AIONNIFTY full-band containment
46.30%
25 / 54 eligible sessions
27 April-23 July 2026.
Veritas / Permanent historical evidence
The historical record.
Misses included.
A dated, partial retrospective audit of opening-window forecasts. These results remain visible alongside the current rolling scorecard, without rewriting the original receipts.
Eight invalid-data dates were wrongly counted.
The 12 November audit included eight dates inside the established invalid-data window of 24 July-17 September 2026. The longstanding eligibility rule excludes those dates from accuracy calculations. We have corrected the denominators and retain every original row and value below. This is an eligibility correction, not model recalibration or the removal of legitimate misses.
The affected dates are 24, 27, 28, 29, 30 and 31 July, and 3 and 4 August 2026. All 62 audited sessions per index remain visible; 54 are eligible and 8 are excluded.
| Metric | Superseded original | Corrected |
|---|---|---|
| Reference-price hit rate | 43/62 = 69.35% | 35/54 = 64.81% |
| Full-band containment | 26/62 = 41.94% | 25/54 = 46.30% |
| WEAK lean accuracy | 19/48 = 39.58% | 19/40 = 47.50% |
STRONG lean accuracy is unchanged at 27/35 = 77.14%. The lean cohort is separate from the range cohort. These corrected historical subsets do not establish present or future accuracy.
Reference-price hits and full-range containment are different tests.
A reference-price hit means the recorded reference price fell inside the predicted band. Full-band containment means the entire observed 09:15-09:30 range stayed inside it. A single price being inside the band does not establish that the entire range was contained.
NIFTY Reference-Price Hit Rate
64.81%
35 / 54 of the same eligible sessions.
Historical reference definitions varied.
Scope of the audit
186
Retained index/session observations.
162 eligible + 24 excluded; 62 audited per index.
What was measured. This is a partial retrospective reconstruction using complete opening-window five-minute bars and dated pre-open forecasts. It is not the full historical population or an original published containment result. The table retains 62 audited sessions per index from 27 April-4 August 2026, including eight invalid-data dates marked Excluded. Corrected metrics use only 54 eligible sessions per index, dated 27 April-23 July. Original values and Boolean results remain in the JSON for provenance, not as valid scores for excluded dates. No forecast or receipt was overwritten. A calibration cutover has not been verified.
Full-band containmentNot loadedAwaiting the published data
Eligible index sessionsNot loadedAll results retained
Reference-price hit rateNot loadedUnknown results are not misses
Loading published results...
| Date | Index | Predicted band | Actual 09:15-09:30 range | Band hit? | Reference price | Reference hit? |
|---|
Scroll within the table for more dates and columns. Excluded records retain original prices and ranges but do not count as hits or misses. Reference-price hits and full-band containment are different tests.
Download the corrected audit data (JSON). The table includes all 186 retained observations: 162 eligible and 24 excluded, not only hits. The original public receipts retain their original definitions.
Directional accuracy, split by reliability
NIFTY 50 archived validations dated 7 April-11 November 2026. These direction results use a different cohort from the range audit above: 75 eligible graded outcomes and eight invalid-data records retained but excluded from rates. Unknown reliability grades are outside both groups.
STRONG Lean Accuracy
77.14%
27 correct / 35 scored sessions.
WEAK Lean Accuracy
47.50%
19 correct / 40 eligible scored sessions.
Excluded records retained
8
Invalid-data WEAK records remain in the audit JSON. They are excluded from corrected accuracy denominators.
WEAK is an explicit uncertainty label. The model specification says the directional lean should not be relied on in that state. Neither group is a current-model forecast of future accuracy.
A permanent record, not a reset.
The 30-session view keeps recent results readable. It does not remove the older record or prove a model improvement. This historical audit and the full public receipt archive remain available so readers can inspect dates, denominators, definitions and misses.