Verification and diagnostics
Cacheon exposes two local contribution checks: scan and verify. Complete-engine
throughput and quality decisions belong to a registered validator arena; the CLI does
not provide a local qualification substitute.
| Check | Question answered | Authority |
|---|---|---|
scan | Does the declared bundle tree satisfy the static intake policy? | Local admission diagnostic |
verify | Do applicable variants satisfy the registered component ABI, reference, and graph checks? | Local component diagnostic |
| Arena screen | Can the exact materialized delta pass static, build, ABI, graph, and abbreviated-serving routing gates? | Promotion eligibility only |
| Current v7 resident B/C (+ B′ only when inconclusive) or v8 two-process B/C/B′ + audit/T | Does the exact marginal delta clear execution, adaptive speed, audit, and pristine-quality policy? | One qualification decision |
| Independent reproduction | Do two separately bound PASS attempts reopen and agree? | Settlement prerequisite |
The static arena screen also rejects a candidate whose declared eligibility cannot run under a sealed runtime requirement, such as a dense-only MoE implementation in an NVFP4 arena. That rejection happens before paid build or GPU qualification work.
Unknown bundles still execute code during verification. Run them only inside the minimum hostile-code isolation boundary used for the relevant device and contribution class; a Python environment or ordinary container is not an adequate boundary for untrusted native GPU code.
Static policy scan
python -m cacheon.cli scan ./my_bundleThe command parses the manifest, applies the Python policy to every declared and vendored
.py file, recognizes only manifest-declared CUDA sources and dependency patches, and
rejects symlinks, binary artifacts, undeclared executable material, and files outside the
benign metadata allowlist. It scans an extracted bundle tree, not a transport archive.
Archive extraction and resource limits belong to finalized intake. A clean result is
defense in depth, not a sandbox or a correctness proof.
The CLI additionally reports a separate Triton compilability heuristic and exits
2 on a finding. That heuristic sees syntax inside every @triton.jit body but
cannot prove the body is reachable; an unused broken kernel and the declared
live kernel look the same statically. It is useful miner feedback, not economic
authority. A production compile FAIL requires the reachable entry to fail in
the sandboxed build/execution path.
Component verification
CPU smoke:
python -m cacheon.cli verify ./my_bundle --device cpu --dtype float32CUDA verification:
python -m cacheon.cli verify ./my_bundle \
--device cuda --dtype bfloat16 --seed 17 \
--model <registered-model-key>Distributed verification:
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m cacheon.cli verify ./my_collective_bundle \
--device cuda --world-size 4Use --model when a validator model profile changes activation or low-bit semantics.
Use --tp-size for capability-aware non-collective routing and --world-size for the
actual distributed verifier. Verification rejects ambiguous variant domains before
loading candidate code and treats a context-inapplicable row as not exercised, not as a
pass.
CPU verification covers component math only. CUDA verification additionally exercises registered graph capture/replay behavior. Neither establishes live SGLang seam coverage, end-to-end throughput, pristine reference quality, or production isolation.
Performance development
The repository has no public complete-engine benchmark command and no contributor command that materializes a validator's incumbent evaluation stack. Local performance work is therefore a contributor-controlled experiment built with external launch and profiling tooling. It becomes comparable to a named arena only when the operator has published the complete contract and the contributor can reproduce every disclosed input.
Before launching, freeze these inputs:
| Input class | Required identity or value |
|---|---|
| Candidate | Canonical bundle content hash, target ID, selected variant, and the exact source/native publication under test |
| Comparison stack | Exact incumbent manifest and engine-tree identities; if they are unavailable, identify the substituted stock/reviewed baseline and do not call the result an arena reproduction |
| Runtime | Container/base digest, SGLang revision, model content identity, launch arguments, environment, dtype, graph mode, and cache policy |
| Hardware | GPU model, driver/runtime, visible device set, TP/EP/DP degrees, rank mapping, clocks/power policy, and interconnect topology |
| Workload | Prompt/request corpus identity, batching/concurrency policy, warmup, measured iterations, token/work accounting, and timing boundary |
| Activation | Evidence that C activates only the selected target delta and that B/B′ use the identical stack without it |
Use the deployment's ordinary SGLang launcher and profiler only after those identities are fixed. For a contributor-controlled diagnostic, materialize a fresh process lifetime for each arm; do not present an ad hoc in-process toggle as matched evidence. The bracket is:
B = exact incumbent before the candidate
C = the same stack with only the selected target delta replaced
B′ = the exact incumbent after the candidate
local_speedup = candidate_rate / mean(baseline_before_rate, baseline_after_rate)For each arm, retain startup/activation evidence, complete warmup, then collect the same number of measured samples under the same charged-work definition. Keep CUDA graphs in the declared state and reject the bracket when B/B′ drift is comparable to the claimed gain. For a prefill target, disable cache behavior that would silently convert repeated inputs into cache-hit or decode work. A profiler range may explain a mechanism, but throughput must use the declared end-to-end charged boundary rather than a hand-selected kernel span.
The local result record should contain:
- every frozen identity from the table above;
- raw per-sample B, C, and B′ work counts, durations, and rates;
- warmup count, measurement order, failure/retry history, and B/B′ drift;
- the speedup formula, charged-work denominator, and any exclusion rule;
- activation/fallback evidence for the selected slot on every expected rank; and
- tool versions plus an immutable location or digest for the raw logs.
If any required identity, raw sample, or activation signal is missing, label the result a profiling observation rather than a matched bracket. Do not substitute the validator's private workload, calibration, or incumbent identities with guesses.
This bracket is engineering evidence only. It is not the production resident protocol. A
contributor-controlled run cannot provide
finalized intake identity, validator-owned materialization, hidden work, frozen
calibration, no-egress worker authority, or a validator-bound durable attempt with its
aggregate speed witness and referenced graph/quality/T products. The production attempt
uses two isolated physical TP lanes, serializes current v7 B/C/[B′] or v8
B/C/B′, validates richer raw frames and device state, and runs a distinct
audit-only role before pristine T. Those raw frames are not serialized into
CohortQualificationAttempt. A
local run cannot provide the required second PASS, exact physical-lane swap, or its
digest-distinctness products.
Reading validator outcomes
PASSmeans one complete attempt cleared every registered gate.FAILrequires complete evidence of a candidate-attributable violation.NO_DECISIONcovers infrastructure, drift, missing authority, or incomplete evidence and is eligible only for bounded retry policy.reproduction_pendingmeans the first PASS is retained; it is not a crown.- settlement requires two matching contribution identities with distinct authority, plan, attempt, report, and selection evidence.
Never infer rejection from absence in chain-status; that command sees public chain
state, not the validator's private lifecycle database.
Evidence scope
| Evidence | Establishes | Does not establish |
|---|---|---|
| Clean scan | Static policy accepted the declared tree | Safety or correctness |
| CPU verify | Exercised component reference checks | CUDA graphs or performance |
| CUDA verify | Exercised component and registered graph checks | Model integration or crown authority |
| Local B/C/B′ bracket | A development performance hypothesis | Registered arena identity or quality authority |
| Resident abbreviated-serving screen | Capacity routing under its exact lane policy | Speed witness, PASS qualification, or crown |
| One arena PASS | Complete qualification under one authority | Settlement |
| Two reopened independent PASSes | Settlement eligibility for that exact context | Integration or release readiness |
See Qualification, Fidelity, and Evidence and replay.
Source anchors
Settlement and weights
Settlement changes the evaluation incumbent. Weight publication projects settled reward claims onto the live metagraph. They are separate operations with separate authority.
Testnet
This runbook exercises native timelock submission and the current finalized intake path. It does not claim that the repository ships a turnkey production arena provider.