Inventory and compare
Scan the synthetic source and compare each inventoried artifact with the declared target capabilities.
Proof of method · Synthetic demonstration
Can an AI-agent move be checked, not just copied?
This early Veld prototype makes a narrow part of an agent-framework move reviewable: inventory what exists, show what can and cannot move deterministically, rehearse changes in isolation, and check one declared target behavior.
What this page proves
This page reports one checked-in synthetic Hermes-to-OpenClaw demonstration. It uses no customer system, production profile, live account, or credential value. It is not evidence of customer use, broad framework compatibility, production readiness, security certification, or complete behavioral equivalence.
The problem being tested
An agent profile can contain instructions, skills, schedules, memories, dependencies, model choices, plugins, channels, and tool-server configuration. In the synthetic fixture, those artifacts do not all transfer the same way: some match exactly, some require translation, some degrade, some require new authorization, and one is unsupported.
A successful startup would not, by itself, show which differences remained or whether a specific translated setting still matched its declared requirement.
The method
Plan, isolate, and verify before calling the move complete.
Plan before applying
Classify every artifact as copy, translate, manual, reauthorize, or skip. Keep unresolved work visible rather than treating it as success.
Apply only in isolation
Apply deterministic actions to a new, empty, disposable target. Do not modify the source profile.
Verify a declared behavior
Check one machine-readable acceptance condition and bind the plan, contract, and receipt with checksums.
What the synthetic run showed
The fixture contained nine artifacts. The plan accounted for all nine:
- 2 copied;
- 2 translated;
- 2 left for manual work;
- 2 requiring reauthorization; and
- 1 unsupported item skipped.
The baseline run passed one declared schedule check. Its receipt reported partial, not complete, because five actions remained unresolved. The source checksum remained unchanged, and the plan accounted for 9 of 9 artifacts.
The negative control changed the translated schedule from 0 9 * * * to 0 7 * * *. The same acceptance contract rejected that target and reported the exact expected and actual values.
What “verified” means here
“Verified” means only that the declared deterministic work, integrity checks, artifact accounting, and one acceptance condition passed at receipt issuance.
It does not mean that every behavior was tested, unresolved work is safe to ignore, credentials were moved, a production cutover is ready, or the target will remain unchanged after the receipt is issued.
Safety boundaries
- This public demonstration uses synthetic inputs only.
- No live credentials, private keys, tokens, production profiles, customer records, or regulated data are requested or used.
- Credential references can be classified, but credential values are not transferred; authorization remains a separate operator action.
- Deterministic actions are limited to a new, empty, isolated target.
- Binary and non-UTF-8 artifacts remain manual.
- Likely-secret detection is heuristic and conservative. It reduces accidental exposure but is not a security audit or proof that arbitrary prose contains no secret.
- Hermes and OpenClaw identify the synthetic fixture formats only. Veld is not affiliated with or endorsed by either project.
- This page is informational and is not a proposal, price, service-level commitment, or promise that a pilot is available.
Help test the question
Have you moved an agent workload?
Veld is interviewing engineering, platform, and AI-operations owners about what broke, what remained manual, and what evidence would have improved the decision. No production access, credentials, customer data, purchase, or endorsement is requested.