Temple of Two
← Instruments

Study + instrument

Phenomenological Compass

A two-stage inference architecture: a small compass reads the epistemic weight of a question and conditions a larger action model.

Study + instrumentPublishedSupported within scope

96% signal accuracy · ΔH = +0.47 nats · one architecture family, one benchmark

The question
Does naming the posture a question deserves change what the answering model actually computes — or only how it talks?
The instrument
A 1.5B LoRA compass reads the geometry, tone, and epistemic weight of a question and issues OPEN / PAUSE / WITNESS to condition a larger action model. Ships as a native macOS app, fully local on Apple Silicon.
The finding
It changes the computation. Token-level entropy profiling shows ΔH = +0.47 nats (JSD = 0.076): the compass measurably restructures the probability distribution rather than restyling the output. Ablation confirms signal-specificity — wrong-signal WITNESS conditioning collapses responses, full wins 31–2.
What failed
The original framing treated the compass as a filter on outputs. That was wrong in a checkable way: a filter cannot move token-level entropy before the tokens exist. The mechanism is field conditioning through attention geometry, which is a different claim than the one first made.
Standing
Study + instrumentPublishedSupported within scope
Verify it
DOI 10.5281/zenodo.19377144 ↗Repository ↗

Metrics with scope

Every number names what was measured, on what sample, and as of when.

  • Signal accuracy

    96%

    800-question HumaneBench, compass signal vs. held-out labels

    As of 2026-04-02 · source

  • Entropy restructuring

    +0.47 nats

    Token-level ΔH, JSD = 0.076, one architecture family

    As of 2026-04-02 · source

  • Ablation

    31–2

    Full-signal vs. wrong-signal WITNESS conditioning, head-to-head

    As of 2026-04-02

From wonder to receipt

Phenomenological Compass

Wonder initiated the search. Rigor changed the answer.

Study + instrumentPublishedSupported within scope
  1. 01Wonder

    Does it matter, to the machine, whether a question is asked with reverence or with haste?

  2. 02Hypothesis

    If posture is real to the model, conditioning on an explicit epistemic signal should change the shape of the probability distribution — not just the wording of the reply.

  3. 03Instrument

    A 1.5B LoRA compass emitting OPEN / PAUSE / WITNESS, wired ahead of a larger action model, with token-level entropy profiling on both paths.

  4. 04Encounter

    ΔH = +0.47 nats, JSD = 0.076. The distribution moved. Wrong-signal conditioning collapsed the response, so the effect is signal-specific rather than a generic prompt-length artifact.

  5. 05Standing

    Supported within scope: one architecture family, one benchmark, not yet replicated across model scales.

  6. 06Receipt

    DOI 10.5281/zenodo.19377144 · phenomenological-compass

DOI 10.5281/zenodo.19377144 ↗Repository ↗