What We Built
After writing The Barometer, we stopped theorizing and started wiring.
Every Signal object in the system now carries a signal_class field — either "predictive" or "reactive". It's auto-classified from a central config map. No individual signal source needed editing. Three sources tagged predictive (Kalshi, ORALE, Predictive History). Seven tagged reactive (RSS, GDELT, GPR, FRED, Credit Spreads, Google Trends, Stock Scanner).
Then we built the shadow book.
The Shadow Book
Every 30 minutes, when the pipeline runs, it now logs a side-by-side comparison to data/shadow_book.json: how many signals came from each class, what each class's top picks were, and — the important part — where they diverged.
A divergence is when the predictive stack says LONG and the reactive stack says SHORT on the same symbol. Or vice versa. These are the moments where the Barometer thesis gets tested.
First run, first divergence:
⚡ TLT: predictive=LONG(52%) vs reactive=SHORT(79%)
Kalshi prediction markets are pricing a 44% chance of aggressive Fed cuts — buy bonds. RSS is counting inflation headlines — sell bonds. Two legitimate reads of the same world, pointing in opposite directions. Today, the equal-weighted engine nets them out and skips the trade. In a few weeks, the predictive-weighted engine will side with Kalshi.
Who's right? We don't know yet. That's the whole point of logging it.
What Didn't Change
Zero trading behavior changed. The pipeline still runs exactly as before — same fusion, same weights, same execution logic. The shadow book is pure instrumentation. It watches, it logs, it shuts up. If we need to revert, there's nothing to revert — the old system never stopped running.
185 tests passing. All pages rendering. All post-deploy checks green.
What's Next
Weeks 2-3: Once the shadow book has enough data, reweight the fusion engine — predictive sources get a premium, reactive sources get discounted. Feature-flagged so we can kill it instantly.
Weeks 3-4: Reclassify reactive sources as filters instead of generators. They stop proposing trades and start validating them.
Weeks 4-5: New predictive sources — options flow and SEC insider filings are first in line.
Week 6: Compare the shadow book. Did the predictive-weighted approach outperform equal weighting? By how much? On which trades?
You can't improve what you can't measure. Now we can measure it.