Skip to content

Final test results and failure analysis

Published: 2026-08-08

This report is the final status of the evidence retained in this repository. PASS means the cited evidence satisfies the experiment's stated acceptance criterion. FAIL means an attempted execution did not satisfy it. INCONCLUSIVE means the retained evidence cannot support the intended claim. The two local contract tests are reported separately from broker-backed experiments.

Results

Scope Status Observed result Evidence
Python contract tests PASS 2 passed pytest -q on 2026-08-08
01 — Multiple AI PASS One work event produced three independent proposals. experiments/01-multiple-ai/results/run-001.md
02 — Human review PASS A human participant emitted ProposalApproved. experiments/02-human-review/results/run-001.md
03 — Event chain PASS An OCI-broker four-event causal chain completed. experiments/03-event-chain/evidence/run-002/
04 — Event replay PASS A fresh replay reconstructed state from all recorded event IDs. experiments/04-event-replay/evidence/run-003/
05 — Failure recovery INCONCLUSIVE Two WorkCreated events yielded only one ProposalCreated. experiments/05-failure-recovery/RESULT.md
06 — Live LLM PASS A real-broker Luna participant produced the required live ProposalCreated event. experiments/06-live-llm/evidence/run-005/
07 — Provider hot swap PASS Deterministic and Luna participants produced proposals at one event contract. experiments/07-provider-hot-swap/evidence/run-002/
08 — Competing consumers INCONCLUSIVE No worker output or benchmark measurement was retained. experiments/08-parallel-scaling/RESULT.md
09 — Audit trail PASS OCI-broker lineage generated a Mermaid audit graph. experiments/09-audit-trail/evidence/run-002/
10 — End-to-end SDLC PASS The corrected OCI-broker run recorded ten events ending in DocumentationCreated. experiments/10-end-to-end-sdlc/evidence/run-003/

Failures and their causes

05 — Failure recovery: inconclusive

The expected recovery result was two proposals for two retained work inputs; only one proposal was recorded. The existing result identifies the likely mechanism as a consumer readiness/partition-assignment race around the worker restart. That mechanism is not proven: raw worker logs, offsets, and assignment telemetry were not retained. The defensible conclusion is that retention alone did not demonstrate reliable recovery in this run.

06 — Live LLM: historical configuration/provider failures

All three retained attempts ended before a proposal event was written:

Run Direct error evidence Cause supported by evidence
run-001 Azure-compatible request returned HTTP 404 DeploymentNotFound. The configured deployment did not exist or was not available to the endpoint at execution time.
run-002 Request returned HTTP 400: invalid configured model identifier. The configured model identifier was invalid for that provider.
run-003 Request returned HTTP 404: configured function not found for the account. The configured provider target was unavailable to that account.

These are provider configuration/availability failures, not model-quality findings. They were superseded by successful real-broker run run-005, which used Azure Sweden and gpt-5.6-luna. See each run's requirements-1.stderr.log and failure.json.

07 — Provider hot swap: historical failed live stage

The live participant in run-001 returned HTTP 404 DeploymentNotFound. The configured Azure-compatible deployment was unavailable at that time, so the sequence could not complete. This was superseded by run-002: the Azure Sweden/Luna participant and two deterministic participants each emitted a proposal from the same input contract. The concurrent run supports contract-level substitutability, not an ordered runtime hand-off.

08 — Competing consumers: inconclusive

Three worker log files were empty and no measurement JSON was produced. The retained evidence therefore establishes neither partition sharing nor throughput. The precise cause is unknown because worker readiness, consumer assignment/rebalance state, host scheduling, and observer instrumentation were not captured. A readiness/assignment problem is plausible, but is not a proven root cause.

06 — Participant identity consistency defect

The passing run-005 event has authoritative top-level source and actor_id values of ai.requirements.azure.v1, but its copied payload.participant_id is ai.requirements.deterministic.v1. The event is retained unchanged. The experiment passed at the event-contract level while exposing a participant-identity consistency defect in the payload. The implementation now writes the live participant ID consistently; Experiment 07's retained live event demonstrates the correction.

10 — First SDLC attempt: harness failure, corrected by retry

The first run failed because the bounded participant raised RuntimeError: consumer ended before receiving the expected event in documentation-9. This is a test-harness lifecycle/readiness failure rather than an SDLC-topology failure: the readiness-checked retry completed and is the basis for the final PASS. The original failure is retained at experiments/10-end-to-end-sdlc/evidence/run-001/.

Implications

The passing runs support causal event exchange, replay, audit lineage, the bounded SDLC topology, and one bounded Azure Sweden/Luna participant through the real broker. They do not establish multi-provider live substitution, reliable recovery, or competing-consumer performance. The inconclusive runs show that a Kafka-compatible event log supplies durable transport but not worker lifecycle coordination, assignment visibility, or execution accounting.