NodusLab

Terminology

The words this repository refuses to use loosely.

Short, so that it is actually read. Where a term already has a standard meaning in the literature, we use the standard meaning and say so rather than inventing a synonym.

Economic units (inherited from Nodus v1)

Term Meaning
CU Compute Unit. A quantity of computational work. CU-v0 = 100 tokens of a fixed reference workload (../nodus/shared/cu.py).
CC Compute Credit. The transferable accounting unit, integer micro-credits.
CU-v0 The current, workload-anchored definition. Its weakness: model_weight is DECLARED, and token counts are provider-reported.
Verified CU (vCU) New here. A CU for which evidence at a stated level supports a stated claim set. vCU is always qualified: "vCU(A, L3, 2^-40)" means one CU whose correctness is supported by an L3 mechanism with soundness error 2^-40. An unqualified "verified CU" is meaningless.
Cost of verified CU cost(CU) + cost(evidence). The metric the whole programme optimises.

Verification vocabulary

Term Meaning
Claim O/A/B/C/D/E Origin / Correctness / Execution / Work / Physical resource / Useful computation. Defined in verification-taxonomy.md.
L0–L6 Evidence strength ladder, from bare assertion to unconditional.
Proof Reserved for L5/L6. Never used for a random audit.
Check An L2/L3 mechanism. Has a soundness error, which must be quoted.
Attestation A signed statement by a hardware root of trust (L4). Not a proof.
Receipt An authenticated summary of a claimed execution (L1). What v1 has.
Soundness error Probability that a cheating prover is accepted, per check.
R cost of evidence / cost of native computation. The headline metric.
R_prove / R_verify The provider-side and verifier-side components of R. R_verify is the one that decides whether a scheme beats re-execution.
Re-execution baseline Running the job again. R_verify = 1 by definition. Every mechanism is quoted against it.
Model binding Cryptographically tying evidence to specific weights, usually via a commitment. Without it, model substitution is undetectable.
Freshness binding Tying evidence to this execution occasion, usually via an unpredictable value chosen after job assignment. Without it, cached results and precomputation are undetectable.

Evidence labels

Inherited from ../nodus/PAPER_BRIEF.md and used identically here:

  • MEASURED — we ran it and recorded it.
  • DERIVED — computed from measured values; the derivation is shown.
  • DECLARED — a human chose the number. Always flagged.
  • SPECULATIVE — a plausible idea with no evidence.
  • UNKNOWN — we looked and do not know. Preferred over a guess.