Skip to content

Evidence

Every claim, and its measurement

Method

Lodestar keeps an append-only log of how it is used, on the Mac it runs on: which gesture, when, how long from the key going down to the destination arriving, and what was corrected. It never keeps content: no window titles, no addresses, no text. The same file the coach reads is the file these numbers were read from.

The sample is one hand, the author's, over fourteen days in August 2026: between seventy and seventy-five thousand events. The typing and pointing minutes in the first result come from two days of the health pulse and are estimates from measured counts; navigation is timed directly. One day with a logging artefact is excluded from rhythm figures. Result 10 is a published study, not this hand.

A sample of one says nothing about people in general. It says a great deal about whether the mechanisms Lodestar is built on are real, because each was measured on the hand that built them before it shipped.

  1. No. 01link

    Window navigation is a small share of the day, even for a practised hand.

    Measurement
    On one fully instrumented day of 280 active minutes: typing 70.5 min, pointing 17.5 min, window navigation 2.6 min. Navigation is 2.8% of input time. Backspaces alone, at a 13.5% rate, cost 9.6 min, 3.7 times the navigation budget.
    Method
    Health pulse, one row per quarter hour of keystroke, click, and scroll counts, with no key identities; navigation timed directly from the gesture log.
    Reading
    The navigation problem is finished for this hand. What remains to win is elsewhere, which is why the coach is built to look elsewhere.
  2. No. 02link

    Practice does not make a gesture faster. Giving it a shorter address does.

    Measurement
    A power law of practice fitted to each of 15 chains (98.4% of gestures) gives a mean exponent of −0.020 and a median R² of 0.0025: flat. The one accepted shortening on August 23 moved the median gesture from 152.8 ms to 97.0 ms (p = 0.0005), and no individual chain improved. A newly created bare letter ran 76 ms in its first 189 uses, faster than a chain with 3,798.
    Method
    Time from lode down to the last key, per chain, over the fourteen days; the shortening compared as before and after with a two-sample test.
    Reading
    Fluency is encoded, not practised. This is why the coach offers single letters first and why it will offer to flatten a chain that keeps stumbling.
  3. No. 03link

    A bare letter under lode misfires at zero. A chain does not.

    Measurement
    Failure to complete the intended gesture: 0.00% for bare letters, 7.9% for chains under b, 11.2% under e, 41.1% under v (n = 124).
    Method
    A start counted as a misfire when the chain was abandoned or corrected within the same hold.
    Reading
    Depth is a tax paid in errors. The coach's flatten offer reads exactly this gradient.
  4. No. 04link

    The hand knows the whole chord before the first key.

    Measurement
    Timed only on the first keypress, the same letter b took 73.4 ms when the chain was bx, 141.9 ms for bd, and 173.1 ms for bg.
    Method
    First-key latency from lode down, grouped by the chain that followed.
    Reading
    The suffix decides the speed of the prefix. Gestures are motor programs, not sequences of decisions, once they are learned.
  5. No. 05link

    Most visits to an application are glances.

    Measurement
    Dwell time is a two-component mixture (ΔBIC −2753): 42.6% of visits are glances averaging 1.14 s, 57.4% are work averaging 18.5 s. 44% of visits account for 2.2% of foreground time.
    Method
    Foreground durations from focus events, fitted as a mixture of two log-normal components against one.
    Reading
    A navigation that is priced by visits overprices glances. The coach prices by the seconds a change would return.
  6. No. 06link

    Where you go next is partly predictable, and one address would cover half of it.

    Measurement
    The predictability ceiling of the transition sequence is 79.9%; a first-order model reaches 46.3%. A single back address would cover 49.2% of switches, and the three most recent applications cover 81.3%.
    Method
    Fano bound and a Markov model over application-to-application transitions.
    Reading
    This is the case for the launcher ranking by recency, and the reason a back gesture is on the list.
  7. No. 07link

    Selecting text by typing it is bound by keystrokes, not by choosing.

    Measurement
    Time to a selection scales with keys typed (R² 0.49) and barely with the number of candidates (R² 0.13), at about 410 ms per key.
    Method
    Regression of selection time on query length and on candidate count over the select gesture log.
    Reading
    The way to make select faster is fewer keys, not smarter ranking, which decided its floor of two characters.
  8. No. 08link

    Arranging two windows side by side dropped from seconds to under one.

    Measurement
    Before the chord, a beside arrangement took 2.1 s median from the first summon to the split standing. On the chord's first day, the whole phrase took 757 ms median, and joins fell from about 1 s to under 200 ms within twenty minutes of use.
    Method
    Event timestamps from lode down to the layout settling, matched to the graph chains within 1.5 s; screen-locked intervals excluded.
    Reading
    Eliminating a navigation beats accelerating one, which is the whole case for layouts as destinations.
  9. No. 09link

    A drawn selection needs a second opinion.

    Measurement
    Of 69 verified copies over two days, the application was asked to confirm the span 50 times: 22 approved, 17 overruled, 11 gave no answer. Twelve of the seventeen overrules were one terminal that takes a press without a position, fixed the next day.
    Method
    The copy ledger: every held span with the app's own selection compared to the pixels, up to the confusable glyphs.
    Reading
    Pixels aim and the app confirms. Neither is trusted alone, and the ledger says why.
  10. No. 10link

    Closing the old path is what makes a hand switch. A nudge does not.

    Measurement
    In Grossman, Dragicevic and Balakrishnan (CHI 2007, n = 42), expert-path use was 28.9% when both paths stayed open and 72.8% when the slow one was disabled. Seven of fourteen control participants never used the fast path at all. A purely visual cue tested below control.
    Method
    Published between-subjects study; not this hand.
    Reading
    This is the evidence behind the coach reshaping the road on accept rather than reminding, and behind pricing the escape hatch in one deliberate confirm.

References

  • Grossman, T., Dragicevic, P., and Balakrishnan, R. Strategies for accelerating on-line learning of hotkeys. CHI 2007.
  • Lally, P., van Jaarsveld, C., Potts, H., and Wardle, J. How are habits formed: modelling habit formation in the real world. European Journal of Social Psychology, 2010.
  • The instrument itself: the observation layer is documented in the repository's design specification and its platform findings in the engineering ledger.