Levandor

Phase-0 Depth-Aware Selectivity Re-Run

4 min read

Phase-0 decisive offline gate: does the candidate edge survive when fills are capped to actually-recorded book depth? Verdict: SURVIVES. The prior t=4.61 result is not a top-of-book mirage.

Context

This run directly follows crypto-shortterm-phase0-depthprofile, which recorded real book depth in the candidate regime and weakened the unfillable-mirage hypothesis. The remaining skeptic question was: does the edge hold when we enforce a two-tier fill model capped to the recorded best_*_sz and 3¢ cumulative depth?

  • Binary: crates/analysis/src/bin/depthaware.rs
  • Report: docs/superpowers/phase0-depthaware-report.md
  • Commit: bba5739
  • Windows tested: 524 candidate-depth windows (τ∈[60,150)s, |p_fair−0.5|≥0.30)
  • Bonferroni significance bar: t ≥ 2.99

Fill Model (Two-Tier Depth Walk)

For Agents

This is the exact fill logic encoded in depthaware.rs. Future analysis runs must use this same model for comparability.

  1. Fill min(S, touch_sz) shares at the touch price.
  2. Fill min(S − q1, depth_3pct − touch_sz) shares at touch ± 3¢ (the sweep tier).
  3. Any unfilled remainder earns nothing (zero credit).

A latency buffer sensitivity check (+0.5¢ penalty on each fill price) was also run — verdict is unchanged.

Decay Table

524 candidate-depth windows, Bonferroni bar t≥2.99:

S (shares)t-stat¢/sharefill_rate
505.525.01¢99.8%
1005.364.87¢99.6%
2505.014.53¢99.0%
5004.584.13¢97.9%
10004.013.55¢94.8%
20003.012.61¢86.8%
∞ uncapped5.675.15¢100.0%

Capacity ceiling: 2000 shares (t=3.01, ¢/share=2.61, fill_rate=86.8%). Beyond 2000 shares the fill rate and ¢/share drop below the Bonferroni bar.

Decay shape: SANE/DECLINING — ¢/share falls monotonically as S grows. This is exactly what a real liquidity-bounded edge looks like. A flat curve or uptick would signal a top-of-book artifact; no such anomaly is present.

Verdict: SURVIVES

T19 gate status: SURVIVES offline depth-cap test — not yet a GO

The prior t=4.61 result (from crypto-shortterm-phase0-selectivity) was NOT a top-of-book mirage. Edge clears t≥2.99 at all tested sizes up to 2000 shares under realistic fill caps.

Next gate: real-fill test on a funded account at tiny size to measure actual latency cost. Only after latency cost is bounded does it make sense to build live machinery.

The +0.5¢ latency buffer sensitivity test was computed and does not change the verdict.

What This Rules Out

  • The unfillable-mirage hypothesis (books so thin the edge is fictitious) is rejected at all sizes ≤2000 shares.
  • The top-of-book artifact risk (edge is flat at ∞ and only appears to decay because of fill capping) is not present — decay is monotonic and physically consistent.

What Remains Open

  1. Latency survival — recorded best_*_sz is a 1s snapshot. Real fills ~150ms after signal may see a degraded book. The latency cost is unknown until a funded real-fill test.
  2. Capacity at scale — 2000 shares (~$2,000/trade) is the offline ceiling. Live capacity may be lower due to fill cancellations and partial fills not captured in the snapshot model.
  3. Live machinery — no execution infrastructure exists. Building it is the next engineering phase, but only justified after the latency cost is bounded.

Window Count Audit (commit a3e0468)

The report header showed “Candidate-depth windows: 524” while the n_win column in the decay table showed 770. These count different things — no double-counting bug.

Definitions

FigureSourceCalibrationFolds included
524is_candidate() on raw p_fairnone (pre-walk-forward)ALL data including fold 0 (train-only)
770is_candidate() on calibrated p_calper-fold isotonicOOS test folds 1–5 only (fold 0 excluded)

No double-counting in the t-stat denominator

Test fold window slices are windows[fold_starts[i]..fold_ends[i]], which are strictly non-overlapping. Each window_id appears in at most one test fold. N=770 in the t-stat denominator is legitimate. T-stats are unchanged.

Why 770 > 524

Calibration sharpens an under-confident raw model, net-adding more windows past |p|≥0.30 than it removes, even though fold 0 (~1/6 of all windows) is entirely excluded from the OOS count. The ~47% increase in candidate windows (524→770) is a legitimate calibration effect, not an artifact.

For Agents

The 524 figure (header) and 770 figure (decay table) are both correct and measuring different populations. Do not flag this as a discrepancy in future audits — it is intentional and documented here.

Code fixes applied

  • n_candidate_windows split into n_raw_candidate_windows (524) and n_oos_candidate_windows (770); both printed in a “Window Counts” section of the report with precise definitions.
  • Hardcoded literal "523" in the build_verdict MARGINAL branch replaced with the live n_windows variable.
  • Local magic number 31_536_000.0 in tau_secs() replaced with the SECS_PER_YEAR constant imported from the types crate (types::SECS_PER_YEAR = 31_536_000.0, 365 days). No numeric mismatch — consistency cleanup only.
  • clippy, fmt, and build all clean post-fix.

Graph View

Backlinks