Back to the Lab
Document investigation · Engineering workbench

Follow the evidence.

Ask a question. Inspect its sources. Change the evidence and compare what happens.

How this works & document preparation

These four Markdown documents were prepared before you arrived: split by headings and paragraphs, embedded, and indexed in a separate collection. Long passages use 400-word windows with 60 words of overlap. Source offsets point back to the original text.

Each live run embeds your question, retrieves from only your selected documents, assembles that evidence, and streams a new answer. No résumé fallback or semantic answer cache is used. Citation links check source IDs; they do not automatically verify a claim. The two webhook documents are explicitly synthetic examples.

Replay follows a captured event log without another model call. Runs stay in this tab's memory; download a report to keep one. The third suggested question intentionally asks for information this collection does not contain.

4 documents · 22 passages · corpus 0e035dfd75bd105b

Evidence collection 4 selected

Uncheck a document to exclude it from your next run. You can still read it.

Synthetic engineering example · v1

Duplicate-delivery incident

Authored synthetic incident; no production claims

Sample incident: one event created two jobs

Status

This is a constructed incident for investigating evidence. All observations below belong to the sample scenario. They are not production metrics, a customer incident, or an account of the portfolio owner's employment.

Observed sequence

At 10:02:00, delivery A arrived for tenant atlas with provider event evt_104. At 10:02:01, the handler queued job job_701. Its acknowledgement was lost. At 10:02:05, the provider retried the same event; delivery B queued job job_702. Both jobs described the same underlying event, but the old implementation used a newly generated job ID as its only uniqueness check.

Diagnosis

The incident's evidence points to event identity being discarded at the enqueue boundary. Checking whether job_702 already existed could not detect job_701 as the same event. The missing invariant was uniqueness of tenant_id plus provider_event_id across deliveries. A queue job ID alone was insufficient in this sample implementation.

Remediation and observed verification

The sample patch persisted the original event identity in an inbox with a unique database constraint and created an outbox entry in the same transaction. In the constructed regression scenario, replaying both deliveries accepted one logical event and classified the second as a duplicate. This result concerns inbox acceptance only; external side effects still require their own idempotency strategy.

Missing evidence

The incident record contains no production p95 latency, financial loss estimate, or guarantee of exactly-once delivery. If asked for those values, an answer should explicitly state that the selected documents do not supply them. Removing this incident document also removes the concrete event ID, job IDs, and observed sequence from the available evidence.

Ready · Every live run uses fresh retrieval and generation.
Execution · inspect each stage

Search only the selected documents in this version of the collection. Scores describe vector similarity, not answer confidence.

Input

Waiting for this stage to start.

Output

No output yet.

Answer

Start with the duplicate-delivery incident. Then exclude that document and ask the same question again.

Retrieved evidence (0)

Scores measure retrieval similarity, not answer confidence.

Passages appear here when retrieval completes.