Why you can defend what Signals gives you.
The differentiator isn't “AI foresight.” It's the method underneath: research kept as structured objects, every source graded, grounding made explicit, models compared rather than trusted, and a workflow you can run again. Here's how it fits together.
- Briefingscope the question
- Signalsstructured findings
- Sourcesjudged with verdicts
- Groundingverified or rejected
- Framessynthesized groupings
- Reporta defensible narrative
Every engagement starts from a scoped brief, not a blank prompt.
A Session opens with a Briefing — the objectives, scope, and constraints that steer discovery. It's drafted in dialogue with an assistant that interviews you, then it stays attached to everything the Session produces.
Because the brief is an object rather than a one-off prompt, every signal, source, and report downstream is anchored to the same stated question — and you can re-read why the research was scoped the way it was.
Findings are structured objects, not bullet points.
A Signal is a candidate indicator of change tracked as structured data: a title, summary, category, tags, scores against your metrics, sources, and a grounding state — whether it's still ungrounded or already verified.
Structure is what makes the research reusable. Signals can be merged when they're duplicates, grouped into Frames, scored against Metrics, and cherry-picked into Collections across Sessions.
Every source is judged, not just attached.
A Source isn't a bare link. Each one carries a verdict describing how it relates to its Signal — supports, contradicts, unrelated, or unreachable — so the evidence is graded, not merely collected.
A source that contradicts a signal is as valuable as one that supports it. Recording both is what separates a research record from a citation list.
Trust is an explicit, recorded outcome.
After weighing its sources together, a Signal lands in a grounding state: Ungrounded, Pending, Verified, or Rejected. The judgment is committed with a confidence and a rationale — never left as an implication.
The system enforces its own honesty: a Verified outcome with no supporting source is automatically rewritten to Pending. A terminal state is never sticky — re-running the evaluation always wins.
Self-correcting: a Verified outcome with no supporting source is automatically rewritten back to Pending. Terminal states are never sticky — the latest evaluation always wins.
Convergence across models, not one model's blind spots.
Tasks and report types can draw on a catalog of models. Run discovery and grounding across more than one, and where they converge you gain confidence — where they disagree, you see exactly where to look closer.
No single model is treated as ground truth. The method is built around comparing them, with every generation recorded against the workflow run that produced it.
The method is repeatable, and it compounds.
A Workflow is an ordered procedure of Tasks defined on a Project. Run it once for a starter pass, or re-run it as the picture changes. The Project owns the shared structure — categories, tag sets, frame sets, metrics, report types — that every Session inherits.
Because structure lives on the Project, each new Session deepens the same theme instead of starting cold. The research is an asset that accumulates, not a deliverable that expires.
See the method produce a result.
Enter a research theme and we'll run a starter Session on our credits — briefing, signals, judged sources, and a starter report.