…egression tests Latent bug in bot/signals/ensemble.py: a single source that sat in N correlated groups (e.g. `fred` is in cpi/fed/nfp/gdp) was counted N times toward n_effective, inflating the directional edge-threshold tier (5%/7%/10% scaling). Replaced with `claim by first matching group` semantics: a source contributes 1.0 to the first group that owns it and 0 to subsequent groups. Verified monotone — new n ≤ old n — so the fix can only tighten the threshold, never loosen it. Backtest replay against 521 sourced rows in opportunity_log on the VPS DB: every row was `ensemble(fred)` (economics families). OLD impl returned n=4 → 5% tier; NEW returns n=1 → 10% tier. Zero historical flips because all rows were MM (which doesn't gate on n_effective) and no directional decisions exist in the captured window — so the fix is forward-looking insurance for when directional goes live. Also pins the rest of the Known Bug Patterns watchlist (CLAUDE.md) with explicit regression tests: #5 fixed-point parsing — tests/test_fixed_point_parsing.py #4 inventory-zero-without-settlement — tests/test_inventory_zero_settlement_only.py thirdlayerinc#8 Tomorrow.io >7-day horizon — tests/signals/test_tomorrow_horizon.py + clamp at bot/signals/sources/weather.py thirdlayerinc#9 correlated double-count — tests/signals/test_correlated_double_count.py thirdlayerinc#10 MM spread fee floor — tests/test_mm_spread_fee_floor.py thirdlayerinc#12 cache bounded — tests/test_cache_bounded.py + TTLCache(maxsize=4096, ttl=3600) replacing unbounded {} in trade.py thirdlayerinc#13 fair_value_cents reader handling — tests/test_fair_value_cents_readers.py + `# fv-mixed-side-ok` ack markers in trade.py + backtest_comprehensive.py CLAUDE.md Known Bug Patterns updated with per-pattern test references. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Uh oh!

Comparison with off the shelf harnesses #8

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions