PRAKHARAI SYSTEMS THAT DEFEND PEOPLE
PS©2026
← Back

03 · Agentic research-integrity gate · 2026

SignalGate

Quant research teams do not lack candidate signals. They lack gates. Dozens of ideas arrive each week from LLM generators, vendor feeds, and papers, and most are spurious. The scarce resource is the researcher hour spent rebuilding each backtest to hunt for lookahead.

SignalGate investigates each candidate like a fraud case: statistical probes as tools, verdicts with receipts, and silence unless a signal deserves a researcher's hour. It recommends, never trades.

0.925
spurious catch vs 0.475 baseline
p=0.00029
McNemar paired test
0.0
false-reject on sound signals
Byte-identical
eval repro, asserted in CI

FIVE-SIGNAL ENGINE · PRECISION 95.33% ON 300 ADVERSARIAL CLAIMS

Internal consistency96%
Grounding check93%
Numeric plausibility89%
Temporal validity97%
Source support91%

Probes, not opinions

A static lint pass and an investigator agent select from four sandboxed verification probes: timestamp alignment, label permutation, regime subsample, and turnover-and-cost sanity. The agent chooses which checks to run; thresholds live in code, never in the model.

Verdicts land as REJECT_SPURIOUS, NEEDS_REVIEW, or PROMISING, each carrying its two strongest numeric receipts and a full evidence bundle.

Measured, honestly

Across 60 seeded cases over six strata: spurious catch 0.925 against a static-lint baseline of 0.475, prose-hidden lookahead lifted from 0.0 to 1.0, and false-reject held at 0.0 on sound signals. McNemar paired test p = 0.00029.

The honest changelog ships too: a bare-prompt agent hit 1.0 catch but falsely rejected 0.875 of sound specs. The four probes are the contribution that restored false-reject to 0.0. Every number regenerates byte-identically from one make command, asserted in CI.

Next caseTruthLayer