Sandbox-only research artifact

When the AI says “just a sec,” what does that actually mean?

A calibration layer that maps LLM time-claims — “a moment,” “quick check,” “let me think” — to the wall-clock seconds they tend to consume. Trained on completed-task logs you keep locally. Gated by accuracy. Honest about its priors.

Lexicon Lens

Paste real AI output. The Lens finds every time-claim, highlights it, and stacks them into one wall-clock distribution.

⌘+Enter to analyse 0 phrases detected

Annotated message will appear here.

Wall-clock probability density. Log-scale x-axis from 1 second to 1 hour.

Log a completed task

Add one row each time an AI promised a duration and you timed the reality. The model recalibrates per category.

Detected category —

Recent outcomes

0 logged
  1. Nothing logged yet. to feel the model move.

Calibration by category

Each phrase family has its own prior and its own gate. Below the gate, output is marked illustrative.

    How the Translator works

    Plain-English methodology — no dark patterns, no hidden state.

    1. Phrase fingerprinting

    A small dictionary of AI hedges — moment, second, quick, instantly, let me think — plus explicit estimates like “5 minutes”. Each phrase resolves to a category with its own prior.

    2. Empirical Bayes per category

    Default priors are log-normal, derived from public anecdotal ranges. As you log outcomes, observed samples take over via a weighted mix — the prior fades as n grows.

    3. Accuracy gate

    A category is “calibrated” only when you have at least 5 samples and the log-standard-deviation falls under 0.9. Below that, output is rendered with an illustrative badge.

    4. Cumulative composition

    The Lens treats each detected phrase as an independent latency draw, then sums their medians and propagates variance to estimate the cumulative range. Sequential, not parallel.

    5. Local-only memory

    Everything you log stays in this browser via localStorage. No network calls, no analytics, no accounts. Hit Reset data any time.

    What this isn't

    Not a performance benchmark. Not a model evaluation. Not advice you should act on. A pedagogical instrument for thinking about AI time-talk — nothing more.