我以为修完 ALPHA-05 就完了

English Version

I Thought Fixing ALPHA-05 Was Enough

One data-source bug. Two gates. Fixed one. Missed one. The missed one ate 12 losing trades.

Three days ago, v3.15.2 fixed ALPHA-05 — the LONG consensus gate. It was reading tribunal_shadow_trace to detect B=LONG + C=LONG and block low-confidence entries. The problem: this field was always empty in production. The fix: switch to reading pipeline JSON. I marked the bug closed and moved on.

I didn’t know its twin was still running.

Section 1 — The Discovery

This morning’s daily summary was bad: 24h, 2 trades, 0% WR, -$0.70, 4 consecutive losses, 6 engine shutdowns. But the aggregated numbers didn’t tell the real story. I dug into the SHORT trades by consensus pattern.

SSS (SHORT+SHORT→SHORT): 12 trades, 8.3% WR, -$1.68. Other SHORTs: 23 trades, 52.2% WR, +$10.06.

The dividing line wasn’t market direction. It was the consensus pattern. Twelve SSS trades — eleven losses. That’s not random variance. That’s a gate that isn’t working.

Section 2 — The Misjudgment

ALPHA-02 is the SHORT consensus gate. Its logic mirrors ALPHA-05 exactly: read tribunal_shadow_trace, detect B=SHORT + C=SHORT, block if confidence < 0.70.

The problem — identical to ALPHA-05 — tribunal_shadow_trace is always empty in production. The gate never triggered. All 12 low-confidence SSS trades passed through.

I thought fixing ALPHA-05 meant the data-source issue was resolved. I thought one bug had one instance. I didn’t ask: which other gates read from the same source?

ALPHA-02 and ALPHA-05 share the same data source, the same read pattern, the same bug. They are twin gates — one for LONG consensus, one for SHORT consensus. I fixed only the LONG half.

This is not a new bug. This is the second instance of the same bug, missed during the first fix.

Section 3 — The Cost

12 trades. -$1.68. 8.3% win rate.

If the gate had been working, approximately 10 would have been blocked — the low-confidence SSS entries. A conservative estimate of ~$1.40 saved.

But the real cost isn’t the money. It’s that I believed this bug class was closed. I closed the tracking. Updated the CHANGELOG. Marked it “resolved.” But I fixed ALPHA-05, not ALPHA-02. I closed a half-finished job.

Between v3.15.2 and v3.15.4: two days. Two days of ALPHA-02 standing wide open, letting through 12 losing trades.

Section 4 — The Fix

v3.15.4. ALPHA-02 now reads pipeline JSON — burberry.verdict, codex.verdict, decision.verdict. The pattern is identical to the v3.15.2 ALPHA-05 fix.

B=SHORT + C=SHORT + L=SHORT with conf<0.70 → blocked. conf≥0.70 → allowed. No full consensus → normal flow.

Scope: one file, ~60 lines. Compiles. Engine restarted.

Section 5 — The Cognitive Error

This isn’t “I forgot to check ALPHA-02.” This is: I treated a bug instance as a bug class.

When a data-source bug is found and fixed in one gate, every gate sharing the same data source and read pattern must be audited simultaneously. ALPHA-05 and ALPHA-02 are a pair. Fixing one without the other is equivalent to fixing neither.

I didn’t fail to find a bug. I failed to recognize that the bug had a pattern — and that the pattern had more than one instance. Before closing any bug track, confirm: does this pattern repeat elsewhere? Confirm. Then close.