Point
Retrieval does not resolve semantic context
Retrieved candidates need interpretation of identity, scope, source sufficiency, and task relevance before they become an answer or an authored change.
- Posture
- asserted
- Lifecycle
- active
Start from the available clue: a source path, Point, subject, or desired outcome. Return relevant context and reasons, not every connected record. A fuzzy concern can justify several evidence-backed directions without declaring one interpretation certain or demanding a fully specified request.
When mapped context is insufficient, consult authorized sources. Distinguish missing evidence from contrary evidence, and a selected policy from observed implementation. Disagreement should remain visible with its scope and sources. The recorded graph is open-world even when maintained; missing edges establish no absence of impact.
Routing remains read-only unless a write is authorized. These instructions guide existing agents; they do not implement a retrieval engine or establish that agents select context correctly. That outcome requires a separate cold-start trial.
Identity searchSimilar passages may express different claims; candidate inspection must precede an identity decision.
Contribution boundariesSelection for an answer does not itself authorize changing the selected records.
RoutingPrecise and fuzzy requests need different useful selections, not a forced single-Map destination.
Current processors preserve exact authored identities and direct authored relation edges, while the candidate-disclosure authoring workflow remains unimplemented.
Both current processors group Point records only by exact id. They carry authored outgoing relations into normalized output and derive only direct incoming reverse entries. Neither implementation uses prose similarity to merge Point identities or constructs reciprocal or transitive edges.
The repository contains no authoring tool that surfaces candidate identities, exposes anchor provenance, and preserves an explicit new-anchor or existing-context choice. That duty remains a conformance boundary for future software rather than a shipped capability.
Reference validatorThe validator groups records only by exact id and emits authored outgoing relations with their direct incoming reverse entries.
Independent processingThe independent reader separately preserves distinct ids and avoids constructing transitive or similarity-derived relations.
Absent and planned toolsNo shipped authoring tool presents candidate identities and anchor provenance before recording an explicit identity choice.
Atlas’s agent-facing useful difference begins after topical retrieval, where explicit identity and standing should improve selection, explanation, and continuation.
Atlas does not need to replace generic document retrieval to support an agent workflow. Its added claim is that a person or agent can inspect why material belongs together, which record anchors the shared meaning, how each perspective contributes, what standing applies, and where an authorized task should continue without minting another account.
That distinction must be tested against a serious same-corpus document and search control. Fluent retrieval alone is not evidence that Atlas-specific identity, state, provenance, or relations improved the task.
Useful differenceThe product claim concerns semantic selection and continuation after retrieval, not whether Atlas can duplicate generic topic search.
Shared navigationInspectable identities, standing, relations, and provenance give people and agents a common basis for choosing what to read or change next.