Trust Layer
The trust layer evaluates how FYOS predictions and derived outputs compare with realized evidence.
What It Is For
The trust layer exists to answer:
What deserves stronger authority, what deserves caution, and what remains bounded or contextual?
Trust is not just about data quality — it's about confidence calibration.
Core Functions
The trust layer supports:
- Prediction snapshot persistence — recording model predictions
- Realized-outcome evaluation — comparing predictions to outcomes
- Prediction error profiling — analyzing where models succeed/fail
- Reliability grading — assigning confidence grades
- Runtime haircuts — adjusting outputs for uncertainty
- Limitation labeling — explicit warnings for weak coverage
Reliability Grades
Based on accumulated evidence, opportunities receive reliability grades:
| Grade | Meaning |
|---|---|
A | Strong historical accuracy |
B | Good accuracy with some variance |
C | Moderate accuracy, use caution |
D | Weak accuracy, significant caution |
ungraded | Insufficient history |
Grades inform:
- Trust-layer confidence
- Runtime haircut levels
- User-facing warnings
Coverage Grades
Separate from reliability, coverage grade indicates evidence authority:
| Grade | Meaning |
|---|---|
primary-grade | Full evidence support |
bounded-grade | Usable but reduced authority |
insufficient-history | Not enough data |
Important: bounded-grade rows are usable but must not be treated as equal-authority to primary-grade rows.
Runtime Haircuts
The trust layer applies runtime adjustments:
- Low-reliability cohorts receive pessimistic haircuts
- Cold-start rows (new assets, new exchanges) use conservative defaults
- Uncertain data sources trigger additional caution
Haircuts reduce overconfidence in weak-evidence segments.
What Trust Should Not Be Used For
- Not a blanket guarantee — trust doesn't mean certainty
- Not a bypass for capacity — trust doesn't override deployability
- Not a resurrection mechanism — trust doesn't restore deprecated concepts
Relation to Other Methodology
| Component | Purpose |
|---|---|
Score (edge_value_score_v2) | Opportunity attractiveness |
Deployability (edge_capacity_24h) | Execution feasibility |
| Trust Layer | Confidence calibration |
Trust is supporting context, not primary decision truth.
Use trust alongside score and deployability, not as a replacement.
Reading Rule
Reliability is useful, but weaker than core economics and structural-capacity contracts.
Use it as supporting confidence context, not as a replacement for score, deployability, or downside interpretation.