This Cross-AI page is English only. Language controls on other site pages are unchanged.
Experimental building blocks. Note: Connecting each AI may require adjustments on your side (API keys, model choice, prompts, and runtime settings). Treat these desks as experimental building blocks, not a guaranteed turnkey multi-cloud product.
Cross-AI · Research Desk
Research Desk
Experimental Research Desk case study. Date 2026-09-10 JST. Earlier RESULT PASS refers to limited workflow checks, not answer reliability. No Library Verified claim. Not a multi-cloud product claim.
Answer quality: 0 / 9 accepted (September 11 battery; incomplete transport counted separately). Kit workflow can run; answers are not reliable for unattended research, publication, or trading.
Experimental local workflow — answer-quality evaluation failed. Earlier RESULT PASS = limited workflow checks only, not answer reliability.
The September 11 review found unsupported publication approvals, false claim attributions and missed numerical inconsistencies. The earlier PASS refers to limited workflow checks, not answer reliability. This kit is for controlled experiments, not unattended research, publication approval or trading decisions. Outputs require checking against the original sources. No external action was executed in these tests. This is sequential local Ollama use, not simultaneous collaboration between cloud AI services.
botshelf-deep-research-3b).
web-fetch-curl, local-file-audit, and executor-synthesis are process seats — not separate LLM providers.
ChatGPT was not used. A later repair run proved a sequential two-model Ollama workflow (botshelf-deep-research-3b → botshelf-doc-review) — same host, two model tags; not concurrency and not multi-cloud.
Setup / handoffs / gates
- How to run: use the local run kit (prerequisites, Ollama model create, automated sequential commands, and Manual handoff when tags are UNAVAILABLE).
- Envelope path: kit runner writes sequential JSON envelopes under
$OUT/handoffs/; redacted samples live at sample-output/handoffs/. Never claim simultaneous agents. - Failure handling: runner
ollamasteps exit nonzero on transport / incomplete / empty failure; labelBLOCKED/FAILED/UNAVAILABLEhonestly — do not invent cloud-provider output; do not run AI-2 after AI-1 failure. - Human approval: corrected final stays
WAITING_ON_HUMAN/AWAITING_HUMAN_REVIEWuntil a human edits/approves. No publish, Verified flip, or external action without Human Yes. Answer quality remains 0 / 9. - Automation vs manual: kit documents both the automated sequential Ollama path and the Manual handoff path (copy/paste between seats when Ollama is unavailable). Manual path ≠ implemented multi-cloud automation.
Where it sits
Cross-AI orchestration package with public proof. Uses a Library Ollama recipe (botshelf-deep-research-3b) as one seat. It is not a Marketplace listing and not a Verified Library badge.
Original verified seats (sequential)
web-fetch-curl— process · 8 public Library URLs, HTTP 200ollama:botshelf-deep-research-3b— LLM · local Ollama · ~13.8s · llama3.2 3.2Blocal-file-audit— process · Proof Engine markdown/JSON auditexecutor-synthesis— process · final brief (NOT ChatGPT; model version unknown)
Runnable local kit (prerequisites, inputs, steps, limitations): /cross-ai/kits/research-desk-local/. Manual copy/paste between model steps is documented. No registration or payment required for the kit.
Hard facts from the run
- RESULT: PASS (limited workflow checks only — not answer reliability)
- Conflicts logged: 2 (recorded, not hidden)
- No X / no Verified flip / no mass DM
- Public proof: research-desk-2026-09-10
Open run kit Public proof Cross-AI hub Library deep-research recipe