In-session inference
Fast and flexible, but probabilistic. It is strongest for temporary task context and weakest when a stable attribute must be known rather than guessed.
Section 02 of 13 · Design constraints
Limits of Common Context-Collection Methods
In-session inference, opt-in memory, behavioral signals and third-party data can all be useful. The design problem is that each carries different limits in consent, provenance, validation, cross-domain coherence and updateability.
Phase-separatedCompanion to 21-slot AnatomyDesign argument · not independent validation
Framework reading
What this section contributes
This section keeps the argument narrow: mechanisms are separated from phase, maturity, evidence type and independent validation.
Opt-in memory
Explicit user-provided memory improves persistence, but it still needs provenance, expiry, correction and confidence rules. What a user says should not automatically become permanent truth.
Behavioral signals
Observed actions can be strong evidence of an event, yet interpretation remains context-dependent. A click is not necessarily preference; a request is not necessarily a purchase; a completed interaction is not expertise.
Third-party data
External data can expand coverage, but introduces dependency on source quality, consent chain, freshness, identity resolution and regulatory interpretation.
The design conclusion
No single method should be treated as a universal truth source. A stronger architecture keeps evidence types distinct and combines them only under explicit rules.
What this section claims
- Different collection methods have different epistemic and consent properties.
- A context system benefits from explicit source typing rather than one undifferentiated profile.
What it does not claim
- This section does not claim that current industry methods are useless.
- It does not claim an information-theoretic impossibility proof.
- It does not claim that more compute can never improve inference quality.
Review discipline
Keep provenance, maturity and validation separate.
Phase 1 provides team-built product and market context. Phase 2 contains the bounded solo AI-native formation work. Phase 3 is where independent technical, legal/IP, compliance, pilot and commercial review decides what survives professional diligence.