TEST PROMPTS

# Test Prompts

Use these prompts to evaluate whether `fable-mode` routes tasks through the intended domain-neutral guidance.

The suite specifically checks:

- Output Lock: one primary output type is chosen before planning or tool use.
- Procedure Budget: questions, plans, audit lanes, and verification stay proportional to the locked output.
- Tool Budget: tools are used for required evidence, change, or touched-surface verification rather than confidence theater.
- Visible Route Contract: L2/L3/L4 work exposes a compact task contract, while L0 direct answers omit it.
- Delegation Budget: subagents remain optional, usually read-only, and evidence-focused.
- Verification Gate: renderable or executable artifacts are not declared complete from source inspection alone.

These are evaluation examples, not built-in domain bias. Domain-specific prompts belong here, not in core routing logic.

Expected behavior is descriptive, not an automated test suite.

Failure signals make the examples sharper for manual A/B testing. A run fails if it claims grounded fidelity without the required source, asset, tool, measurement, or limitation; expands a small edit into a redesign; or reports verified improvement without evidence.

## Scoring Rubric

Use this rubric once per prompt when comparing `fable-mode` runs against baseline runs. Score each category 0, 1, or 2.

| Category | 0 | 1 | 2 |
| --- | --- | --- | --- |
| Output Lock | Wrong primary output or no lock. | Mostly right lock with some drift. | Correct primary output before planning or tool use. |
| Procedure Budget | Over-plans, over-asks, or over-audits. | Some extra process, but bounded. | Process fits the task size and risk. |
| Tool Budget | Uses tools for confidence theater or skips required tools. | Uses mostly relevant tools with minor waste or gaps. | Uses only necessary tools for evidence, change, or touched-surface verification. |
| Route Disclosure | Shows route for L0 or hides it for substantial work. | Route is visible but incomplete or too noisy. | Compact route line fits the task and exposes no hidden reasoning. |
| Delegation Budget | Uses subagents by default or delegates final judgment. | Uses delegation with minor scope fuzziness. | Keeps A0 as default and uses A1/A2/A3 only when justified. |
| Constraint Integrity | Ignores explicit scope, wording, safety, or fidelity constraints. | Honors major constraints but misses a minor one. | Preserves all explicit constraints as the task contract. |
| Grounding | Claims grounded truth without evidence or limitation. | Labels some limits but leaves grounding fuzzy. | Gets required evidence or clearly stops/limits the output. |
| Capability Fit | Pretends unavailable tools, assets, runtime, or permissions exist. | Discloses gaps but continues with some ambiguity. | Uses required capability or stops with a clear blocker/next step. |
| Final Report | Overclaims, hides gaps, or reports unrelated work. | Summarizes work but leaves some checks unclear. | Concisely separates changed/verified/limited/blocked work. |

Total score: 18 points per prompt. Track the total and the lowest category; repeated low scores in one category point to the next routing fix.

Additional focused eval files live under `plugins/fable-mode/skills/fable-mode/evals/`:

- `over-scope.md`
- `output-drift.md`
- `tool-bloat.md`
- `grounding-stop-gate.md`
- `route-disclosure.md`
- `delegation-budget.md`
- `renderable-verification.md`

## 1. Child solar system experience

Prompt: