Published by Gigabrain Labs · Updated
Follow the source behind an answer
See how Gigabrain connects project records, finds source passages and handles a correction. These examples use fictional documents. No customer material appears here.
A connected project and a changed date
| Step | Observed result |
|---|---|
| Connect the documents | Project Alder depends on an access review, which follows a separate schedule. |
| Retrieve the source | The Monday schedule sets the inspection for 18 October. |
| Change the schedule | The old citation fails verification after its source changes. |
| Refresh and retrieve | The Monday schedule sets the inspection for 25 October. |
The result quotes the source passage. It does not measure a model’s answer quality.
Setup and source details
Executed on 2026-09-14 in both System and Full archives. The local memory service used hybrid retrieval and graph expansion.
The graph links remain agent-authored interpretations. Their source support stays inspectable.
Inspect the recorded quotations, paths and both archive hashes
Reproduce the correction with the licensed archives and public test files.
Inspect the 8 September reproduction check. The original execution date remains attached to the original result.
Memory and MCP component checks
Current component result. Execution date: 2026-09-14.
These tests exercise local memory operations and the MCP protocol. They do not run a language-model conversation.
Inspect the results and skipped-test count, then follow the matching reproduction instructions.
Inspect retrieval results and source files
The editorial experiment compares retrieval modes on fictional project records. It measures recovered sources and citation checks, not generated-answer quality.
Read the four-mode retrieval comparison for the question set and measurements.
Read the source files, inspect each result and check the runtime and archive identity.
Every tested mode returned passages for the unsupported query. A related passage still needs a relevance check.
A configured ingestion loop
The scheduled run read approved fictional source updates and refreshed their search and graph records.
A repeat run made no extra authoring calls. The test removed its schedule afterwards.
Schedule, agent and permissions
The demonstration used the native macOS scheduler and an approved Codex graph-authoring runtime.
Agent version: 0.153.4. This optional authoring step used OpenAI through the configured account.
The source scope, schedule and provider access were approved for this run. It does not establish scheduled execution of every pre-built loop or support on every agent.
Reproduce the demonstrations
Download the fictional sources, recorded results, public test tools and a review worksheet in one package. Use your licensed Gigabrain archive to run the checks in an isolated installation.
Download the evaluation package or inspect its file manifest.
Gigabrain Labs assembled this material. Each result retains its original test date and environment.
Use the same checks on your own work
Start with a question you can check. Inspect the source, add a correction and ask again. Include a question the documents cannot answer. A related passage may still fail to support the answer.
Choose your agent setup and follow the first-task guide.
