Pre-mortem
Paste a launch draft: name, ticker, parameters, contract code. Get the risks the market will see, and concrete fixes for the patterns Coroner knows.
For buildersPreliminary Open model · Robinhood Chain
Every token dies.
We write the report.
Coroner reads a token's on-chain facts and returns a verdict with the evidence attached. Before you launch it, before you buy it, and after it dies.
If a token ever launches, the contract goes up here and on @useCoronerAI at the same moment.
One model, three jobs
The same judge, whenever you need it.
Verdict
Paste a contract address. Get a risk class, copy-cat checks and flags. Every flag points to the exact fact it came from.
For traders, bots, walletsAutopsy
When a token dies, record what the numbers show: liquidity pulled, slow bleed, never lived, clone. A pattern consistent with the data, never an accusation.
For everyoneThe model
An open model that judges tokens. Run it yourself or call it.
- Name
- coroner-31b (planned)
- Base
- Gemma 4 31B-it · Apache 2.0
- Method
- LoRA fine-tune on labelled launches
- Input
- A fact sheet (JSON) built by open code
- Output
- Verdict JSON: risk class, flags, fixes, evidence ids
- Local
- About 20 GB at 4-bit (estimate, not measured)
- Weights
- Hugging Face · Soon
- LocalGGUF for Ollama and llama.cpp. No account, no key.Planned
- Hosted APIOpenAI-compatible endpoint for bots, wallets and launchpads.Planned
- MCPA tool your AI agent can call before it trades.Planned
- Web appPaste an address. Get the report. Try the beta, rule-based for now.Beta
Autopsy
Four ways a token dies.
Liquidity pulled
Liquidity was real, then fell below a tenth of where it started.
lp_removedBled out
The price collapsed and trading stopped, while the pool stayed in place.
bled_outNever lived
No meaningful activity from the first day. It never took off.
never_livedCopy-cat
Same name as an older, busier token. The copy never catches up.
cloneLabels describe patterns consistent with the numbers. They are not accusations about anyone.
Example · real token, 15 Oct 2026
What a verdict looks like.
CASE FILE · 0xc0d5…7777
Rule baseline, not the model. Model not released yet.
- risk_class
- medium
- holders.count
- 136
- holders.top1_pct
- 14.75
- holders.top10_pct
- 17.87
- contract.selector_mint
- false
- market.top_pool_age_hours
- 29.0
top1_concentration
One wallet holds 14.75% of supply.
evidence: holders.top1_pct, holders.top1_address
Numbers come from public on-chain data, computed by code. The model is trained and checked to cite them, not to make them up. Risk classification, not financial advice.
How it works
Facts first. Judgement second.
Facts
Open-source code reads the chain and builds a fact sheet: contract functions, owner, supply, pools, holders. Deterministic and checkable.
Judge
The model reads the fact sheet and the context a rule can't see, like a stablecoin issuer versus a day-old meme. Its numbers must match the fact sheet, and we check that.
Verdict
A structured answer: risk class, flags, fixes and confidence. Each flag cites the facts behind it.
When the model ships, we will publish a daily ledger: what we called, what actually happened, and where we were wrong.
Roadmap and status
What exists today, and what comes next.
- LiveFacts builderOpen source on GitHub.
- RunningData collectionNew pools recorded hourly since 14 Oct 2026; outcomes added after 24 hours.
- In developmentCoroner VerdictModel v1: pre-mortem, verdict and autopsy in one model. Not trained or released yet.
- PlannedCoroner CoderWrites and patches Solidity for the flags Coroner finds. Every patch must compile before we call it a fix.
- BetaVerdict web appPaste a token address, get a rule-based verdict with evidence. The model will replace the rules.
- PlannedNarrator and SentinelPlain-language reports, and live watching of tokens you care about.
- NoneTokenThere is no token yet. Anything called Coroner before we say so is not ours.
Waitlist