Skip to content

Observer value-channel window is a fixed 10 samples: a slow mode under a fast cadence reads 'diverging', confidently #1044

Description

@InauguralPhysicist

Found by phugoid rung 1 (2026-08-24), the first external grading of observer verdicts against a mode-structure oracle (phugoid GAPS.md G4, tests/observer_check.eigs).

The value-channel predicates classify over a window of the last 10 observed samples (docs/PREDICATES.md), with no per-binding control. A mode slower than ~10 samples of the consumer's cadence cannot fold inside the window, and the failure is not mere insensitivity — the verdict is confidently wrong:

  • The 747 phugoid (T = 46.9 s, physics truth oscillating throughout) replayed through an observed binding at 1 Hz reads diverging on its rising quarter-cycles, stable/improving elsewhere — measured at three probe points, all wrong.
  • The same signal decimated to 5 s cadence (9.4 samples/cycle) reads oscillating correctly at every probe point.
  • dynamics already documents the entropy-channel flavor of cadence sensitivity (physics.eigs header: sampling every step makes everything read equilibrium); the value-channel version is sharper because diverging is an alarm verdict a consumer would act on (stall warning, conflict detection).

A frame-locked consumer — the flight-sim thesis: live verdicts at a frame deadline — cannot decimate per binding without hand-building a shadow binding per timescale, which is exactly the multi-timescale gray band (#187 lineage) the flight-sim proposal predicted would break first.

Question for the design: should the window be per-binding configurable (samples, or better, seconds — flight modes are specified in time units), or should a slow-fold detector widen it adaptively? The oscillating folding rule (>= 2 reversals, net travel <= 0.3x path) already almost expresses the right idea; it just never sees a fold at this cadence.

Repro (phugoid repo, branch rung-1-longitudinal): eigenscript tests/observer_check.eigs — the three O.ph1s.* rows are pinned as class divergence and will flip loudly the day the windowing changes, which is phugoid's re-grade signal.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions