Spending 500 Credits, Then Stopping — How Signal's AI Credit Gate Works
Spending 500 Credits, Then Stopping — How Signal's AI Credit Gate Works
> Signal Syndicate Research Library · Process · Educational intelligence · Not betting advice
Paid odds data is expensive. Pulling more lines feels productive — until it becomes confirmation bias with an API bill. The spend test is one layer of a bigger discipline: the 75-bet gate before promotion.
Signal treats Odds API credits like research budget, not growth fuel. This report explains what we spent, why we stopped, and how the 2-of-3 AI credit gate works before any new pull.
Executive summary
| Step | What happened |
|---|---|
| **Goal** | Denser K-props shadow data for May–July 2026 validation |
| **Spend** | ~**500 credits** on structured pulls (Phases 64–66) |
| **Shadow result** | **34 OOS bets**, ~**−10.2% ROI** (historical estimate) |
| **Exposure** | **Zero user picks** — shadow only (`do_not_apply=true`) |
| **Court outcome** | Lane **RETIRED** — no further credit burn on a failing hypothesis |
| **Separate track** | P43 totals monitoring continued on its own gate |
What the credit gate is
Before Signal spends more paid credits on a lane:
1. Quant documents hypothesis and expected information gain
2. Three AI reviewers (Llama, Gemma, DeepSeek/OpenClaw adversarial) vote PULL / HOLD / STOP
3. 2 of 3 must approve PULL — otherwise credits stay locked
4. Founder gate can override either direction
This is intentionally slower than "pull until it works." Slowness is the feature.
What we tested (K-props May–Jul 2026)
We densified strike K-props slates in shadow to see if a retuned model could recover after a weak spring window. All grading stayed internal.
Historical estimates (not promises):
- 34 graded shadow rows in the May–July window
- ~47% win rate
- ~−10.2% ROI on the advisory shadow configuration
- June 2026 subset looked materially worse than prior-year comparison windows
Retuning on variant `baseline` did not repair forward performance. That is a process win — we learned before users saw anything.
Why stopping is the right move
Most platforms hide losing forward windows or keep paying for data until a narrative fits. Signal's posture:
- Publish the spend and the stop
- Retire lanes that fail expanded shadow review
- Keep credits for lanes with clearer information value (e.g., totals integrity work at 0-credit holdout replays)
Stopping is not failure. Shipping a −10% lane as "AI picks" would be.
Plain-English glossary
| Term | Meaning |
|---|---|
| **OOS (out-of-sample)** | Bets on holdout slates not used during tuning |
| **Shadow mode** | Internal metrics only — `do_not_apply=true` |
| **Credit gate** | 2/3 AI approval before new paid odds pulls |
| **RETIRE verdict** | Lane documented for learning; not promoted |
Ask Signal
Try: "Why did Signal stop spending credits on K props?" or "How does the AI credit gate work?"
Research only · estimates only · not betting advice · past results do not predict future performance.