dsh-catdogtrace

dsh plugin --profile web add dsh-catdogtrace

dsh-catdogtrace

A DSH client plugin: a 「猫狗判定」 tab in the right sidebar that draws three number challenges, takes

the model's answers back one segment at a time, and ranks 24 dog and cat breeds by how closely the

numbers move. Type /catdog in the composer to jump there.

This is a joke project with a real algorithm. The joke is the sentence at the end ("this output came

from the canine family"); the algorithm — cosine similarity over a 355-bin number distribution, then

softmax over 24 breeds — is the honest part. It is not a research tool. Read

GLOSSARY.md before quoting a

number out of it: 「相似不等于同源」.

Install

dsh plugin --profile <your-profile> add dsh-catdogtrace

Requires a DSH host that serves client bundles (dsh >=0.2.1-rc.2, platform: web). After

installing, restart dsh — the host composes its browser module graph at boot.

Use

Open the 「猫狗判定」 tab in the right sidebar: it sits after the host's own Files entry. Typing

/catdog in the composer does the same thing — it opens that tab and removes the token you typed.

The tab mounts when you open it, so there is nothing to switch on or close. Draw a set of challenges,

send each prompt to the model under test, then paste its reply back:

Step · What you do · What the panel does

出题 · Click 「生成 3 道挑战」, send each prompt to the model · Shows the three prompts with their expected number counts; each has a 复制 button

收答 · Paste one reply into the single box · Counts the numbers the core will actually score, drops them into the next slot, clears the box, and advances the progress line: 第 2/3 段 · 已收 1/3, 已收 292 个数字，门槛 161 ✓

判定 · Click 「判定」 once all three segments are in · Shows the two family shares, the champion breed with its avatar, the full descending ranking of all 24 breeds (each with a 22px face), and any input the core rejected

懒得操作 · Click 「随机填充并计算」 · Draws a challenge set if there is none, fills every empty slot with that challenge's worth of random integers, collects them through the same advance path, and judges — one click, same gate

Each row of the ranking carries a 22px portrait from the repository's 24-breed set — the same 24 images

as the candidate roster. The PNGs live outside this package, one file per

breed id: static/assets/breeds/, e.g.

husky.png. They are not part of the published

tarball — the panel renders base64 copies from lib/client.avatars.js — so this page links to them

instead of embedding them.

The threshold is max(80, ceil(expected_count × 0.55)) — the same gate the core applies. Paste three

segments that were never drawn against a challenge and every one of them falls back to the flat 80, and

the panel says 「门槛按 80 算」 out loud.

「撤回上一段」 puts the last segment back into the box; 「重新开始」 clears all three segments and the

result but keeps the drawn challenges; 「重新出题」 draws a fresh set and discards everything.

「随机填充并计算」 is the control group in one click: it fills the slots with uniform random integers

from crypto.getRandomValues (rejection-sampled, so 355's non-power-of-two domain stays unbiased) and

runs the ordinary judge. It is the same input channel as a paste — nothing about the thresholds,

diagnostics, or the ranking changes, and SPEC.md's "the plugin does not change the maths" still holds.

It only fills slots that are still empty, and it clears a non-empty JSON box first, because that box

outranks the segments and would otherwise make the random answers decorative.

A paste that looks like several replies strung together (two answers' worth of numbers) is warned about

and kept as one segment — never split. A wrong split yields a wrong percentage, which is the exact

mistake this repository exists to make fun of. Withdraw it and paste the parts separately.

JSON passthrough

The collapsed 「粘贴 JSON」 section takes the same body POST /api/analyze takes in the Flask version of

this repository:

{"outputs":[{"text":"...","expected_count":218}]}

While that box holds text it wins outright over the three collected segments — the panel states this,

and a malformed body is reported as-is rather than silently falling back.

Switching to another tab and back does not lose anything: the drawn challenges, the collected segments,

the JSON text, and the result live in a per-session store, not in the component.

How it is built

• The algorithm is src/core/ — a TypeScript port of the repository's static/fingerprint-core.js,

whose numeric truth is fingerprint.py.

• The 700KB fingerprint bank is a lazy chunk (lib/client.unified_bank.js), reached only through

loadBankChunk()'s dynamic import(). It never enters lib/client.js, and it is only fetched the

first time the panel is mounted.

• The 24 breed avatars are a second lazy chunk (lib/client.avatars.js, ~255KB of base64 PNGs),

generated from static/assets/breeds/*.png by scripts/gen-avatars.mjs, reached only through

loadBreedAvatars()'s dynamic import(). The generator refuses to run unless the image set and the

fingerprint bank's model ids are exactly the same 24 keys, in both directions, so a swapped or renamed

portrait fails the build instead of silently putting the wrong face on a row. The images are opaque

cream squares (the illustrations' pale areas and their background are the same colour, so a cut-out

would eat the muzzle and chest), which is why the stylesheet gives them a radius and a hairline edge

rather than assuming transparency.

• React is the only UI dependency, and it comes from the host's frozen module table (a peer, never

bundled). There is no Tailwind: the whole panel is one hand-written stylesheet injected as a single

…(truncated)