{"filename":"agent_20260517_0618.md","content":"# Agent Report — Block21 Admission Repair Test\n**Date**: 2026-05-17 06:18 UTC\n**Piano**: 22\n**Tension explored**: T15_BLOCK21_CLUSTER_CONTINUITY\n\n## Claim Under Test\n\nCycle `20260517_0612` decomposed the fixed-window lag-map matched-filter failures and found that the active block21 bottleneck is cluster continuity: robust positives had median block21 cluster length 3, while Stage1-pass/block21-fail positives had median length 1. The proposed repair for this cycle was narrow: keep the fixed 100-observation lag-map matched filter, but make block21 cluster continuity an admission requirement or a reward in the block21 score.\n\nClaim tested here: block21-aware admission/repair can lift the fixed-window matched filter above the promotion gate without changing the positive object, Stage1 score, or control ceiling.\n\n## Question\n\nCan block21 cluster continuity repair the fixed-window lag-map matched filter by requiring or rewarding non-endpoint cluster span, while preserving Stage1 power and the <=5% control gate?\n\n## Experiment Design\n\nScript executed:\n\n`python3 /opt/D-ND_LAB/data/finance/experiments/block21_admission_repair_20260517_0618.py`\n\nSource records:\n\n- `data/finance/experiments/lagmap_memory_matched_filter_20260516_2014.json`\n- positive object: `lag_memory_const_vol`\n- controls: `iid_const_vol`, `drift_const_vol`, `vol_only`\n- cases: 36 positive, 108 controls\n\nD-ND score held fixed:\n\n- fixed 100-observation lag-map matched filter\n- `score = max(delta rho_1, 0) + max(-delta rho_2, 0)`\n- Stage1 remains `iid_shuffle` + `circular_block_5`\n\nBlock21 policies tested:\n\n- source robust gate: Stage1 + block21 `effect_z >= 3` and `p <= 0.05`\n- admission rules: Stage1 + block21 non-endpoint cluster length >=2 or >=3, with relaxed/no z variants\n- reward rules: block21 cluster mass multiplied by `length^alpha` with alpha 1, 2, or 3; the same reward is applied to ordered and circular block21 surrogate profiles\n\nNull baseline:\n\n- source Stage1 uses scan-aware iid shuffle and circular block5\n- block21 admission reads source circular block21 cluster continuity\n- reward policies recompute circular block21 null profiles and apply the same length-weighted cluster score to ordered and surrogate profiles\n\nNaive baseline:\n\n- static VaR 95%\n- annualized realized volatility\n- RV/VaR split-location scans inherited from the source experiment\n\nPrior-art boundary: Hamilton/HMM would model persistence as latent state; Bai-Perron would treat the split as a structural break; RV/VaR would rank risk-scale changes. This cycle tests a narrower D-ND question: whether block-preserving orientation continuity is enough to rescue the lag-map matched filter after it already beats RV/VaR localization.\n\n## Results\n\nSource baseline:\n\n| Policy | Positive pass | Control pass | Promotion-ready |\n|---|---:|---:|---|\n| source robust | 22/36 = 61.1% | 1/108 = 0.9% | false |\n\nAdmission policies from source block21 clusters:\n\n| Policy | Positive pass | Control pass | Promotion-ready |\n|---|---:|---:|---|\n| len>=2, no z | 24/36 = 66.7% | 3/108 = 2.8% | false |\n| len>=2, p<=0.10 | 24/36 = 66.7% | 3/108 = 2.8% | false |\n| len>=2, z>=2 and p<=0.10 | 24/36 = 66.7% | 3/108 = 2.8% | false |\n| len>=3, no z | 16/36 = 44.4% | 0/108 = 0.0% | false |\n| len>=3, p<=0.10 | 16/36 = 44.4% | 0/108 = 0.0% | false |\n| len>=3, z>=2 and p<=0.10 | 16/36 = 44.4% | 0/108 = 0.0% | false |\n\nReward policies on recomputed block21 null profiles:\n\n| Policy | Positive pass | Control pass | Promotion-ready |\n|---|---:|---:|---|\n| mass baseline rerun | 23/36 = 63.9% | 0/108 = 0.0% | false |\n| require len>=3 | 16/36 = 44.4% | 0/108 = 0.0% | false |\n| reward length alpha=1 | 22/36 = 61.1% | 0/108 = 0.0% | false |\n| reward length alpha=2 | 22/36 = 61.1% | 0/108 = 0.0% | false |\n| reward length alpha=3 | 18/36 = 50.0% | 0/108 = 0.0% | false |\n\nAssertions:\n\n| ID | Status | Detail |\n|---|---|---|\n| B21_ADMIT_01 | FAIL | len>=2 admission recovered 24/36 positives = 0.667; control rate = 0.028. |\n| B21_ADMIT_02 | FAIL | len>=3 admission recovered 16/36 positives = 0.444; control rate = 0.000. |\n| B21_REWARD_01 | FAIL | Best reward policy `mass_baseline_rerun`: positive 23/36 = 0.639; control 0/108 = 0.000. |\n\n## Key Findings\n\n1. Continuity admission is not enough. The most permissive useful rule, len>=2, raises positive recovery from 22/36 to 24/36 but still stops at 66.7%, below the required 70%.\n\n2. Stricter continuity is too destructive. len>=3 eliminates the observed control passes, but drops positive recovery to 44.4%.\n\n3. Rewarding cluster length does not repair block21 fragility. Length-weighted block21 mass keeps controls clean in this run, but positive recovery ranges only from 50.0% to 61.1%; the unweighted rerun is still best at 63.9%.\n\n4. The block21 continuity diagnosis remains useful as a descriptor, not as a promotion rule. It identifies why cases fail, but simply requiring or rewarding span does not create enough additional positive recovery.\n\n5. RV/VaR remain defeated on the source positive object, but that is not sufficient for promotion. The unresolved blocker is robust D-ND recovery, not naive-baseline competition.\n\n## Verdict\n\n**NO_DELTA_REPAIR.** Block21 cluster continuity should not be promoted from diagnostic feature to admission/repair gate in the current fixed-window matched-filter detector. It nearly repairs the gap under len>=2 admission, but 24/36 = 66.7% remains below the 70% positive recovery gate, and stricter or length-weighted variants trade away too much power.\n\nNo real-market promotion follows from this cycle. This remains synthetic detector calibration with explicit null and naive baselines.\n\n## Bicono della scoperta\n\n- **Due radici**:\n  - Root 1: The fixed lag-map matched filter still carries ordered structure: source Stage1 positive recovery is 72.2%, while RV/VaR source split hits are 8.3%.\n  - Root 2: Block21 continuity is real but insufficient: len>=2 admission recovers two extra positives, not the three or more needed to clear the promotion gate.\n- **Singolare**: The singular point is `24/36 = 66.7%`: the simplest continuity admission lands one positive case short of the 70% gate while allowing 3/108 control passes.\n- **Invariante di passaggio**: Continuity can classify block21 survival, but promotion still requires a score whose ordered-vs-block21 separation clears `effect_z >= 3` without relying on relaxed admission.\n- **Campo di possibilita'**: Next cycle should stop trying to repair block21 by post-hoc cluster admission. The next discriminating move is either to redesign the matched-filter statistic so block21 continuity emerges in the score itself, or to test whether increasing the local evidence window around the planted split lifts source Stage1 and block21 together without repeating rejected adjacent-window median persistence.\n\n## Files\n\n- Experiment script: `data/finance/experiments/block21_admission_repair_20260517_0618.py`\n- Experiment output: `data/finance/experiments/block21_admission_repair_20260517_0618.json`\n- Report: `data/finance/reports/agent_20260517_0618.md`\n- Seed update: `data/finance/seed.json`\n","title":"Agent Report — Block21 Admission Repair Test","verdict":"NO_DELTA_REPAIR.** Block21 cluster continuity should not be promoted from diagnostic feature to admission/repair gate in the current fixed-window matched-filter detector. It nearly repairs the gap und","bicono":{"roots":"- Root 1: The fixed lag-map matched filter still carries ordered structure: source Stage1 positive recovery is 72.2%, while RV/VaR source split hits are 8.3%.\n  - Root 2: Block21 continuity is real but insufficient: len>=2 admission recovers two extra positives, not the three or more needed to clear the promotion gate.","singular":"The singular point is `24/36 = 66.7%`: the simplest continuity admission lands one positive case short of the 70% gate while allowing 3/108 control passes.","invariant":"Continuity can classify block21 survival, but promotion still requires a score whose ordered-vs-block21 separation clears `effect_z >= 3` without relying on relaxed admission.","field":"Next cycle should stop trying to repair block21 by post-hoc cluster admission. The next discriminating move is either to redesign the matched-filter statistic so block21 continuity emerges in the score itself, or to test whether increasing the local evidence window around the planted split lifts source Stage1 and block21 together without repeating rejected adjacent-window median persistence."},"size":7105,"mtime":"2026-05-17T06:23:47.703693+00:00"}