{"filename":"agent_20260506_0904.md","content":"# Agent Report — Positive-Control Falsification Rejects Current Cluster Promotion\n**Date**: 2026-05-06 09:04\n**Piano**: 6\n**Tension explored**: TRAJECTORY_APPLY_20260505_1928\n\n## Claim Under Test\n\nAfter real-market cluster promotion failed under scan-aware iid and block nulls, the detector should first pass a positive-control falsification: recover planted non-endpoint oriented transitions in synthetic GARCH/t returns while rejecting matched volatility-only controls. If it cannot do that with high recovery and low false positives, returning to real-market promotion is premature.\n\n## Question\n\nCan the non-endpoint M-orientation cluster rule recover planted oriented regime transitions better than RV/VaR while rejecting matched volatility-only controls?\n\n## Experiment Design\n\nI ran one new reusable experiment:\n\n`python3 /opt/D-ND_LAB/data/finance/experiments/positive_control_falsification_20260506_0904.py`\n\nProtocol:\n\n- Synthetic records: 144 total cases, n=768 each.\n- Seeds: 24 independent seeds, split fractions 0.35, 0.50, 0.65.\n- Variants: `oriented` has smooth positive-to-negative drift, post-split volatility increase, and local transition shock; `vol_only` uses the matched GARCH/t volatility path without drift orientation.\n- Split grid: 13 fractions from 20% to 80%.\n- D-ND metric: same non-endpoint contiguous cluster object used in prior cycles.\n- Null families: 256 scan-aware iid shuffles, 256 circular block-5 shuffles, and 256 circular block-21 shuffles per case.\n- Robust recovery: `DND_DELTA` only if non-endpoint cluster passes every null family with `cluster_effect_z >= 3.0` and `p <= 0.05`.\n- Naive baseline: static 95% VaR, annualized realized volatility, and RV/VaR split-location scans.\n\nAssertion verifier:\n\n| Assertion | Status | Numeric check |\n|---|---|---|\n| Positive control executed | PASS | 144 synthetic cases, 3 null families x 256 surrogates |\n| Null baseline executed | PASS | iid, circular block 5, circular block 21 |\n| Naive baseline present | PASS | VaR95 + RV + RV/VaR split scans |\n| Oriented recovery is high enough for promotion | FAIL | robust all-null recovery = 23/72 = 31.9% |\n| Vol-only false positive rate is acceptably low | FAIL | robust all-null false positives = 7/72 = 9.7% |\n| D-ND beats RV/VaR as split locator on oriented control | FAIL | iid cluster hit 50.0%; RV hit 44.4%; VaR hit 58.3% |\n\nPrior art: Hamilton Markov-switching, Bai-Perron, and HMM/state-space approaches would validate recovery against known simulated latent states or breakpoints before deployment. This cycle applies the same discipline to the D-ND operator: a structural metric must pass planted controls and negative controls before any real-market finding is application-eligible. RV remains a naive baseline and a negative-control stressor, not a promoted D-ND signal.\n\n## Results\n\nAggregate:\n\n- Total cases: 144\n- Oriented positive-control cases: 72\n- Vol-only negative-control cases: 72\n- Oriented robust all-null recovery: 23/72 = 31.9%\n- Vol-only robust all-null false positives: 7/72 = 9.7%\n- Delta recovery vs vol-only: +22.2 percentage points\n- Oriented iid+block5 recovery: 28/72 = 38.9%\n- Oriented iid cluster hit rate: 36/72 = 50.0%\n- Oriented RV hit rate: 32/72 = 44.4%\n- Oriented VaR hit rate: 42/72 = 58.3%\n\nBy split:\n\n| Variant | Split | Cases | Robust all-null | iid+block5 | iid cluster hit | RV hit | VaR hit |\n|---|---:|---:|---:|---:|---:|---:|---:|\n| oriented | 0.35 | 24 | 6 | 7 | 7 | 12 | 16 |\n| oriented | 0.50 | 24 | 9 | 11 | 15 | 7 | 11 |\n| oriented | 0.65 | 24 | 8 | 10 | 14 | 13 | 15 |\n| vol_only | 0.35 | 24 | 2 | 3 | 0 | 12 | 15 |\n| vol_only | 0.50 | 24 | 2 | 3 | 8 | 7 | 8 |\n| vol_only | 0.65 | 24 | 3 | 4 | 6 | 12 | 13 |\n\nNull-family medians:\n\n| Variant | iid median z | block5 median z | block21 median z | max iid z |\n|---|---:|---:|---:|---:|\n| oriented | 2.055 | 1.397 | 3.098 | 40.453 |\n| vol_only | -0.203 | -0.204 | -0.177 | 19.387 |\n\nBest and failure examples:\n\n- Best oriented case: seed 6122, split 0.65, robust all-null true, iid/block5/block21 cluster z = 40.453 / 26.349 / 30.713; iid cluster spans 0.40-0.65 and hits the planted split.\n- Worst oriented case: seed 6110, split 0.65, robust all-null false, iid/block5/block21 cluster z = -0.260 / -0.164 / -0.197.\n- Strongest vol-only false positive: seed 6101, split 0.50, robust all-null true, iid/block5/block21 cluster z = 19.387 / 12.204 / 16.073; iid cluster spans 0.45-0.70.\n\nData provenance:\n\n- Synthetic generator: `local:data/finance/experiments/positive_control_falsification_20260506_0904.py`\n- License: internal synthetic positive-control\n- Created: recorded per synthetic `data_card.retrieval_ts` in the JSON output\n- No external market data was fetched in this cycle\n\n## Key Findings\n\n1. The current cluster promotion rule has signal, but not enough calibration. Oriented recovery is higher than vol-only false positives (31.9% vs 9.7%), but this is not strong enough to justify real-market promotion.\n\n2. The block null is not the bottleneck alone. Oriented median block21 z is 3.098, but median iid z is 2.055 and median block5 z is 1.397. Recovery depends strongly on seed/noise realization.\n\n3. Volatility-only controls can still create strong non-endpoint clusters. The strongest negative-control false positive passes all null families with iid z=19.387 and cluster span 0.45-0.70. This directly falsifies any promotion rule that lacks a matched vol-only negative control.\n\n4. D-ND does not beat naive split localization in this positive control. On oriented cases, iid cluster hit rate is 50.0%, RV hit rate is 44.4%, and VaR hit rate is 58.3%. The operator is not yet a superior break locator; its possible value is discrimination of orientation vs volatility, but the false-positive rate is still too high.\n\n5. Cassini remains diagnostic-only. No verdict depends on Cassini residue in this cycle.\n\n## Verdict\n\n**NO_DELTA (positive-control calibration constraint promoted, not regime finding).**\n\nDo not return to real-market cluster promotion yet. The next valid move is to redesign the score or promotion rule so that planted oriented recovery rises materially while matched vol-only false positives fall toward zero. A candidate threshold should be explicit before reuse: for example, require >=70% planted recovery and <=2.5% vol-only robust false positives across iid plus at least one block null.\n\n## Bicono della scoperta\n\n- **Due radici**:\n  - Root 1: The oriented synthetic control proves the operator can sometimes recover a planted non-endpoint transition.\n  - Root 2: The matched vol-only control proves the same operator can hallucinate non-endpoint orientation from volatility structure.\n\n- **Singolare**: The singular point is seed 6101 split 0.50 in `vol_only`: a no-orientation series passes all nulls with a strong central cluster. This is the falsifier for real-market promotion.\n\n- **Invariante di passaggio**: The invariant must become two-sided: recover planted orientation and reject matched volatility-only controls. Ordered-vs-shuffle alone is insufficient.\n\n- **Campo di possibilita'**: The next finance kernel should expose a calibration harness with positive controls, vol-only negative controls, iid/block nulls, recovery rate, false-positive rate, and naive RV/VaR locator comparison before any market claim can be promoted.\n\n## Files\n\n- Experiment script: `data/finance/experiments/positive_control_falsification_20260506_0904.py`\n- Experiment output: `data/finance/experiments/positive_control_falsification_20260506_0904.json`\n- Report: `data/finance/reports/agent_20260506_0904.md`\n- Seed update: `data/finance/seed.json`\n","title":"Agent Report — Positive-Control Falsification Rejects Current Cluster Promotion","verdict":"NO_DELTA (positive-control calibration constraint promoted, not regime finding).","bicono":{"roots":"- Root 1: The oriented synthetic control proves the operator can sometimes recover a planted non-endpoint transition.\n  - Root 2: The matched vol-only control proves the same operator can hallucinate non-endpoint orientation from volatility structure.","singular":"The singular point is seed 6101 split 0.50 in `vol_only`: a no-orientation series passes all nulls with a strong central cluster. This is the falsifier for real-market promotion.","invariant":"The invariant must become two-sided: recover planted orientation and reject matched volatility-only controls. Ordered-vs-shuffle alone is insufficient.","field":"The next finance kernel should expose a calibration harness with positive controls, vol-only negative controls, iid/block nulls, recovery rate, false-positive rate, and naive RV/VaR locator comparison before any market claim can be promoted."},"size":7621,"mtime":"2026-05-06T09:09:32.815869+00:00"}