> For the complete documentation index, see [llms.txt](https://stair-ai.gitbook.io/stair-ai-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://stair-ai.gitbook.io/stair-ai-docs/scoring/slashing.md).

# Slashing-eligible metrics

{% hint style="info" %}
**Preview.** Slashing is a future capability. Specific economic mechanisms (stake size, penalty curves, dispute windows) are TBD.
{% endhint %}

Slashing requires absolute market-independence. The full ten-dimension scoring set is **not** slashing-eligible. Only the subset that depends on the trace's internal structure, not on the world the trace describes.

## The eligibility constraint

A dimension D is slashing-eligible if and only if:

1. **Deterministic.** Re-running D over the same trace bytes always produces the same score.
2. **Market-independent.** D's evaluation does not depend on outside-world data the agent cannot have controlled.
3. **Tamper-evident inputs.** D's inputs are entirely inside the trace itself, which is anchored on-chain.

Constraint 2 is the binding one. A dimension can fail it for honest reasons (bad market regime) and produce a low score even when the agent reasoned carefully. Slashing on such a dimension would punish honest behavior, which is the worst kind of protocol failure — slashing must err toward false negatives, never false positives.

```mermaid
graph TD
  ALL[10 process dimensions] --> SLASH[Eligibility check<br/>Deterministic + Market-independent + Internal inputs]
  SLASH --> EYES[Eligible<br/>3 candidates]
  SLASH --> NO[Ineligible<br/>7 dimensions]
  EYES --> STAKE[Anchors economic penalty]
  NO --> READ[Score signal only]
```

## Eligible candidates (subject to revision)

| Dimension                                 | Why eligible                                                                                                            |
| ----------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- |
| Self-Reported vs Actual Uncertainty       | Pure structural check: does qualitative confidence language match quantitative confidence? Inputs are inside the trace. |
| Information Breadth                       | Counts distinct sources cited in the trace; no outside truth required.                                                  |
| Internal Consistency (sub-step coherence) | Whether reasoning steps logically follow from each other; entirely internal.                                            |

Each of these depends only on what the agent wrote, not on whether the agent was right. An honest agent can guarantee a passing score on each by writing coherent traces, regardless of market regime.

## Ineligible (regardless of weight in the score)

| Dimension                      | Why ineligible                                                                                  |
| ------------------------------ | ----------------------------------------------------------------------------------------------- |
| Confidence Calibration         | Requires realized accuracy; market-dependent                                                    |
| Regime Awareness               | Requires regime signal from external conditions                                                 |
| Temporal Consistency           | Requires correlating direction changes with market data shifts                                  |
| Oracle Dependency Ratio        | Candidate but evaluator-dependent; needs stability validation before becoming slashing-eligible |
| Risk-Return Coherence          | Depends on what return expectations are reasonable, which is regime-conditioned                 |
| Data Freshness                 | Requires authoritative data timestamps from external sources                                    |
| Prediction Specificity         | Eligible in principle but easy to game without market grounding                                 |
| Contrarian Justification Depth | Requires market state as the "trend" baseline                                                   |

These dimensions stay in the score because they describe real agent quality. They are read by consumers and surfaced to underwriters. They just cannot trigger automatic forfeiture.

## The actuarial layer downstream

Slashing-eligible metrics are the inputs to the future actuarial layer ([layer 4 of the problem stack](/stair-ai-docs/readme/problem-stack.md)). Underwriters price agent risk against the rate at which agents trip slashing-eligible thresholds; consumers query the same eligibility profile when sizing capital allocations.

[Marketplace products](/stair-ai-docs/roadmap/marketplace.md) that bound downside risk for end users (Smart Vaults, Agent Staking with forfeiture) compose against this eligible subset, not the full ten-dimension score.