KLM: why unknown is not zero

A recorded 0 and an unavailable measurement are different statements. Zero supplies a value. Unknown says there is not enough information to supply one.

Language-model systems can lose that distinction easily. A source has no date, code substitutes zero for a missing freshness score, and a dashboard presents the number as a measurement. Missing evidence has become an apparent low score.

KLM's honest-null rule aims to preserve the distinction: an unavailable signal remains null, with an explicit reason.

What zero cannot say

If freshness was not measured, a zero freshness score does not mean that the source was measured to be old. The source's effective date still needs investigation.

This excerpt illustrates a repository example. It is not a complete KLM record:

{
  "freshness_score": null,
  "signal_status": {
    "freshness_score": "unavailable"
  },
  "nulls": {
    "freshness_score": {
      "reason": "no_varve_age_metadata"
    }
  }
}

The system avoids reporting a measurement it never made. The reason leaves a trace that can be investigated later.

Unknown can disappear at a boundary

A JSON record may preserve null while another component erases it. A default such as value ?? 0 replaces an explicit unknown with zero. Some wire formats cannot distinguish absence from a scalar's default value without additional metadata.

KLM therefore requires the distinction to survive construction, encoding and decoding boundaries that the implementation controls. An invalid signal is also different from an absent one. A malformed or out-of-range value must not silently become “unknown.”

Different numbers carry different kinds of knowledge

KLM distinguishes four statuses:

• measured: deterministically computed from a recorded event and replayable.
• heuristic: an approximation based on observable signals, still replayable.
• synthesized: produced through model or evaluator interpretation.
• unavailable: not produced; represented by null and a reason.

A model's declared confidence is separate from confidence supported by evidence. Model-based assessment can enrich a record, but it cannot replace the mechanical floor required for a conformance claim.

The limit of an honest record

An honest unknown does not make an answer true. It preserves a gap instead of concealing it. KLM conformance concerns record structure and traceability, not a general guarantee of factual correctness.

Reading a system therefore requires more than comparing scores. It requires knowing how each score was produced, what status it carries and what happens when it is unavailable.

See the KLM project page and the validation guide for the record and toolchain behind this distinction.

Sources and revision

This article is based on the public repository inspected on 9 October 2026. The links below are pinned to that revision; the project may subsequently change.

• Specification §4.5
• Honest-null example