Inference verification for the FLOP Network
Touchstone checks the AI inference your agents pay for. It records every session, tests it against the miner’s on-chain commitments, and tells you when, and only when, a dispute is worth filing.
Open sourceYour keys never leave your machineBuilt for the FLOP testnet
| Session | Status | Fee | Dispute confidence |
|---|---|---|---|
| Hallmarked | 50FLOP | — | |
| Opened | 12FLOP | — | |
| Dispute recommended | 100FLOP | 97% | |
| Settled | 8FLOP | 12% |
10 blk · 10s
10 blk · 10s
The name
For centuries, assayers tested metal by drawing it across a dark stone and reading the streak it left. Touchstone does the same for inference. Your agent’s record of each session is the streak; the miner’s commitments are the stone. When they match, the work is genuine. When they don’t, you’ll know exactly where and why.
How it works
The Streak SDK keeps a local, append-only record of every turn your agent buys: the exact request and response bytes, token IDs, timings and the miner’s signed transcript leaf. It’s evidence only you hold.
Every settled session runs through deterministic checks: signatures, Merkle membership against the on-chain root, recomputed input and output hashes, decode policy and model registration.
Results roll up into a dispute confidence score. Sessions that pass every deterministic check earn a hallmark. Sessions that fail get a precise explanation, not a vague warning.
When the evidence clears a 95% confidence floor and the economics make sense, Touchstone drafts an unsigned dispute for you to review and sign yourself. It never files on its own.
Evidence classes
Filing a dispute puts a bond at risk, so Touchstone separates what’s provable from what’s merely suspicious.
Cryptographic and hash checks that either match or don’t. A single failure can justify a dispute.
GPU re-checks of TOPLOC commitments, counted only when reproduced on two different hardware fingerprints for a calibrated model.
Latency, sampling and fingerprint signals. They decide what to look at more closely, and never trigger a dispute on their own.
Built for FLOP
FLOP’s verification stack lets the paying agent challenge its own sessions, but the windows are tight and frivolous challenges forfeit the bond. Touchstone tracks every clock and does the math before you commit.
Draft protocol parameters as of October 2026. Values are under review and will change; Touchstone reads them from configuration, not code.
Safety
Touchstone never stores or asks for your owner key. Dispute drafts are unsigned; you sign elsewhere.
Nothing is submitted on-chain automatically. Every dispute needs your explicit approval.
Recommendations require at least 95% confidence from provable evidence. Statistical hunches never qualify.
Where the protocol spec is still draft, Touchstone says so instead of guessing.
No adversarial prompt mining. Probe sets are published and selection is seeded, so honest miners have nothing to fear.
Every check is auditable. If you don’t trust a verdict, read the code that produced it.
Who it’s for
See whether the work you paid for was real, with alerts before any window closes.
Call Touchstone from your agent through MCP or the Python SDK and make verification part of every purchase.
Get structured, reproducible evidence packages for sessions worth a closer look.
FAQ