Cucumbs demo

Cucumbs / demo / part 4 of 6: the claims

Fact, guess or unknown

Every claim that moves between agents carries a label saying how it is known. The label travels with the claim, so nobody further along mistakes a rumor for a checked fact.

Layer 4: the claims. Every message has an EVIDENCE field. Here the claims from the rent letter and the laptop search come out one at a time, and each gets a label.

The six labels

Each label answers one question: how do we know this? The examples come from the sample tasks used on the loop page.

  • FACT

    Checked against a primary source that anyone can look at, and it held up.

    Example: Section 4 of the lease requires 60 days' written notice for a rent increase.

    Can it change? Only if the source turns out to be wrong or out of date.

  • OBSERVED

    Seen first-hand by an agent during the task, at a particular time.

    Example: The seller's page showed $849 at 9:40 this morning.

    Can it change? Yes. It is true of that moment, and the next look may differ.

  • REPORTED

    Someone else says so. The team hasn't checked it.

    Example: A review site says the battery lasted 10 hours in its test.

    Can it change? It becomes a FACT only when someone checks the original source.

  • INFERRED

    Follows from other evidence by a step of reasoning.

    Example: The letter came 45 days ahead, so the notice is shorter than the lease allows.

    Can it change? It is only as strong as the evidence and the reasoning under it.

  • HYPOTHESIS

    A guess worth testing. Nothing supports it yet.

    Example: The landlord might agree to a later start date if asked.

    Can it change? Testing it turns it into one of the other labels.

  • UNKNOWN

    Nobody knows yet, and the message says so.

    Example: Whether the city caps rent increases.

    Can it change? Yes, once someone finds out. Until then it stays visible.

Sort the claims

Ten claims from the sample tasks. Pick a label for each one and you'll see straight away whether it matches, and why. Some are close calls, and the feedback says when a second label is defensible.

The sorter needs JavaScript. Here are the same claims, with the label hidden until you open it.

  1. The lease, section 4: rent increases need 60 days' written notice.

    Show the label

    FACT It was read in the lease itself, and anyone with the lease can check it.

  2. The seller's page showed model A at $849 at 9:40 this morning.

    Show the label

    OBSERVED An agent saw it first-hand at a set time. The price may be different now.

  3. A review site says model A's battery lasted 10 hours in its test.

    Show the label

    REPORTED The team didn't run the test. Someone else says so.

  4. The letter came 45 days before the increase, so the notice is shorter than the lease requires.

    Show the label

    INFERRED It is reasoned from two other pieces of evidence: the letter's date and the lease clause.

  5. The landlord might accept a later start date if asked.

    Show the label

    HYPOTHESIS Plausible, and worth asking, but nothing supports it yet.

  6. Whether the city caps rent increases.

    Show the label

    UNKNOWN Nobody has found an answer yet, so the message says so.

  7. A tenant forum post says increases above 5% need the city's approval.

    Show the label

    REPORTED It is one second-hand source. It stays REPORTED until someone reads the actual rule.

  8. Model B costs less than model A at both sellers checked this morning.

    Show the label

    INFERRED Each price was observed, but comparing them is a step of reasoning, so INFERRED fits best.

  9. The maker measured its 14-hour battery figure with the screen dimmed.

    Show the label

    HYPOTHESIS It is a guess about how the maker tested. Nobody has found the test conditions.

  10. How much model A's battery will wear down after two years.

    Show the label

    UNKNOWN No source the team found answers it. Saying so is better than guessing.

Rules of thumb

  • A label moves toward FACT only when someone checks a source. Repeating a claim doesn't make it stronger, however many agents repeat it.
  • An OBSERVED claim is tied to a moment. Write the time down with it.
  • UNKNOWN is a useful answer. It shows the person where the gaps are.
  • When two agents disagree, both claims keep their labels and both go forward.

Where the labels change

Labels change at the compare and reconcile stages of the loop, when results are checked against their sources. Sentinel does most of that checking. A label can also go down: a FACT whose source turns out to be out of date goes back to REPORTED or UNKNOWN.

By the time work reaches the approval gate, the person can see which parts of the draft rest on checked facts and which rest on someone's word.

Go deeper: the first axiom, seen from here (layer 4 to 6)

This is Sol's reading, not part of the axioms. The first axiom says truth is uncovered, not created. The labels are a working version of that: a claim moves toward FACT only when someone uncovers its source, and repeating it doesn't count.

The axioms themselves are at the root, on the last page.