# The Slop Index: Methodology v1

_A benchmark for the specific failure patterns of language-model prose. Built by Pure Light Ltd._

## The numbers

Clotho v1 and the base model it was trained from, scored by the same detector,
the same 20-prompt suite, the same day (11 September 2026):

| Metric | Clotho v1 | Base (Mistral-7B) |
|---|---|---|
| Mean slop score | **0.0** | 0.6 |
| Max slop score | **0** | 2 |
| Clean generations (score 0) | **20/20** | 14/20 |
| Triple/n-gram repetition | **0** | 1 occurrence (gate: FAIL) |
| Summary endings | **0** (bar: <=1/20) | 1 |

Collection conditions: temperature 0.9, top_p 0.95, single-shot, 512 new tokens,
identical seeds. The detector is our in-house implementation; its source and the
prompt bank are proprietary. This page is the publication of record.

## One look at the data

Same prompt to both models:

> Show the typing pool emptying at five on a Friday.

Base model:

> I'm an AI language model and don't have the ability to display images or
> videos. However, I can describe the scene of an emptying typing pool...

Clotho v1:

> At 4.57 the typewriter room was filled with women sitting tensely at their
> machines, poised for the last burst of work. The stenographers were to finish
> at five precisely...

No preface, no summary ending, straight into the scene.

## Long context, measured

A 32,768-token context is a claim until it is tested. Ours was, end to end:

| Check | Result |
|---|---|
| Needle retrieval at 10k / 20k / 28k tokens (early, mid, late placement) | **9/9 verbatim** |
| Instruction held at depth (3/3 at each depth) | **9/9** |
| Style suites at depth: slop score | **0** |
| Warm retrieval latency | **2.5 to 7.6 s** |

## Serving audit

Contract tests against the live API, all serving paths, streaming and
non-streaming: **23/23** audit suite, **22/22** user-simulation calls across
**11/11** scenarios, **0** errors, context correctly reported as 32,768 on the
public model endpoint.

## What the detector weighs

slop_score is the sum of triggered signals per generation (0 = clean):

| Signal | Weight |
|---|---|
| Sentence-level triple repetition (3 identical consecutive sentences) | 2 |
| Echo finale (closing sentence echoes the opening) | 2 |
| Summary endings ("And so, with that...") | 2 |
| Cliché bank (shivers-down-spine, "air crackled", "orbs", "mirth", ...) | 2 |
| Negation stacking ("not ... not ... not") | 1 |
| Pseudo-precision in narrative prose | 1 |

An independent MIT-licensed writing-register detector is used separately for
copy-model register validation; it does not score the results above.

## Publication rules

1. Our models are scored under the same protocol as any contender; wins and
   losses publish alike.
2. Sample output ships with every result set on this page.
3. Models are re-scored after material updates; history is kept.

## Honest caveats

- Heuristic, English-language detector: it catches these patterns, not all badness.
- Single-shot 512-token runs under-reveal repetition that emerges over long
  multi-turn sessions; a multi-turn protocol is planned for v2.
- Snapshots decay: cite the collection date with any comparison.
- Public cohort expansion (baseline models scored under the same protocol) is
  planned post-launch.