JARVISLabs

Research that produces a cited report and opens it

On 12 September JARVIS researched a question from two named documentation pages, wrote a Markdown report and an HTML copy through its governed file owner, and read both back by digest. Six findings each cite a named source. One live defect was repaired.

LIVE-PROVEN What this run supports. Recorded 12 September 2026.

A rendered research artifact answering a question about Python's asyncio.timeout, with sections for question, retrieval, sources and findings, each finding quoted with its source and fetch time.
A research artifact produced by JARVIS’s research desk on 12 September 2026. Findings are quoted as documented facts with their sources; what could not be established is listed separately.

Asked

What was asked

Research a Python standard-library question using exactly two given sources, then write the research up as a named artifact.

Result

What happened

research.run fetched two current docs.python.org pages through the governed research desk (both HTTP 200, fetch dates recorded) and kept a durable dossier. research.artifact wrote a Markdown report and an HTML copy into the owner's designated research folder through the governed file owner and read both back by digest. The HTML rendered with all nine required headings. The HUD file-results widget showed four governed receipts.

Pipeline

Architecture path

  1. Owner request (typed inlet)
  2. Hosted cognition proposes research.run
  3. Governed research desk (isolated browser, robots, fetch dates)
  4. Dossier
  5. research.artifact
  6. Governed file writes
  7. Read-back by digest
  8. HUD file-results widget

What this proves

  • On the normal launcher, a research request produced a durable, cited artifact that was read back by digest and opened correctly.
  • Each of six findings quotes a named source with its URL and fetch date.
  • The report separates evidence from assumption: contradictions are stated as not detected rather than as agreement, and one assumption is marked NOT ESTABLISHED.

What it does not prove

  • Two public documentation pages are not a broad research task.
  • Recorded on 2026-09-12; a recorded run is not a current reading.
  • The sources did not state publication dates, so recency was not assessed.
  • The drive used the typed inlet with hosted cognition, not voice.

Context

Acceptance journey J5 asks for research that “produces a cited artifact that opens correctly and distinguishes current evidence from assumptions.” The research desk browses in an isolated context, obeys robots.txt and refuses private hosts. Sources become evidence in a claim graph. Lexical matching can mark a claim CONSISTENT at most. Only an entailment check may call it ENTAILED, because “a keyword match must never be laundered into ‘verified’.”

The run

On the normal launcher with the conversation loop, hosted cognition and the typed inlet, the owner asked for research on a Python standard-library behaviour using exactly two named docs.python.org pages. research.run read both through the governed desk (HTTP 200, with fetch dates recorded) and made the dossier durable. The owner then asked for the research as a named artifact. research.artifact wrote a Markdown report and an HTML copy into the owner’s designated research folder through the governed file owner and read both back by digest.

The driver checked the result independently:

  • the HTML rendered in headless Chrome with all nine required headings;
  • the Markdown held six findings, each quoted from a named source with URL and fetch date;
  • the contradictions section stated that none were detected, and said explicitly that this is not agreement;
  • one assumption was marked NOT ESTABLISHED; there were no retrieval failures, and the untrusted-content notice was present;
  • the HUD’s file-results widget showed the four governed receipts (two creates, two read-backs), without paths by design.

What failed first

The first drive exposed a real defect. The desk summarised each page’s opening prose instead of answering the question, and cited an interpreter start-up speed-up as if it were relevant. The repair makes the desk take question-focused sentences first. It reads 120,000 characters of a page instead of 40,000, and marks a sentence as relevant only if it contains two of the question’s terms. The drive was then repeated.

The evidence

Journey J5 in docs/orders/post-lm/ACCEPTANCE_JOURNEYS.json records the run, the checks and the repair. Commit b9c9b3dec marked it LIVE. The rendered artifact and HUD capture are local evidence files and are not published. No footage exists yet.

The limits

Two documentation pages make a narrow test. Neither source stated a publication date, and the artifact says so instead of inventing one. All journey readings predate the source of 17 September, so the project does not treat them as proofs of the current tree.