Skip to content

From Data to Knowledge

AI systems are often described as one mysterious step:

Input → AI → Answer

Spillway uses a more structured model because source material, observations, interpretations, and user decisions should not collapse into one thing.

At a high level:

Artifact → Observation → Evidence → Knowledge

That is enough structure to keep the system understandable without requiring users to learn the internal architecture.

Artifact

An artifact is source material the system can inspect: a message, attachment, URL, calendar item, document, or similar input.

Artifacts matter because later interpretations should still be able to point back to what they were based on.

Observation

An observation is something Spillway can establish directly from the artifact or another trusted source.

Examples include:

  • a sender address;
  • the presence of a PDF attachment;
  • a URL host;
  • a date appearing in the text; or
  • a literal identifier in a reference source.

When ordinary software can establish a fact reliably, Spillway should prefer that over asking a language model to guess it.

Evidence

Evidence is information that bears on an interpretation.

A date is an observation. Whether it is a deadline is an interpretation supported by evidence.

A sender's relationship to a project may support one category without proving every future message from that sender belongs there.

Evidence retains provenance so Spillway can distinguish where it came from and whether several apparent signals are actually independent or are copies of the same underlying source.

Different properties also answer different questions. A fact can be highly certain but only weakly relevant to a decision. A human correction can carry special authority without making every related inference certain.

Knowledge

Knowledge is a durable, revisable representation built from evidence.

It can help Spillway remember useful structure: recurring relationships, domain knowledge, project associations, or other context that should not have to be rediscovered from scratch each time.

Knowledge can be updated as evidence changes. The goal is not to freeze a conclusion forever, but to preserve enough basis and provenance that later reasoning can revise it responsibly.

Memory and context

Some information is useful because it changes how other evidence should be interpreted rather than because it is another fact to add to a pile.

Spillway treats that kind of remembered context separately so it can condition later reasoning without pretending it is an independent observation.

The same principle applies when building prompts for AI: the goal is to provide relevant context, not the largest possible context window.

Suggestions and human decisions

Knowledge can support suggestions: a category, a possible task, a date, an opportunity, or an explanation.

A suggestion is still not the same thing as a human decision.

When a person confirms, corrects, rejects, or later changes something, Spillway should preserve that action as its own evidence with its own provenance. The system should not rewrite the history so that an earlier model appears to have "known" what the person later decided.

Derived views are rebuildable

Many useful product surfaces are projections over underlying information: focused views, work lists, attention queues, and other summaries.

Those views should be rebuildable from the evidence and user-authored state beneath them. That makes it safer to improve the interface or reasoning without turning each projection into another source of truth.

Why this separation matters

Keeping these layers distinct supports:

  • explainability: important suggestions can point back to their basis;
  • correction: the system can distinguish a bad observation from a bad interpretation;
  • rebuildability: derived views and predictions can change without rewriting source history;
  • privacy boundaries: useful abstractions can sometimes move without copying all source content;
  • learning: later reasoning can use prior evidence without silently inventing new authority.

Deterministic reasoning and AI have different jobs

Stable structure is usually best handled deterministically. Ambiguous interpretation is where AI is useful.

A parser may establish that a date exists. A model may help decide whether that date is a deadline. A person may decide whether the deadline matters enough to act on.

Those are different roles, and Spillway is designed to preserve the difference.

Why deterministic systems still matter · Explainable AI · From Evidence to Action

Documentation provenance: Iterative human–AI construction. See Documentation Provenance.