlier ordering should be re-measured rather than assumed. Predictions carry the family's measured bias correction (bias_train -0.0073, bias_valid -0.0071 over n=320: I have been mildly optimistic, so each arm is shaded down from the naive parent-plus-effect estimate). The tight arm is the informative one: I am predicting a large valid loss, and a flat result there would falsify the trail-dose story and reassign the H-E147 spread to the cap. loose trail: the value that won the 0.9/0.8-base curve (H-E147.v3 +0.1448/+0.2287). Expect it to add a little train and cost nothing in valid relative to the H-E150.v3 parent.

Rule 7 says judge rules by removing one at a time and letting the pipeline speak; the H-E016 line has instead been built by accretion - eighteen flags, 339 trials, 290 adoptions - and only no_restack has ever been removed from a winning stack (H-E145.v3, -0.000/-0.039). Every component in the parent stack was adopted in a different configuration from the one it now sits in, so the family's per-flag 'best' numbers in findings are attributions, not measurements at this operating point. With FDR at 0.159 and rising with each trial, the honest next step is to find out which of these eight settings the +0.123/+0.239 actually depends on, because a rule that is inert here should not be carried into any promotion case. Predictions are the parent's realized deltas minus my estimate of each component's contribution, nudged down ~0.008 for this family's measured calibration bias (-0.0074 train / -0.0075 valid, i.e. I have been mildly optimistic). I expect at least one of these four to come back INSUFFICIENT_EVIDENCE - i.e. the component does nothing - and that is the useful outcome. Drawdown headroom is ample (parent dd_delta +0.0069 train / -0.047 valid against a 0.02 ceiling), so removals are unlikely to trip the constraint. drop core_weight 1.0 back to the file default. core_weight climbed 0.85 -> 0.95 -> 1.0 across H-E098/H-E136/H-E138 with small gains each step, so I expect a small loss, not a collapse.