How each answer is built and checked.

Source retrieval

A research plan produces bounded queries. Results are fetched through a provider adapter on the server, normalised and de-duplicated before they are stored.

Each stored source keeps its type, host and retrieval date. The run keeps the queries that produced it.

Atomic claims

Sources are broken into claims that each assert one proposition. This makes a claim checkable on its own and comparable across sources.

Claim extraction is automated and can be wrong. The source excerpt stays beside the claim so a person can verify the wording.

Evidence relationships

Each claim links to sources as supporting, contradicting or contextual evidence, with a short note.

Contextual evidence covers cases such as a newer changelog entry that may explain why an older review no longer applies.

Contradiction handling

When sources disagree, the claim is marked contested and the evidence on each side stays visible.

The system does not choose the more convenient source. Resolution is a reviewer's decision.

Insight synthesis

Insights are written from stored claims only. Each links to its claim IDs.

Downstream wording must respect contested claims: a battlecard should not state a contested claim as fact.

Confidence and freshness

Confidence is a label describing how well the stored evidence supports an insight. It is not a probability and not a guarantee of accuracy.

Freshness reflects the age of the evidence an insight depends on, so stale intelligence is visible.

Monitoring and baselines

A watchlist holds public pages and search queries. Pages are fetched on the server, limited to public web addresses, and stored as bounded text snapshots with a retrieval date.

The first check of a page is a baseline: nothing is compared yet, so no signal is raised. Later checks are compared with the last stored snapshot, and snapshots are kept so history is available.

A source that fails is reported as a warning. A check where every source failed is reported as failed and does not count as a baseline.

Change detection

Detection is deterministic. A price that moved, a free plan that disappeared, a launch line that was added or a headline that was rewritten is a fact about two stored snapshots, found by comparison rather than by asking a model what it thinks changed.

Pages containing instruction-like text are excluded from analysis and flagged, because scraped content is treated as data and never as instructions.

Only pages whose text is available without signing in or running scripts can be read today.

Signals, importance and confidence

Each change becomes a signal with the before and after and a stored claim describing it. Importance (low, medium, high, critical) reflects the type and size of the change. Confidence (low, medium, high) reflects how directly it was observed and how reliable the source is.

A model may reword a signal and suggest actions using your workspace context. It cannot add facts, it can move importance by at most one level, it cannot create a critical signal from a lower one, and a response that cites an ID that does not exist is discarded in full.

A signal never becomes an action directly. Promote it to a draft insight; a person approves it first.

Briefs

A brief is assembled by rules from the signals of a period: ranked by importance and then confidence, with dismissed signals left out, a summary and a short list of recommended actions.

A model may rewrite the summary and prioritise the actions. Every signal it cites must exist, otherwise the rule-built brief is used. Each brief records which of the two produced it.

Human review

The product proposes; a person approves or rejects. Only approved insights feed actions and Ask.

When evidence is missing

When stored evidence cannot answer a question, Ask says so and offers a new research run. It does not present generic model knowledge as workspace evidence.

Questions about a time window, such as what changed this week, search the stored signals and claims for that window and say when nothing was recorded. “What don't we know?” is answered from stored state: contested claims, single-source claims, subjects with no claims and insights waiting for review.

Outcomes

After an approved action is used, a person can record what happened. The workspace tallies outcomes by action type so you can see what has worked. It does not yet use those tallies to rank future intelligence automatically.

Citation validation

Answers and actions cite IDs of stored claims and sources, so a citation always resolves to an object in the workspace.