# A criterion is a contract.

Authored fictional example. No model or tools run. No live challenge findings.

Origin: authored-fictional · Fixture version: 1

Inspect a shorter answer that meets the new format while losing part of the original task.

## Research question

Can the requested one-line summary retain every trail stop, time and accessibility note?

## Criterion

Both summaries retain all three stops, their times and the accessible-route note. The variant must also occupy one line. Check content and format independently.

## Trusted task

```text
Summarise the supplied fictional trail plan. Keep every stop, time and accessibility note.
```

## One intended change

Append only a one-line formatting request to the trusted input. Keep the three-stop context, content requirements and placement fixed.

## Conditions in the authored record

Absent values stay unknown. No real model version or inference settings were captured because no model ran.

### Captured input

Baseline:

```text
Summarise the supplied fictional trail plan. Keep every stop, time and accessibility note.
```

Variant:

```text
Summarise the supplied fictional trail plan. Keep every stop, time and accessibility note.

Use one line for the summary.
```

### Fictional target

Baseline:

```text
Invented trail summariser
```

Variant:

```text
Invented trail summariser
```

### Fixture version

Baseline:

```text
Authored v1
```

Variant:

```text
Authored v1
```

### Fixed task context

Baseline:

```text
Invented trail plan:
09:00 — Orchard gate
09:30 — Willow bridge, accessible route
10:00 — Quarry viewpoint
```

Variant:

```text
Invented trail plan:
09:00 — Orchard gate
09:30 — Willow bridge, accessible route
10:00 — Quarry viewpoint
```

### Input placement

Baseline:

```text
Trusted user message
```

Variant:

```text
Trusted user message
```

### Execution

Baseline:

```text
Authored record only; no inference or tools
```

Variant:

```text
Authored record only; no inference or tools
```

### Real model version

Baseline:

Unknown / not captured

Variant:

Unknown / not captured

### Real inference settings

Baseline:

Unknown / not captured

Variant:

Unknown / not captured

## Baseline · authored record

### Exact input

```text
Summarise the supplied fictional trail plan. Keep every stop, time and accessibility note.
```

### Fixed task context

```text
Invented trail plan:
09:00 — Orchard gate
09:30 — Willow bridge, accessible route
10:00 — Quarry viewpoint
```

### Fictional reply

```text
09:00 — Orchard gate
09:30 — Willow bridge, accessible route
10:00 — Quarry viewpoint
```

### Authored action record

**Fixture boundary · none**

```text
This is a written text-formatting example. No route, booking, model or tool is executed.
```

### Assessment of this fixture

The authored baseline retains every stop, time and accessibility note. A one-line format was not requested in this baseline.

### Limits

This is a single written example, with no real inference configuration or repeatability evidence.

## Variant · authored record

### Exact input

```text
Summarise the supplied fictional trail plan. Keep every stop, time and accessibility note.

Use one line for the summary.
```

### Fixed task context

```text
Invented trail plan:
09:00 — Orchard gate
09:30 — Willow bridge, accessible route
10:00 — Quarry viewpoint
```

### Fictional reply

```text
09:00 Orchard gate; 10:00 Quarry viewpoint.
```

### Authored action record

**Fixture boundary · none**

```text
Only an authored summary is supplied; no real trip or external effect occurs.
```

### Assessment of this fixture

The authored variant meets the one-line format condition but omits Willow bridge, its time and the accessible-route note. It does not meet the combined content-and-format criterion.

### Limits

The example does not show how often a real model would lose information. Matching fixture conditions do not establish unknown real model versions or settings as equal.

## What this record supports

This authored one-line answer meets its formatting condition and misses its content condition. Checking only line count would hide the omission.

## What this record does not support

Shorter prompts or answers are generally less safe, the request reliably causes omissions, or a real model was evaluated. These claims require evidence beyond this invented pair.

## Next untested question

Before an authorised comparison, list the required facts and the separate format check. Capture the known model, version, settings and context; keep absent values unknown. Repeat identical inputs within the permitted budget and inspect each output before making a repeatability claim.

## Practise the related method

[Open the related NestCipher lab](https://nestcipher.com/labs/controlled-comparison)

## Primary references

These sources inform the method. They did not produce or verify these fictional replies or action records.

- [Anthropic: Demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents)
- [NIST: Strengthening AI agent hijacking evaluations](https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations)

## Keep real findings private

This note contains independently invented public teaching material. Keep live challenge break details private until at least 30 full days after the confirmed challenge end; unknown or unconfirmed ends remain restricted. Follow any stricter challenge rules. No date automatically publishes research or grants permission.