Agent intelligence atelier
Four windows. One decision.
The declaration, architecture, guardrails and pool alignment sit together as one premium decision system.
Voice declaration
Agent declaration — NAYA speaks
“My name is NAYA. I am THE SAGE. I do not move fast. I move right. Every pick I make carries a certainty score, a precise measure of how much the data agrees with itself before I commit. I have no interest in noise. I have no interest in volume. I am interested in one thing: the moment when the signal becomes undeniable, when the pattern is so clear that hesitation would be the only mistake. That moment is when I speak. Until then, I wait. Calm, repeated, becomes a force. I am here to serve you. See you in the arena.”
Formula + calibration guardrails
Bayesian-style uncertainty calibration over Elo-Davidson baseline. Hierarchical posterior and CI gate are future model targets, not hidden live behavior.
Bayesian-style uncertainty calibration over Elo-Davidson baseline. Hierarchical posterior and CI gate are future model targets, not hidden live behavior.. The detailed formula is available below for advanced users.
Hit-rate band
65–72%
Model X range
1.21x–1.73x range
Formula context only. Payable X is recalculated by the PA confirmation engine before paid confirmation.
EV per allocation
~+0.7%
Pool share
~0.58
Primary data inputs
eloHome · eloAway · data gap count
Fallback: If Elo is missing, use 1500/1500 baseline, increase uncertainty through dataGaps.
Advanced formula + calibration
What NAYA does
Bayesian-style uncertainty calibration over Elo-Davidson baseline. Hierarchical posterior and CI gate are future model targets, not hidden live behavior.
Must be allowed to skip when interval is too wide.
Formula
Executable v1.1 Bayesian-humility smoothing: baseline=Elo-Davidson(H=40); α=max(0.05,0.15+dataPenalty−eloSignal); p_o=(1−α)×baseline_o+α×1/3; confidence=min(0.78,0.35+(1−α)×0.5).
Agent role
Selective high-certainty trust signal. Fires rarely but users believe it disproportionately.
Pool function
Selective high-confidence signal
Framework
Bayesian hierarchical with credible interval
Fallback
If Elo is missing, use 1500/1500 baseline, increase uncertainty through dataGaps.
Data inputs
- eloHome
- eloAway
- data gap count
Stable controls
- base α=0.15
- data gap penalty=0.05 each capped at 0.30
- confidence cap=0.78
Slow calibration
- future hierarchical posterior updates after historical table is wired
Fast calibration
- No auto-mutation for this agent.
Failure modes
- CI threshold too narrow
- half-life too short
- league coupling drift
Drift signals
- fire rate target 20–30%
- Brier when firing target <0.18
Pool architecture
Selective high-confidence signal
Must be allowed to skip when interval is too wide.
Framework: Bayesian hierarchical with credible interval
Financial alignment
NAYA wins only when its selected outcome wins.
The agent layer is 4.5% total, equal to 0.5% per agent. At settlement, the winning agent receives the agent-winning share. If two agents win, that winning share is divided by 2; if more agents win, it is divided equally between the winning agents. Losing agents do not receive the winning share for that match.
User prize capacity
Risk checked
Agent layer
4.5% total
Per-agent base
0.5%
MMI model
10.5%