Temporal state
Ask what applies now, what applied at a chosen time, and what was known then. Include late evidence, withdrawal and supersession.
Research & Benchmarks
Intelligence should improve outcomes—not just retrieval scores.
Useful intelligence must help people finish work with the right evidence. Inspect our methods and limitations, and distinguish an illustrative example from a measured result.
Evidence status
Know what you are looking at.
- Public contracts
Available to inspect
The public MCP documentation describes mapped knowledge access, source inspection and preview/approval boundaries. It is a technical reference, not a benchmark.
- Synthetic example
Available on this page
The downloadable Atlas sources and prepared review illustrate the reasoning to inspect. No product service, retrieval algorithm or model runs when you explore them.
- Pure-function evaluation
Under review
An isolated correctness baseline is being reviewed. No result from that work is presented here as production, connected-workflow or cross-vendor evidence.
- Connected workflows and outcomes
Results not published here
Connected persistence, live GitHub evidence, MCP authorization, real-client continuity and validated workflow outcomes still require separate evidence. No efficiency or superiority result is claimed.
Canonical workflow traces, validated outcome/cost reporting and procedures learned from approved traces are separate research work. Their presence on a roadmap does not establish availability or a verified end-to-end workflow.
Methods before claims
Test the evidence chain, including where it breaks.
Contradictions and source fidelity
Distinguish a real conflict from a later replacement. Check source revisions, missing evidence and fabricated references.
Context selection
State required and excluded evidence before running a query. Report omissions, stale indexes, source coverage and scope boundaries.
Agent continuity
Retrieve in a separate authorized session after a source changes. Exercise missing and revoked access; record the actual client and transport.
Workflow outcomes and cost
Count validated completions, failed and unknown attempts, intervention and elapsed time. Keep missing tokens or cost unknown; label estimates.
What a published result must include
A versioned corpus, independent expected outcomes, code revision, run date, environment and declared mocks. Report each pass, failure, skip and unsupported case, with category denominators. Compare the same tasks and settings. Publish source and corpus hashes, reproducible commands, limitations and measured versus estimated costs.
A pure-function test does not establish connected persistence, real-client authorization or successful work completed. Prepared answers do not measure retrieval or model quality.
Inspect the example
A new constraint. A decision worth revisiting.
Fictional project, prepared sources and review. These controls reveal the example; they make no AI requests, save no notes and grant no permissions.
Original decision
Atlas/Decisions/014-import-queue.md
Revision: synthetic-v1 · Recorded: 2026-09-01
Keep imports asynchronous so a browser closing does not interrupt the job. Retain the original upload for seven days. Reuse a stable job key when retrying. The seven-day retention period is an assumption to review if storage policy changes.
Inspect the context-selection case
Question: Can I replace the queue with a longer browser timeout?
Expected observation: A retention change does not remove the need to survive a browser closing. Cite both sources, preserve the queue requirement and flag the unresolved recovery question.
Excluded: an unrelated project launch note. These expected observations are authored examples. No retrieval algorithm, authorization boundary or model was executed to produce them.
Inspect the contract behind the workflow.
The public MCP reference describes authorized mappings, decision evidence and write previews. Documentation explains the contract; the prepared example above is not a connected test or a measured result.
A technical evaluation you can inspect
Start with one decision your team keeps explaining.
Use a synthetic project first. Agree on the expected answer and a failure condition before connecting a client. This is a self-directed evaluation, using your existing account and plan.
01
Choose the evidence
Select an ADR, a changed constraint and a reviewed brief. Keep source paths, revisions and unresolved questions with the record.
02
Verify an authorized read
Follow the client configuration in the MCP reference, discover the mapping and inspect one retrieved source. Hosted MCP requires Premium or higher; local-only and device-encrypted remote notes are unavailable through that endpoint.
03
Return with a changed constraint
In a fresh session, ask which assumption changed and which decision remains in force. Inspect both sources. Record missing, stale or incomplete evidence as a failure to investigate.
04
Review the proposed next step
Ask for a proposed note update with its sources and exact destination. Keep it read-only unless you explicitly approve a write in your synthetic sandbox. Check current revisions and the resulting receipt.