AGENT 01build_evidence_ledger
Evidence Scout
Separates supported facts, inference, and open questions.
INPUT · source packetIndependent release evidence
Agentic systems,
tested as systems.
Run reproducible black-box checks for contracts, evidence, approval gates, live boundaries, idempotency, and latency.
01 / LIVE EVALUATION
Choose a deployed system
Ready. The lab calls only the three allowlisted public deployments.
- / 100
02 / RESULT
-
03 / CREWAI TEACHING TRIAL
One brief.
Four accountable agents.
Visual replay
This teaching view mirrors the four-agent crew deployed in CrewAI AMP. A deterministic no-key model proves the live handoffs and named human-review boundary; it is not a model-quality benchmark.
CREWAgentic Systems Editorial Crew
Ready to show the handoff.
AGENT 02map_the_system
Systems Mapper
Turns evidence into roles, handoffs, gates, and receipts.
CONTEXT · evidence ledgerAGENT 03draft_content_pack
Field Note Editor
Drafts the personal post and a distinct company adaptation.
CONTEXT · ledger + mapAGENT 04review_release_packet
Claims Gate
Checks every claim and stops at human review.
CONTEXT · complete draftHUMAN GATEAhmad owns the publish decision.
LOCKED
Execution timeline
Sequential process- 01Evidence ledgerQueued
- 02Systems mapWaiting for context
- 03Channel-separated draftsWaiting for context
- 04Claim audit + release stateWaiting for context
Truth boundary: this page replays the architecture in the browser and never claims that animation is an AI run. The deployed AMP automation is the source for live execution status and traces. Its deterministic teaching model proves orchestration without an external model key; it does not prove generative quality. The hosted run ends with a review packet; the locked gate is enforced by having no publishing tool or connected social account.
Technical case study 04Published by Ahmad Bukhari. Built and verified by AiXCEL Solutions.
04 / SCENARIO BRIEF
A live evaluator that proves it can fail.
Agent demos often validate only the happy path. This lab calls three allowlisted public systems, verifies seven contract and governance controls, repeats requests for idempotency, probes forbidden live mode, measures latency, and exposes evidence for every score.
A permanent score of 100 proves nothing.
The original interface only showed a passing baseline. Missing evidence, approval bypass, and latency breach can now be injected explicitly and each lowers the score.
The evaluator cannot fetch arbitrary URLs.
Target names resolve through a code allowlist. This prevents the public service from becoming a server-side request-forgery tool.
Public checks are not production certification.
Privacy, provider quality, semantic task success, and real-data mutations require authenticated project-specific evaluation.
05 / ARCHITECTURE
Probe, compare, score, and gate.
06 / DECISIONS
Binary controls stay deterministic.
The public URL, headers, latency, and boundary behavior must work outside the repository.
An approval bypass or missing evidence has one repeatable result. A model judge is reserved for calibrated qualitative output.
The public evaluator has no reason to fetch arbitrary hosts.
Idempotency is observable only when the same key is used more than once.
The bounded request surface is small enough for the platform runtime.
Current checks finish synchronously. Scheduled suites and long model graders are the trigger for a queue.
07 / EVALUATION
Seven weighted checks
- Availability
- Typed run contract
- Resolvable evidence
- Human decision gate
- Live-provider boundary
- Idempotent replay
- Bounded trace and latency
08 / OBSERVABILITY
Current monitoring path
- X-Trace-ID and JSON logs
- Per-probe latency and evidence
- Vercel runtime logs
- GitHub Actions regression tests
- Score, grade, and fault label in the response
09 / NEXT GATE
Production evaluation layer
- Durable run history and schedules
- OpenTelemetry trace correlation
- Langfuse or LangSmith qualitative suites
- Sentry application errors
- Human-labelled golden sets and drift alerts