Skip to Content
Getting StartedConceptsHonest Numbers

Honest numbers

Morse AI would rather tell you it doesn’t know than show you a number it made up. Once you learn to read its silences, the whole product becomes more trustworthy — an empty cell is information, not a gap. Three conventions carry this.

An em-dash means “nothing measured” — not zero

Where a value was never captured, Morse AI renders an em-dash (—), never $0 or 0. The distinction is deliberate and pervasive: a null renders the dash; a real zero is a real number.

So “cost: —” means we have no priced cost for this run, which is a very different thing from “this run was free.” Reading the dash as zero is the one misread the convention exists to prevent.

Gate lines say why a panel is empty

When a panel can’t show a verdict yet, it doesn’t blank out — it tells you the reason, in words. You will see lines like:

  • “Not instrumented — no LLM span data for this agent in this window.”
  • “Insufficient data. Need at least 50 traces in each arm before a memory verdict.”
  • “No fleet infrastructure cost yet — hosted-only workloads won’t show a fleet here, and we never fabricate a number when there’s nothing to measure.”

A gate line distinguishes the two reasons a panel might be quiet: not instrumented (the data was never sent) versus insufficient data (it’s coming, but there isn’t enough yet to be honest). The account lifecycle is what decides which one you see.

Estimates are walled out of the totals

Every cost is tagged with a provenance: provider_reported (exact, from the provider’s own billing), estimated, or unpriced. Only spans that are actually priced contribute to a trace’s total — unpriced spans are counted for a coverage score but never summed into the number you budget against.

Infrastructure estimates get the same treatment: an estimated per-trace infra share is stamped estimated and must never be summed into the priced total or into any cost-per-successful-outcome denominator. Estimated and exact never mix in the headline figure; approximations surface separately, clearly labeled.

This is what makes the Coverage Score meaningful: it tells you how much of your spend is exact versus estimated, so you know exactly how far to trust the total — instead of taking one blended number on faith.

Where Morse withholds a verdict

The same rule — say so rather than guess — appears as named states across the product:

  • Baseline forming — an agent without enough history stays neutral. The gate line explains the evidence it needs. See Lifecycle.
  • “—” instead of $0 — cost per successful outcome shows a dash when no success has been confirmed. See Cost.
  • Not enough labeled outcomes — the weekly brief withholds the success rate and verdict while showing coverage. Unlabeled runs are not failures. See Dashboard.
  • Inconclusive — a regression comparison below its required sample is labeled directional-only. P-values are corrected across scorers. See Regression.
  • n/a — scorers needing context or a span tree cannot score a frozen input/output replay. Missing scores are not fabricated. Review your policy’s readiness before relying on a build gate. See CI Gate.
  • Needs review — a completed investigation without a high-confidence, nonempty hypothesis needs a human check. The confidence is the model’s self-assessment, not a calibrated probability or a verified cause. See Investigations.
  • Cost Intelligence — where provenance and the Coverage Score are surfaced.
  • Lifecycle — why a panel says “insufficient data” and when that clears.
  • Data model — priced vs unpriced spans, by type.
Last updated on
Data ModelLifecycle