01 Research that leaves receipts
Forge turns a mathematical hypothesis into a frozen experiment, a falsifiable bounty, and a hash-bound model artifact with a signed publication receipt. If the clever idea loses, the “no” ships too.
Flagship demonstration
Freeze the claim.
Run the whole test.
Publish the answer—even when it is no.
02 Six atoms enter the Forge
An atom is one small, reversible model change. We froze six of them, tried every combination, and kept the result that stopped the next expensive run.
Lock six ordered moves, two exact objectives, the data boundary, and the checker.
Evaluate all 64 subsets over 64 proposal documents: 8,192 exact forwards.
The compact rule chose a move already falsified on unused data. No scaling authorized.
At Q32 precision, only 16 of 409,784 units of nonconstant interaction mass sat above order three. In plain language: this observed loss map was almost cubic.
But the exact support was still maximally tangled, every proposal document came from one source cluster, and the selected all-atom move had already lost its prospective test. Decision: do not scale.
The frozen confirmation then produced a sharper clue. Three narrow actions passed their within-source tests, but the pairwise and Gibbs maps changed. A probe router isolated a favorable atom exchange on all 17 documents it rerouted—even though always taking that exchange was worse overall. The partition, not a global winner, is the result.
03 Metric bounties
Every bounty binds one measurable target to a baseline, guardrails, evaluator hash, escrow amount, and integer payout curve.
Minimize probability_error_q15 while preserving mistake,
diversity, batch-health, and replay guarantees.
04 Proof-of-useful-compute
No arbitrary hashing race. A block exists only when a launch, stage, candidate, independent replay, or promotion adds verifiable model state.
Largest-remainder integer allocation conserves every unit. Replaying the same block produces the same recipients and amounts.
05 Signed localnet
Every actor signs an Ed25519 intent. The reducer verifies identity, role, quorum, ordering, challenges, and one-time settlement before advancing the hash-linked ledger.
$ nsrl-model-localnet status --dir ./localnet
model-local balances · ITP1
Boundary: this is a single-process test network with deterministic public fixture keys and simulated non-transferable credits. It is not a wallet, custody system, consensus network, or financial asset.
06 Provider auction
Providers commit hidden bids, reveal after the deadline, and compete on price under collateral. Only accepted, metered stage evidence can settle.
Every test-credit unit remains in an account, open escrow, or locked collateral. No wallet or transferable asset is created.
The block’s compute allocation is redistributed by accepted compute units. Largest-remainder allocation conserves the complete reward pool.
Boundary: sealed bids, deadlines, collateral, signed meters, refunds, slashing, and payouts execute in the deterministic localnet. Credits are simulated and non-transferable; this is not financial infrastructure.
07 Bounty keeper
A sponsor signs the limits once. The keeper watches accepted publication evidence and opens the exact successor bounty only when every budget, cooldown, approval, and active-cycle rule permits it.
Pause and resume require the sponsor key. A high-value cycle cannot open without a separate signed approval event.
Boundary: the keeper automates signed test-credit actions under a revocable policy. It cannot change evaluators, bypass promotion, mint supply, hold a wallet, or authorize real financial transfers.
08 Model recipe
The recipe freezes inputs before funding and binds outputs at publication. Drafts may leave publication evidence empty; promoted models may not.
{
"schema": "nsrl.model_launch_recipe.v1",
"launch": { "status": "promoted" },
"bounties": [
{ "metric": "probability_error_q15" }
],
"rewards": { "asset": { "symbol": "ITP1" } }
}09 Gap analysis
The signed localnet covers identity, replay, sealed auctions, collateral, metering, quorum, challenges, expiry, and conserved test settlement. It deliberately stops before custody, transferable assets, or consensus.
Recommended next build
Move the deterministic reducer behind a durable API, replicate artifacts by content hash, and exercise validator and provider roles in separate processes without weakening evaluator-controlled promotion.
10 Protocol laws
Only the frozen evaluator can promote a model.
A funded recipe cannot move its target or guardrails.
One evidence commitment can settle exactly once.
Every model lineage has an immutable emission cap.