AGI Scorecard中文
All sections
JarvisResearch pilot

One goal. A useful next step.

Find the evidence, work through the trade-offs, and leave with something you can act on. Your next visit starts where you left off.

Your private workspace

Try one task free, then register to continue. A private browser key keeps your tasks separate.

Register or sign in free

Restore another workspace (optional)

Progress you can inspect.

  1. A clear goal

    Bring a real task and choose the context that matters.

  2. Work with evidence

    Jarvis runs a fixed evidence workflow and records what actually ran.

  3. A testable next step

    Inspect sources, record whether you tried, completed or were blocked, and edit a corrected attempt when needed.

Built for AI work, learning and product research. Current tools return editorial records, public metadata and arithmetic; they do not read entire linked articles or execute your action plan.

Explore the source library

Before you hand over a goal

How does the free trial work?

Guests can submit one task, then register a free AGI account to continue. A saved task uses the trial even when AI is unavailable. Reading, exporting and deleting its result remain available. An interrupted first run can resume saved work, but a new run or daily watch needs an account. One trial is recorded per browser key; the same IP can start at most one guest trial in 24 hours. This is abuse prevention, not verified identity: shared networks may need to register sooner. Registration uses the existing AGI display name and private login-key backup, with no email, payment or extra model quota.

Can I continue a saved Future Guide plan?

On the distribution-opportunity page, explicitly attach the reviewed business evidence and prepare one selected plan in Jarvis. A single-use browser-session preview expires after 30 minutes. Review it and apply it to an empty goal, then choose cloud consent and Start yourself. Only the final visible goal and selected evidence receipt are submitted; editing it does not change the original local plan. The saved task and export retain the evidence version, source links, limits and self-reported fit even if AI is unavailable. This prepares an unverified draft or source pack; it does not diagnose, repair or execute a workflow. CSV execution records belong in the existing Earn delivery lab’s workflow mode, which flags counts, duplicate run IDs and empty outputs rather than reconciling individual business records.

What can Jarvis do today?

Research AI work, learning and product ideas using our editorial catalog. With your choice, it can search GitHub and Hacker News metadata. It can calculate, compare findings and draft concrete next steps. It cannot control your computer, access your accounts, send messages or buy things.

Does it continue when I leave?

Submitted tasks are stored on the server. A background runner checks the queue about every two hours; scheduled checks can be delayed. After free registration, daily watch runs up to seven times over seven days; results appear here, with no email or push notifications. You can pause or delete a task. Resuming an interrupted run reuses saved steps; a model call with an unknown outcome is not silently repeated. The guest trial can finish in the background; further runs need free registration, not a paid membership. Previously paused tasks need an explicit resume.

Will my last task reopen when I return?

Yes, on this device and browser. Opening or creating a task saves a strictly validated pointer for up to 30 days, so the same task can reopen after a reload or later visit. The pointer contains only its task identifier, workspace scope, language and time—not the goal, result, memories, consent or key—and it never authorizes access; the server still verifies the private workspace and returns the task before it can reopen. An unsent goal draft takes priority. New task, delete, sign-out, workspace switching, an expired or missing task, and any malformed or cross-workspace pointer clear or ignore it.

How does Jarvis suggest where to continue?

For a report with next actions, Jarvis derives one optional review cue from the current run’s fixed self-reported states: a blocked step first, then the most recently tried unfinished step, then the first unmarked step. Completed steps are skipped. The cue contains only an action number and fixed state; it does not create memory, call AI, change progress, scroll automatically or prove that the suggested step is correct. Review every step before acting.

What happens when AI is unavailable?

You get the matching source pack and an explicit limited status. No preset is passed off as AI output. A new run gathers evidence with code and uses at most one synthesis call; an older checkpoint can total two calls. Explicit GitHub metadata reading-candidate requests can receive a limited rule-based packet with no new model call. It shares the existing site allowance of 12 attempts per 24-hour window and 3 per IP window. Failed attempts count too.

What does a rule-based metadata packet establish?

For explicit one-candidate GitHub metadata screening, Jarvis selects the first valid repository in the returned order, copies attributed fields and retains executed arithmetic. This is an inspection order, not a quality ranking. The packet proposes a bounded checklist; it does not read README or code, establish licensing or safety, or prove that all goal conditions are met. Other tasks still produce unverified AI drafts.

Can I record what happened after a suggested action?

Yes. For each current next action, you can mark Tried, Completed or Blocked. A blocked step offers an editable recovery draft for the blocker, the constraint to preserve and a smaller safe alternative; it does not retry or call the model until you review, consent and submit. This is your self-report, not independent proof that the real-world goal succeeded. Jarvis stores only the fixed state, action number, run number and update time; it does not duplicate the action or task text in this progress receipt. A new daily report gets a new set of progress states. Deleting or expiring the task removes these receipts.

Where are my goals and memories stored?

An unsubmitted goal draft can stay for up to two hours in this browser tab, scoped to this workspace and language. Only the goal text is restored; consent, public search, keywords, cadence and memory selections are not. You can discard the draft, and it is cleared after submission, sign-out or workspace switching. Browser storage is readable by scripts on this origin, so do not enter sensitive data. Editable memories stay in this browser; only the relevant memories you select are sent with a task. Submitted tasks and their selected context are stored in Cloudflare D1 for 30 days and sent to Workers AI for inference. Enabled public searches send only your explicit public search keywords to GitHub or Hacker News. A private key saved in this browser protects task access; clearing it can lose access. Existing member keys still open their original task history, even after membership expiry. Device memories are separated by workspace. Delete a task to remove it from our active store. A separate minimal trial receipt (workspace hash, task identifier and date, without task text) is retained to prevent resetting the trial by deletion or expiry. Network-limit records stop counting after 24 hours and are removed during bounded maintenance. Network digests attached to a task are removed with that task, at deletion or 30-day expiry. Choosing registration and continuation links this browser’s trial history to your account; device memories can move with it. It does not publish tasks to the discussion room.

Is this AGI, and is it a paid product?

No AGI capability has been established. This bounded research-agent pilot is free, with one guest trial followed by free account registration. No paid membership is required. There is no separate Jarvis plan. Shared model and search limits still apply, and availability is not guaranteed. We are testing whether people find these results useful and return with real tasks.

Does my exported report keep action progress?

Yes. The export includes the current report’s fixed self-reported action states and recorded times. Unmarked steps remain unmarked. It is a static snapshot, not verified completion or a runnable checkpoint; it does not update automatically or include your private memories or access key.

Follow what changes in the AI era.

Subscribe to AGI Scorecard