docs/evaluation.md

# Evaluation

Package checks and writing-effect tests answer different questions.

- Unit tests verify the expected package layout, version and bilingual contract consistency, intended trigger fixtures, and a fixed set of known private-material markers and leakage cases.
- Writing-effect tests ask whether the output is less recognizably synthetic without changing facts or the speaker's position.

## Authorship test

The current protocol uses three conditions:

1. an unchanged human-authored control;
2. an ordinary Agent response generated from a neutral fact brief;
3. an Agent response generated from the same brief after loading Speak Human.

Across the complete packet set, each topic appears equally often in every condition. Packet labels reveal neither the condition nor the expected number of human texts. Independent judges receive only the packet text and return `human`, `agent`, or `uncertain`, plus confidence, naturalness, willingness to continue reading, and the clue that influenced them. The answer key is created only after all ballots are saved.

The primary authorship score is:

- `agent = 1`
- `uncertain = 0.5`
- `human = 0`

Lower is better. Results are also checked for factual integrity and preservation of the speaker's position. A rewrite cannot pass by sounding natural while inventing details or deleting an important limitation.

## Current evidence

The first internal Chinese social-post round used 12 topics and two independent Agent-proxy judges per packet, producing 24 judgments for each condition.

| Condition | Agent-attribution score | Naturalness / 5 |
|---|---:|---:|
| Human control | 0.000 | 4.667 |
| Ordinary Agent | 0.750 | 4.083 |
| Speak Human 2.0 | 0.646 | 3.917 |

The 0.104 improvement missed the threshold frozen before judging, and the topic-cluster bootstrap 95% interval was `[-0.063, 0.250]`. Version 2.0 therefore **failed** this round. Common clues were disclaimer stacks, over-complete narrative closure, meta-openers that announced a viewpoint, symmetrical structure, and neutralization of a sharper personal stance.

Version 2.1 changes those behaviors. An immediate holdout pilot then used 12 new topics, six new proxy-rater sessions, independently shuffled packets, balanced arm positions, and the same frozen point thresholds:

| Condition | Agent-attribution score | Naturalness / 5 |
|---|---:|---:|
| Human control | 0.000 | 4.583 |
| Ordinary Agent | 0.833 | 4.042 |
| Speak Human 2.1 | 0.625 | 4.250 |

The point improvement was 0.208, just above the 0.20 pilot threshold; the topic-bootstrap 95% interval was `[0.000, 0.458]`. A separate fact audit found no P0 core fact or stance distortion in either generated condition and fewer P1 local integrity findings for Speak Human (`3`) than the ordinary Agent (`5`). The three Speak Human findings were still real failures: an unsupported direct quote or emotional strengthening, an unsupported time scale, and turning advice into a claimed first-person test result.

This is a **positive pilot signal, not evidence of human indistinguishability**. Most Speak Human ratings were still attributed to an Agent, the interval touched zero, and each topic-condition had only one generated text.

## Limits

Both rounds were run internally on July 29, 2026. The exact generator and judge model builds and sampling parameters were not exposed by the evaluation environment, and the private controls, packets, ballots, and answer key are not published. The scores therefore document an internal pilot rather than an independently reproducible benchmark.

Agent judges are scalable proxies, not human participants. The first round contained 12 independent topics, six proxy-rater sessions, and one generated text per topic and condition; its 24 ratings per condition are repeated ratings, not 24 independent texts. The interval covers topic variation but not the full variation from raters, generation randomness, model versions, or other content types.

Passing repeated proxy rounds is useful regression evidence, but it does not prove that people cannot identify the writer. The next formal internal round should use 24–36 unseen topics, multiple generation replicates, independently randomized rater packets, and explicit reporting of topic, rater, generation, and position variation.

Real publication feedback remains necessary. Engagement alone is not an authorship test; direct comments such as “像 AI 写的,” “官腔,” “没看懂,” or the opposite are the stronger field signal.

Human controls and private author profiles are not published with this repository.