From an open question to a reviewable result.

One investigation, one durable record

An investigation is the unit of work. It belongs to an organization and project, carries a hard spending ceiling, and advances through explicit states instead of disappearing into a single opaque model call.

The record grows over time: the original question, protocol revisions, approvals, source manifests, execution receipts, checkpoints, progress events, artifacts, and review decisions all stay attached to the same investigation identifier.

• Tenant-scoped from creation
• Versioned protocol and approvals
• Append-only progress and audit history
• Private artifacts unless an export is approved

The research lifecycle

Each phase has a clear entry condition and a safe stopping point. A failure, expired lease, or exhausted budget pauses work at a checkpoint rather than silently widening the scope.

1. 01Frame the question and success criteria
2. 02Prepare the protocol and evidence boundary
3. 03Approve the protocol, model, compute, and ceiling
4. 04Gather sources and execute bounded analyses
5. 05Checkpoint, recover, and continue when needed
6. 06Review claims, null results, and limitations
7. 07Assemble the approved artifact set

Long-running work without a fragile browser session

Cloud execution is coordinated by a durable workflow. The workflow claims a time-limited lease, reserves the worst-case provider charge, calls GPT-5.6 Luna only through OpenAI’s official Responses endpoint, settles actual usage, and records a hash-verified checkpoint between every phase.

The governed launch phases are evidence mapping, protocol design, adversarial review, and a review-ready final report. If an executor disappears, scheduled maintenance detects the stale lease. ApolloBot resumes only from a verified checkpoint and only when the retry policy, budget, approval, credentials, provider receipt, and stored phase result still agree.

What completion means

Completion is not merely a generated answer. A hosted investigation reaches human review with its evidence map, proposed methods, adversarial findings, source receipts, provider usage, limitations, and hash-verified report attached. Proposed experiments are not represented as executed work.

• Claims linked to supporting and conflicting evidence
• Proposed methods separated from completed work
• Nulls, uncertainty, and limitations preserved
• Artifact hashes verified before retrieval
• Publication, repository writes, and compute held behind separate approvals