A summary is not knowledge: How AI decides whether an idea should be saved, researched, or turned into a product

The problem I wanted to solve sounded simple. Videos, interviews, articles, and documents contain useful ideas. Could AI read them at scale and keep the parts that fit KOF?

The easiest output is a summary. Each source becomes a few smooth takeaways that preserve the general meaning. As the summaries accumulate, however, the real questions simply move into a new pile. Is this claim credible? Why does it matter to KOF? Should it be saved or investigated? Can it support an article, or does it describe a problem that deserves a product?

A database full of summaries is closer to a filing cabinet. A useful brain must also judge how mature an idea is and assign its next step.

The same idea can have five destinations

Suppose I receive one sentence: “AI can turn expert intuition into rules.” It may be provocative, but sounding plausible does not determine where it belongs.

If the source is clear and the idea matters to KOF but needs no action now, save it. If important uncertainty remains and there is a way to resolve it, research it. It becomes an article only when it answers a public question and KOF has an original analysis to add. It becomes a product experiment only when the problem repeats, the user is identifiable, the value is measurable, and the process can be reproduced. If the source and boundaries cannot be established, discard it.

Discard: untraceable, irrelevant, repetitive, or unacceptably risky
Save: worth remembering, with no current action required
Research: important uncertainty with a method for resolving it
Article: a public question with original KOF analysis and boundaries
Product: a recurring job with a clear user, value, and validation method

These destinations are not ranks. A saved item is not below a product, and discarding material does not mean it was poorly made. Routing keeps an idea in the state justified by its current evidence and maturity.

Extract claims before summarizing the whole piece

One passage can mix observations, experience, interpretation, and prediction. Compressing it into three takeaways can erase those differences. A safer starting point is to extract claims that can be checked independently.

Fact: verifiable through external evidence or direct observation
Viewpoint: the speaker’s judgment, preference, or interpretation
Hypothesis: unproven but testable
Proposal: a recommended action that still needs cost and authority checks

“The company released three products this year” is a verifiable fact. “Its speed comes from its culture” is an interpretation. “The same method works for independent developers” is a hypothesis. “KOF should build a tool” is a proposal. All four may appear in the same interview, but they require different evidence.

Bind the source to the claim

A list of links at the bottom of a note is not enough. Every claim should retain who made it, where it appeared, whether it came from direct observation or retelling, and whether later evidence contradicts it. Sources may remain internal rather than appearing in every public article, but they cannot disappear during summarization.

The most dangerous model behavior is completing a gap with a plausible answer. The routing system must allow unknown values, and those unknowns must be able to block a claim from becoming a published conclusion or product requirement.

Claim: a statement that can be checked independently
Type: fact / viewpoint / hypothesis / proposal
Source: origin and location
Evidence for: current supporting material
Evidence against: counterexamples or contradictions
Unknowns: what is still missing
Owner: who is accountable for the judgment
Status: save / research / article / product / discard

Do not give a polished score to something that failed a hard gate

AI can score content easily, but scores can create false precision. A claim with no source does not become a fact because it received 87 points. A popular topic that has no relationship to KOF does not become the right product because the market is large.

Routing should apply hard gates before priority scores. Can the source be traced? Is the claim type explicit? Does it connect to work KOF actually does? Are the largest unknowns and risks visible? If any answer is missing, the next step is usually more evidence or rejection, not more model reasoning time.

The decision turns on what KOF can add

Valuable material does not automatically become valuable when KOF repeats it. The important question is whether we can add something the original material did not contain.

That addition may be a first-party experiment, a reusable template, a counterexample, executable software, or a decision framework that connects scattered ideas. If the only contribution is paraphrasing, saving the source is often the better decision.

Five questions help with priority: Does it fit a problem KOF is already working on? Is it new relative to existing content? Can a reader take a useful next step? Is the evidence strong enough for the claim? Can KOF test it through its own work or products? These questions support comparison, but they do not replace the hard gates.

What each route requires

Saving requires traceability. Keep the original material, why it matters, where it applies, and the event that should trigger another review. A note with no revisit condition easily becomes a collection item that never opens again.

Research requires an answerable question. “Read more” is not a research plan. “Test the same rule in different situations and see whether it reproduces the original decisions” has a stopping condition. Evidence that rejects the original idea is still a useful research result.

An article requires public value. It should answer a problem readers encounter, separate evidence from the author’s judgment, mark what remains unverified, and add an original KOF contribution. Many sources without a new decision only produce a digest.

A product has the highest gate. The problem must recur, an identifiable group must be willing to change an existing behavior, success must be observable, and the manual process should already be understandable. Article interest proves that people care about a topic; it does not prove that they need another app.

Make the idea classify itself

Return to the earlier claim: “AI can turn expert intuition into rules.” Under this framework, it is currently a hypothesis rather than a fact.

It fits KOF’s knowledge workflow and can produce interview templates, decision tables, and validation methods, so it deserves to be saved and researched. It can also support a methodology article if the article admits that cross-context validation is unfinished. A product is premature because repeat use, real user response, and the boundaries of what cannot be formalized remain unknown.

Current type: hypothesis
KOF relevance: high
Possible KOF contribution: routing framework, validation format, counterexamples
Main unknowns: cross-context reliability and model-invented values
Current routes: save + research + article
Gate not passed: product

This classification does not make the idea sound more impressive. It prevents a smoothly written article from being mistaken for evidence that a product already works.

AI can propose; it should not approve itself

AI is good at reading in volume, extracting claims, finding related material, surfacing contradictions, generating missing questions, and suggesting routes under an established policy. Those tasks can reduce a large amount of clerical work.

Required fields, rejection conditions, risk limits, write permissions, and publication decisions belong outside the model. People should still approve high-risk or irreversible actions. If the same AI produces an answer and declares that answer trustworthy, the workflow has no independent checkpoint.

AI proposes: claims, types, evidence, counterexamples, suggested route
System enforces: required sources, unknowns, risk rules, state transitions
People decide: official records, public claims, research investment, product development

A good brain is not measured by its note count

The value of a knowledge system is not the number of summaries it stores. Every important idea should have a clear state: why it was kept, how far we currently trust it, what evidence is missing, and who owns the next action.

A summary makes content shorter. Routing tells an idea where it should go next. That is the job I want AI to perform inside the KOF brain.