DOOMCON：可独立复算的实时 AI 活动指数

Superintelligence Real-time Early Notice. Formerly DOOMCON; the repository and the web address keep the old name.

Live: https://messagegabrielhere-lgtm.github.io/doomcon/ · verify the readings in your browser · method

A real-time index of AI activity tempo, computed from public data by open code.

Five pillars, one 0–100 score, one level from DOOMCON 5 (DORMANT) to DOOMCON 1 (UNPRECEDENTED). Every number is recomputable by a stranger.

We don't know anything. We just count.

What this is

Every other live AI-risk index — DoomBench, IMD's AI Safety Clock, skynetcountdown, takeofftracker, pdoom.ai — scores by human or model judgement. None of them can be independently recomputed.

This one can. Clone the repo, run two commands, compare your number to the published one. If they disagree, that's a bug.

It measures observable activity tempo relative to the field's own history — not probability of harm, not a forecast. The full formula, every constant, and an honest "how this index could mislead you" section are in docs/METHODOLOGY.md.

Run it

No Node installation required — everything runs in Docker:

docker run --rm --network host -v "$PWD":/app -w /app node:20-alpine node collector/collect.mjs

Step	Command	Writes
Build the frozen reference	`node collector/backfill.mjs`	`data/reference.json`
Collect live readings	`node collector/collect.mjs`	`data/raw/<iso>.json`
Score it	`node collector/engine.mjs`	`data/state.json`, `data/receipts/`, `data/history.ndjson`
Render share cards	`node collector/card.mjs`	`public/cards/`
Generate post text	`node collector/posts.mjs`	`data/posts.json`
Build the site	`node site/build.mjs`	`public/`
Build the posting console	`node site/post-sheet.mjs`	`public/post-sheet.html`

Prefix each with the Docker wrapper above. Backfill runs once — the reference distribution is frozen on purpose and must not be regenerated casually.

Architecture

collector/fetch.mjs        shared HTTP: timeout, retry, User-Agent, structured errors
collector/sources/*.mjs    one adapter per source — drop a file in, it is discovered
collector/collect.mjs      runs every adapter concurrently -> data/raw/
collector/engine.mjs       normalise -> smooth -> compose -> level, with anti-flap
collector/receipts.mjs     hash-chained, append-only audit trail
collector/card.mjs         share cards as dependency-free SVG
collector/posts.mjs        post text, with URL and future-tense guards
site/build.mjs             static site generator -> public/

Zero npm dependencies. Node 20 built-ins only. There is no node_modules, no lockfile, and no supply chain.

Adding a source

Drop a file in collector/sources/. It is discovered automatically.

export default {
  id: 'my-source',
  pillar: 'capability',          // capability | compute | attention | governance | markets
  label: 'Human-readable name',
  async collect(fetchJson) {
    const data = await fetchJson('https://…')
    return { value: 42, unit: 'things/day', meta: { … } }
  },
}

Three rules: value is a finite number, higher always means more activity (negate inside the adapter if your source points the other way), and throw on failure — never return a fallback. A dark source is a first-class state.

A new source does not enter the index until it is added to the frozen reference distribution, so adding one cannot retroactively move history.

AI battle

/arena.html is a separate side project: Claude, GPT, Grok, Gemini and DeepSeek each trade $1,000 of paper money against live Coinbase order books, alongside two no-AI baselines. The models only propose. A server prices each proposal into a ticket and checks it against fixed rules before filling it, and stops run every 5 minutes whether or not a model is awake. Every fill and every rejection is public. It shares nothing with the index pipeline. See docs/ARENA.md.

Data and API

Path	What
`/api/state.json`	Current score, level, pillars, per-source freshness
`/api/history.json`	Scored observations over time
`/api/receipts/<id>.json`	One hash-chained receipt
`/embed.html`	Embeddable widget
`/feed.xml`	RSS of index moves

All CC-BY 4.0. Attribution appreciated, not enforced.

Verifying a receipt

Every scored observation carries every input exactly as fetched, plus the hash of the previous receipt. verifyChain() in collector/receipts.mjs walks the chain and re-derives every hash.

Note that arXiv counts are not immutable — papers get cross-listed after publication — so we snapshot the value and the query string at fetch time and never recompute history. A receipt records what we saw, not what the source says today.

Credits

Historical training-compute and notable-model data from Epoch AI, CC-BY 4.0.

Concept debt to pizzint.watch, whose Pentagon Pizza Index proved that a single named scalar plus a distribution channel beats a better dashboard with neither. The full teardown is in docs/TEARDOWN.md.

Licence

Code MIT. Data and index values CC-BY 4.0.