Gate B: how much does bar-derived fill detection overstate real fills?
Question
How much bar-derived fill detection overstates real fills, calibrated against recorded order-book depth.
Verdict
NO OVERSTATEMENT MEASURED — the median item's bar-derived fill count differs from recorded depth by -14.29 percentage points of attempted cycles, and the bar panel MISSED fills the recorded book shows, which is a defect in the opposite direction and is not a reason for confidence; 28% of items overstate. Under the touch-only rule ... bar data is fit for purpose for fill detection at this configuration. This is a FIRST calibration over 51.9 hours of recorded depth, 400 items, 2,770 resolvable attempted cycles ... not a settled constant.
In plain language (2 terms)
This site's own explanation of the words above. The verdict itself is published exactly as the research index records it and is not rewritten here.
- attempted cycle
- One full attempt at the strategy: place a buy order, wait, and sell if it fills. Attempts that never filled are counted too, so an average per attempted cycle cannot be flattered by dropping the ones that did not work.
- touch_only
- A deliberately generous rule for deciding an order filled: it counts a fill whenever the market price reached the order's price at all. Real fills are a subset of these, so a result measured this way is an upper bound on what could have happened, not an estimate of what did.
- Population
- 400-413 items overlapping the 2h panel.
- Resolution
- 2-hour patience, 2h panel; touch-only and volume_confirmed fill rules.
- Cutoff
- Recorder history 2026-08-04 20:31 -> 2026-08-07 00:26 UTC, read under the named sealing exemption gate_b_calibration (#63).
- Baseline
- Recorded order-book depth (the frozen panel/recorder overlap) as the ground-truth fill count.
Limitations
FIRST calibration over 51.9 hours / 2,770 resolvable attempted cycles, 0.080% of the 2h panel -- not a settled constant; covers only 2/6/12-hour patience, not 1/3/7-day horizons.
Notes
Regenerated in place twice (after #67's sell-leg resolution-window defect, then after #126's sub-bar lookahead fix). The document records its own prior superseded verdicts inline; there is no separate archived file for those prior figures, so no cross-file supersedes/superseded-by entry applies. Covers only 2/6/12-hour patience -- does NOT calibrate the 1/3/7-day horizons Gate A's long-patience region or any 3/7-day recommendation would need (ADR-0010).
Source
- Citation
- po-research:gate-b-calibration:cb4ddea11be3db9b
- Write-up
- docs/research/gate-b-calibration.md
- Data
- docs/research/gate-b-calibration.json, docs/research/gate-b-overstatement.csv
- Artifact
research-index.jsonv1 — sha256:cb3d976779a5cbcb4188039a8d26d5e558a2d844030a638c2f16fe5a8eb0a2fe- Build data cutoff
- 2026-08-12
- Issues
- #63, #65, #67, #70, #126