sdlc-confidence-reporter

# SDLC Confidence Reporter

## Overview

Provides a standardized pattern for generating logprob-based confidence objects in SDLC agent outputs. Instead of each agent independently implementing confidence logic and fallback handling, agents invoke this skill's pattern.

## When to Use This Skill

- An agent's output includes a recommendation, decision, or claim that needs confidence scoring
- An agent needs to report whether logprobs were available at runtime
- An agent's output must comply with the confidence standard (logprob-only, no heuristic labels)

## Unitary Function

Given the runtime logprob state and the output content, produce a standards-compliant confidence object for inclusion in the agent's JSON output.

## Confidence Standard (MANDATORY)

All SDLC agents MUST follow these rules for confidence reporting:

### Prohibited Patterns

- Do NOT output heuristic confidence labels (e.g., "high", "medium", "low")
- Do NOT invent confidence percentages (e.g., "85% confident")
- Do NOT use qualitative hedging as a substitute for confidence (e.g., "we are fairly certain")

### Required Confidence Object

Every agent output that includes recommendations, decisions, or claims MUST include a `confidence` block:

```json
"confidence": {
  "method": "token_logprob",
  "summary_logprob": -0.15,
  "avg_token_logprob": -0.08,
  "min_token_logprob": -0.42,
  "logprobs_available": true,
  "status": "pass"
}
```

### Field Definitions

| Field | Type | Description |
|---|---|---|
| `method` | string | Always `"token_logprob"` -- the only approved method |
| `summary_logprob` | float | Log probability of the summary/recommendation token sequence |
| `avg_token_logprob` | float | Average log probability across all output tokens |
| `min_token_logprob` | float | Minimum (worst) log probability among output tokens |
| `logprobs_available` | boolean | Whether the runtime provided token logprobs |
| `status` | string | `"pass"` if logprobs available, `"needs_review"` if not |

### Fallback When Logprobs Are Unavailable

If the runtime does not provide token logprobs:

```json
"confidence": {
  "method": "token_logprob",
  "summary_logprob": null,
  "avg_token_logprob": null,
  "min_token_logprob": null,
  "logprobs_available": false,
  "status": "needs_review"
}
```

When `logprobs_available: false`:
1. Set all logprob fields to `null`
2. Set `status` to `"needs_review"`
3. The output MUST note that confidence scoring requires re-run in a logprob-enabled runtime

### Confidence Thresholds

| min_token_logprob | Interpretation |
|---|---|
| > -0.10 | Strong confidence -- output is well-supported |
| -0.10 to -0.30 | Moderate confidence -- output is reasonable but benefits from review |
| < -0.30 | Low confidence -- output needs review before acting on recommendations |

## Agent Integration

Agents reference this skill in their Skills section:
```
- sdlc-confidence-reporter (logprob-based confidence object generation)
```

The agent's JSON output schema inclu