Replay levels
Exact replay
Inputs, versions, environment, state, and outputs match within a declared deterministic boundary.
Functionally equivalent replay
The same objective and policy outcome are reached although token-level or timing details differ.
Partial replay
Only selected model, tool, or policy steps can be reconstructed.
Unavailable replay
Required external state, model version, source snapshot, or side-effect history is absent.
Research-grade evidence package
- Request and interpreted objective.
- Runtime profile, model route/version, and prompt-contract version.
- Context source IDs/hashes, memory scopes, tool/schema versions, and policy decisions.
- Approvals, checkpoints, artifacts/hashes, errors, recovery, and unresolved uncertainty.
- Random seed where available, environment/dependency versions, and UTC timestamps.
Evidence should minimize sensitive content while preserving the facts required for review. OpenTelemetry conventions can support correlation but do not replace product-level evidence semantics. Source: OpenTelemetry GenAI conventions
Claim register
| Field | Purpose |
|---|---|
| Claim | Specific, falsifiable statement rather than a broad narrative. |
| Source type | Foundational work, formal proposal, prototype, benchmark, production evidence, or editorial inference. |
| Evidence level | Current position on the evidence ladder. |
| Counterevidence | Known results or interpretations that weaken the claim. |
| Limitations | Scope, assumptions, missing data, and boundary conditions. |
| Status and reviewed date | Active, downgraded, deferred, rejected, or superseded with UTC review date. |
Negative and null results
Publish adaptation that did not improve outcomes, structural edits that increased cost, critic loops that failed, no-op thresholds that were too conservative, and benchmark tasks with inconclusive results. Negative evidence is necessary to distinguish a reusable mechanism from selective reporting.
Open research agenda
- A formal, falsifiable definition of teleodynamic runtime behavior.
- Safe goal revision and resource variables that generalize across workloads.
- Independent evaluators and causal attribution for recovery.
- Operator control under self-maintenance and long-horizon drift.
- Benchmark portability and evidence interoperability.
- Criteria that separate useful metaphor from implemented mechanism.
Source record
References
National Institute of Standards and Technology. NIST. Published 2023-01-26; last reviewed 2026-06-20 UTC. Government framework.
National Institute of Standards and Technology. NIST. Published 2024-07-26; last reviewed 2026-06-20 UTC. Government profile.
OpenTelemetry project. Cloud Native Computing Foundation. Published Current specification repository; last reviewed 2026-06-24 UTC. Official specification.
Enrique ter Horst and Juan Diego Zambrano. arXiv. Published 2026-03-11; last reviewed 2026-06-24 UTC. Research preprint.