What the tool can do
- Freeze the system, authority, tasks and baseline for one evaluation
- Keep per-run results, traces, critical failures and transfer boundaries together
- Import, copy and download handoff-ready JSON locally in the browser
FUURAA AI Knowledge Library · Browser-local tool
AI Agent Evaluation Run Record Builder
Turn one agent-system evaluation into a reconstructable, comparable and expirable JSON record. Entries are processed only in this browser page unless you choose to download or copy them.
Run-record workspace
Six sections connect system identity, run evidence and a scoped release decision.
Start blank or load a clearly marked fictional example. Completion checks only field presence—not study quality, evidence truth or deployment safety.
0%38 fields remain incomplete.
Decision and system identity
Freeze the exact deployment question and the complete acting system.
Task distribution and baseline
Describe the work sampled, who it affects and the comparison that makes the result meaningful.
Evaluation design
Predeclare the data, fixtures, scorers, thresholds, repetitions and accountable evaluator.
Results and process evidence
Keep outcomes and trajectories together so averages cannot hide critical failures.
Transfer and applicability boundary
Name tested contexts and transfers that remain unsupported.
Scoped release decision
Bind the decision to an owner, monitoring, rollback, expiry and rerun trigger.
JSON
Machine-readable run-record preview
The export uses RFC 8259 JSON. The field structure is a FUURAA public method—not an international standard, audit or certification format.
{
"schema": "fuuraa.ai-agent-evaluation-run-record/v1",
"record_id": "",
"deployment_decision": "",
"system_identity": {
"agent_release": "",
"model_stack": "",
"tools_and_connectors": "",
"execution_environment": "",
"memory": "",
"authority_boundary": "",
"human_controls": ""
},
"task_context": {
"intended_users": "",
"workflow": "",
"task_distribution": "",
"failure_consequence": "",
"frozen_baseline": ""
},
"evaluation_design": {
"dataset_or_task_set": "",
"fixtures_and_initial_state": "",
"scorers_and_rubrics": "",
"thresholds": "",
"repetitions_and_seeds": "",
"evaluation_owner": ""
},
"results_and_process": {
"per_task_outcomes": "",
"process_trace": "",
"variance_and_uncertainty": "",
"recovery_and_handoff": "",
"cost_and_latency": "",
"critical_failures": ""
},
"transfer_and_boundary": {
"languages": "",
"geographies_and_domains": "",
"operators": "",
"excluded_uses": "",
"residual_risks_and_unknowns": ""
},
"release_decision": {
"state": "",
"decision_owner": "",
"monitoring_plan": "",
"rollback_and_stop_plan": "",
"expires_at": "",
"rerun_trigger": "",
"rationale": ""
}
}Applicability boundary
Structural completeness does not make an agent deployment-ready.
What the tool cannot prove
- That a run occurred, logs are untampered or an owner identity is valid
- That tasks are representative, scorers valid or conclusions transferable
- That a system is safe, lawful, compliant, audited or certified
FUURAA analysisA useful agent run record serves one concrete deployment decision and lets another reviewer reconstruct the model, tools, authority, environment, tasks and failures. No aggregate score can cancel an authority breach, unrecoverable side effect or other critical failure.
Primary sources and boundaries
Use frameworks to design an inspectable record—not to impersonate release approval.
26 January 2023
NIST AI RMF 1.0
Connects system context, measurement, governance and continuing risk management across the AI lifecycle.
BoundaryVoluntary and use-case agnostic; it does not prescribe this record or certify a deployment.
Open primary source ↗26 July 2024 · page updated 8 April 2026NIST AI 600-1 · Generative AI Profile
Adds generative-AI risks and actions around measurement, provenance, information integrity and human oversight.
BoundaryIts actions still require prioritisation for the actual system, affected people and deployment context.
Open primary source ↗Released May 2024 · documentation checked 4 August 2026UK AI Security Institute · Inspect
Provides an open evaluation framework organised around tasks, datasets, agents, tools, scorers, sandboxes and logs.
BoundaryA framework can make runs inspectable; it cannot make a weak task distribution representative or a scorer valid.
Open primary source ↗Version 1.1 · last updated 8 May 2026METR · Task-Completion Time Horizons 1.1
Illustrates a reproducible way to relate success probability to the length of coherent software tasks.
BoundaryMETR says measurements above 16 hours are currently unreliable; the metric is not wall-clock autonomy or universal deployment evidence.
Open primary source ↗Internet Standard · December 2017IETF RFC 8259 · JSON
Defines the portable structured-data interchange format used by the exported run record.
BoundaryValid JSON syntax does not establish that a run occurred, a trace is authentic or a release decision is sound.
Open primary source ↗Method, architecture and review