From Data to Knowledge

AI systems are often described as if they perform one mysterious step:

Input → AI → Answer

That is not how Spillway is being built.

Spillway separates several different kinds of information because each deserves a different level of trust, a different method of interpretation, and a different role in the final decision.

Raw artifacts
    ↓
Observations and facts
    ↓
Evidence
    ↓
Knowledge
    ↓
Suggestions and decisions

These layers are related, but they are not interchangeable.

Raw artifacts

Raw artifacts are the material the system can inspect:

message bodies and headers;
sender and recipient addresses;
attachments;
URLs and link text;
calendar invitations;
thread history;
dates and times;
prior user actions.

At this layer, Spillway has not yet decided what anything means.

An attached PDF is simply an attached PDF. A date is a date. A link is a link.

Keeping this layer distinct matters because later interpretations should remain traceable to the material that produced them.

Observations and facts

Many useful things can be established deterministically.

For example:

a message contains a PDF attachment;
the sender address belongs to a particular domain;
a calendar invitation includes a Zoom link;
a date appears in the body text;
a URL points to a known survey platform;
the user previously filed messages from this sender into a particular category.

These observations do not require a language model to rediscover them.

When a fact can be established reliably through parsing, lookup, or a stable rule, Spillway should do that work directly. Deterministic processing is faster, easier to test, easier to explain, and less likely to vary unpredictably.

Evidence

Facts rarely determine the final meaning of a message by themselves.

A PDF attachment could be an invoice, an agenda, a paper to review, or a receipt. A date could describe a deadline, a meeting, a historical event, or the day something was sent. A question may or may not require a reply.

Evidence is information that supports or weakens an interpretation.

Examples include:

nearby language that makes a date look like a deadline;
a familiar sender combined with an attached manuscript;
an unanswered question in a thread;
a survey link paired with language asking for participation;
repeated user corrections that suggest a category rule is unreliable;
several independent signals pointing toward the same conclusion.

Evidence has provenance, scope, and strength. It may conflict with other evidence. It may become stale. It may support a conclusion without proving it.

But provenance alone is not enough. Evidence also needs weight.

Two observations can come from equally well-documented sources and still deserve very different influence. Spillway may eventually consider factors such as:

reliability — how often this source, model, or rule has been right for this kind of judgment;
certainty — how strongly the observation supports what it claims rather than merely suggesting it;
specificity — whether the evidence applies to this exact message, a sender, a domain, or only a broad class of cases;
recency — whether the evidence still reflects the user's current projects, relationships, and priorities;
independence — whether several signals genuinely corroborate one another or repeat the same underlying clue;
authority — whether the evidence came from a deterministic observation, a model proposal, a recurring pattern, or an explicit human correction;
contradiction — whether strong counterevidence points in another direction.

A human-confirmed category may deserve more weight than one model suggestion. A deterministic finding may be nearly certain about a narrow fact—such as the presence of a Zoom link—without settling the broader question of whether the message is important. Ten repeated model outputs may add little if they all depend on the same mistaken premise.

This idea has parallels in research on human judgment. Work by Josh Pasek and colleagues has shown that an incorrect answer held tentatively is not the same phenomenon as a confidently held misperception, and more recent work models how belief certainty changes the influence a belief has on an overall attitude. The relevant lesson for Spillway is not that human cognition supplies a ready-made software formula. It is that what is believed, how confidently it is supported, and how much influence it should receive are separate questions.

That distinction is central to Spillway: model output is evidence, not automatic truth, and evidence should not all count equally.

Knowledge

Knowledge is the system's current, revisable understanding after evidence has been accumulated and interpreted.

Examples might include:

this sender is usually associated with a particular project;
this recurring message is probably a newsletter;
this institution is important to the user;
this type of attachment usually requires review;
this category assignment is strongly supported;
this old pattern should now receive less weight.

Knowledge is not permanent merely because it was once useful.

People change roles. Projects end. Mailing lists change purpose. Old evidence may become less relevant. New evidence may contradict an earlier interpretation.

A trustworthy system therefore needs to preserve uncertainty, allow correction, and revise what it believes over time.

The same person communicates differently in different places

Evidence also depends on where and how it was produced.

People do not express one context-free version of themselves. The same person may answer a survey question carefully, post a short reaction on social media, write a formal email, make a tentative comment in a meeting, or take an action without explaining it at all.

Those expressions are not interchangeable measurements.

Research comparing Twitter activity with survey responses illustrates the point. Josh Pasek and colleagues found that Twitter sentiment about a president was not a simple proxy for presidential approval measured in surveys. The two sources could tell broadly similar long-term stories while diverging sharply at more granular levels and over shorter periods. The lesson is not that one source is real and the other is noise. Each reflects a different communication environment, population, behavior, and measurement process.

That matters as Spillway eventually learns from more than email. A calendar acceptance, an email reply, a document edit, a task completion, and a chat message may all concern the same project, but they reveal different kinds of information. The system should preserve those source differences rather than flatten every trace into one undifferentiated score.

Suggestions and decisions

The final layer is action:

suggest a category;
surface a possible task;
highlight a date;
identify something that may need a response;
rank work for attention;
explain why a suggestion appeared.

These outcomes should be based on the available facts, weighted evidence, and revisable knowledge—not on one opaque model response treated as authoritative.

Spillway also distinguishes a machine suggestion from a human decision. A person confirming, correcting, or dismissing a suggestion changes the authority of that information.

Why some work should be deterministic

A language model can often infer that meet.zoom.us is associated with Zoom. But asking the model to rediscover that fact every time is wasteful and less reliable than a deterministic lookup.

The same principle applies to:

file types;
known domains and services;
URL structures;
calendar metadata;
sender and thread direction;
exact dates and identifiers;
explicit user preferences.

When something can be known directly, the system should know it directly.

Why fixed rules are not enough

Other questions are too context-dependent for a practical fixed-rule system:

Is this request genuinely important?
Does this date function as a deadline?
Is the sender asking for action or merely sharing information?
Does a document require review, filing, or no response?
Is this message part of an existing project?
Which of several plausible interpretations best fits the whole conversation?

It is possible to write rules for fragments of these problems, but the number of exceptions and interactions quickly becomes unmanageable.

This is where AI is useful: not as a replacement for facts, but as a tool for interpreting ambiguous evidence and combining context that would be impractical to encode exhaustively.

The design principle

Spillway's goal is not to choose between deterministic software and AI.

It is to use each where it is strongest:

deterministic systems establish stable facts;
evidence systems preserve provenance, weight, conflict, and uncertainty;
AI helps interpret ambiguity and synthesize context;
human judgment remains authoritative for consequential decisions.

That layered approach is slower to design than simply sending everything to a model. But it produces a system that can be tested, corrected, explained, and trusted.