Skip to content
JANUSOBSERVATORY

Research companion · bounded evidence workflow

Evidence first. Answers second.

Search the allowlisted Janus corpus, preserve what the instruments did not evaluate, and keep every factual paragraph tied to a resolvable source. The Story, Observatory, and Atlas never depend on this service.

Lexical corpus196 passage records
Scenario coverage10 of 10
Canonical releaseindependent review pending
Independent replayfixed 20,000-run preset verified

Ask the corpus

A cited answer—or an honest stop.

Questions are limited to the approved corpus and deterministic Janus records. The service does not store raw questions by default.

69 / 600

Preview disabled.

Runtime boundary

One short workflow. No autonomous research swarm.

01

Route

Reject probability rankings and requests outside the Janus evidence boundary before any provider call.

02

Retrieve

Search verified-open papers and candidate canonical summaries. Unknown-rights scenario PDFs never enter the public corpus.

03

Audit

Validate every citation marker against retrieved source pages. Invalid or uncited provider output falls back to deterministic evidence.

Allowed at runtime

Small, explicit tools

Corpus search and passage fetch surround six typed domain tools: scenario lookup, two-to-three-scenario comparison, the categorical observability matrix, the published observer lookup, reported collapse metrics, and a fixed 20,000-run independent replay. One bounded round may make at most six domain calls. Each answer carries call names, statuses, and deterministic SHA-256 receipt hashes. There is no arbitrary web, SQL, filesystem, code execution, account mutation, or free-parameter simulation.

Release gate

Prompt surface held closed

JANUS_RESEARCH_PREVIEW_ENABLED controls the prompt surface; the lexical, deterministic path can run without provider credentials. Live generation separately requires the server-only live-call flag, valid credentials, and deployment canaries. Hybrid retrieval additionally requires a matching prebuilt index and embedding service. Shared multi-replica request budgets remain a deployment gate.

Independent research

Reported values and reimplementation never share a label.

The resilience experiment independently implements the paper equations and compares its ensembles only with prose-reported aggregates. Its outputs remain noncanonical and its year-step choices remain explicit implementation decisions.

Inspect resilience validation →