Shadow Mode Explained — Tracking Lanes Users Never See
Shadow Mode Explained — Tracking Lanes Users Never See
> Signal Syndicate Research Library · Process · Educational intelligence · Not betting advice
Shadow mode is how Signal tests ideas without turning them into products overnight. The model runs, ROI and drawdown get logged, AI reviewers debate the evidence — but `do_not_apply=true` means no picks, no policy change, no marketing.
This page explains why that separation matters — and how it differs from "paper trading" hype.
What shadow mode is (plain English)
Imagine a research lab behind a locked door. Experiments run daily. Dashboards update. Nothing on the public site changes until:
1. Holdout sample is honest and large enough (see 75-bet gate)
2. Multi-AI courtroom clears promotion risk
3. Founder gate approves JSON / credit / exposure changes
Shadow is that locked lab.
What shadow mode is not
| Myth | Reality |
|---|---|
| "Shadow picks you can tail" | **No.** Results are observational only. |
| "PASS in shadow = live tomorrow" | **No.** Small-n PASS is **hypothesis** until expanded holdout. |
| "Shadow overrides a broken base model" | **No.** Model layer and filter layer are scored separately. |
Real example — two layers, one shadow track
On NBA totals we fixed a look-ahead stats bug (full-season refresh inflating holdouts). After the fix, an edge filter called B_tight passed on a tiny locked holdout — while the base projection model still failed at scale.
We logged B_tight in shadow. We did not adopt it in production JSON. Full write-up: When the Filter Passes and the Model Doesn't.
That is shadow mode working as designed: catch the tempting small-sample win before it becomes a product.
The `do_not_apply` flag
Hard internal flag meaning:
- Metrics feed Research Library and Ask Signal answers
- No user-facing signal cards
- No auto-exposure in dashboard pick feeds
- No "we're live on this lane" marketing
If you see impressive ROI in a library report, check whether the lane was shadow-only. We label it.
When shadow graduates
Shadow results alone never graduate. Typical chain:
Shadow logging → expanded holdout → 2/3 AI court → founder gate → publish verdict → (maybe) promote
Promotion gates like 75 OOS bets apply even after shadow PASS.
Why bettors and creators should care
Creators need repeatable research stories — discovery, validation, conflict, decision — not lock graphics.
Bettors need to know whether a platform tests in the dark before shipping in the light. Shadow mode is the proof.
Ask Signal: "What does shadow mode mean?" or "Why track a lane if users can't bet it?"
Research only · estimates only · not betting advice.