# Funded audit integration evidence — September 11, 2026

These bounded experiments exercise NeuroShard-native settlement, funded full
replay, model evolution, data evaluation, generation and recovery. All accounts
and both machines have one owner. The experimental chains have been stopped
with state preserved. Their allocations are separate from the public 0.4.0
network; the recorded candidate peer is not a permanent service endpoint.

Read the [results and limitations](https://github.com/neuroshard-ai/neuroshard/blob/main/docs/FUNDED_AUDIT_RESULTS.md)
and the [candidate operations guide](https://github.com/neuroshard-ai/neuroshard/blob/main/docs/CANDIDATE_OPERATIONS.md).
The trial scripts were frozen at commit `3710226`. Both genesis files commit
package hash `2875b6e2de443c136cfffe3288b35d03a3f763ee521ce4f231d6704043e1a125`.
Later script changes and distribution checks are identified separately.

- `synthetic.json` and `real.json`: final lifecycle results, audit accounting,
  quality decisions, quorum failure checks and common-block application hashes.
- `*-accounting.json`: supply invariants checked independently against all four
  retained validator databases. Final snapshot heights may differ; common-block
  agreement is a separate measurement.
- `*-genesis.json`: the public experimental genesis bytes. Public validator keys
  are identities; no private signing material is included.
- `*-audit-reports.json`: exported local replay records, without commit salts.
  Their `object_bytes` include cached object reads and are not measured wire
  bandwidth. Timings include object handling and shared-host contention.
- `auditor-crash-recovery.json`: an interrupted fraud upload retained exactly
  the same signed bytes and nonce after the process was killed and restarted.
- `real-fraud-cost.json`: unique signed dispute-envelope sizes and consensus
  timestamp span, including that crash. This excludes retransmission, RPC
  framing and peer gossip; it is not a client-latency or total-bandwidth measure.
- `follower-*.json`: cross-host replay, retained-identity restart, public peer
  connectivity, and reconnection after the driver outage. A follower is not a
  fifth validator or a second independent owner.
- `hardware.json`, `script-provenance.json` and `distribution.json`: environment
  and source/artifact identities. No new PyPI or public runtime release is implied.

The real trial exceeded the frozen driver's one-hour second-cohort wait.
`real.json` records its explicit retained-state continuation, including the
resume height and helper digest. Existing genesis, keys, signing state, worker
results and transaction outboxes were reused. No new training or genesis reset
was used to finish its remaining evaluation. The follower needed a manual
restart to reconnect promptly after the long outage. This is not a claim of
unattended end-to-end recovery from every failure.

These JSON records are inspectable experiment evidence, not a cryptographic
proof that independent operators performed the measurements. Signed reports
alone do not prove computation; colluding auditors can attest to a false claim.
The model quality decisions apply only to the disclosed fixed public cohorts.

`SHA256SUMS` lists every published file except itself. Download the files named
there into one directory and run `sha256sum -c SHA256SUMS`. Matching checksums
verify consistency with this manifest, not the truth of the experiment. Preserve
the commit and genesis identities when reproducing the source procedure; never
initialize a new chain over an existing validator home.