# 0002 — Researcher rulings on the engineer's five questions

2026-08-21. The researcher's answers, verbatim, then the design
consequences extracted from each. Per §110 the native concepts are
preserved; formalizations below are candidates adjacent to them, not
translations replacing them.

## 1. Layer-2 encoding

> "my default is that anything has multiple descriptions. words, shapes,
> images, vectors, math, sequence, time. what we really dont want to do
> is compress it all into the same LANGUAGE aka language. this has to
> hold many types"

**Ruling**: the dilemma (fixed vocabulary vs freeform relations) is
dissolved by rejecting its premise — there is no single Layer-2 scheme.
**Consequence**: `representation.content` is scheme-TYPED, and the store
holds many *types* (graph, sequence, geometric, numeric, prose...) of the
same artifact simultaneously. Prose is one type among several, never the
canonical one (§74 hardened from warning to rule). Comparators are
scheme-pair objects; the *absence* of a comparator between two schemes is
a recorded fact, not a bug to fix. Experiment 0 changes: parses must span
at least two genuinely different types, not two readings within one type.

## 2. Parser variance vs artifact variance

> "read something once it may or may not land, think about it later and
> it might. doesnt mean you reread it, you retreaded the dents it made."

**Ruling**: my binary (artifact structure vs parser habit) is missing the
third term. A parse is artifact × the dents it left × later traversal of
those dents. Part of "parser variance" is the artifact's *operative
payload* acting on the parser (§4, §39) — which is measurement of the
artifact, not noise.
**Consequence**: variance decomposition gets three terms — artifact
main effect, parser main effect, artifact×parser interaction (the dents).
Deferred experiment (recorded, not built): the retread protocol — a
second pass over one's own earlier parse *without the artifact present*,
vs a fresh reread; what develops from dents alone? (jane's
reconstruct.py was this experiment run on a self.)

## 3. Weighting / GED costs

> "the thing is going to have subjective opinions based on what its
> exposed to thats the point. disagree feed more, still disagree? maybe
> im wrong."

**Ruling**: weighting is not contamination to minimize — it is the
instrument's *taste*, exposure-dependent, arguable-with, and permitted to
outlast the researcher's disagreement.
**Consequence**: weighting objects are versioned STATE with exposure
provenance (what the instrument had been fed when this weighting held),
updated by feeding, never by editing. Hand-set GED costs are only the
day-zero prior. The dispute protocol is native: disagree → feed more →
persistent disagreement is a result about researcher or instrument, and
the record shows which corpus grew the taste.
**Held tension (deliberate, §104)**: instrument-with-taste vs §32
(self-conviction) and §30 (shared priors). Resolution is NOT neutrality —
it is taste-with-provenance, confronted by critics and nulls like every
other producer. Subjective ≠ unaccountable.

## 4. Nulls

> "nulls should be a self-evident occurance you see a shape like another
> shape but its missing a piece. this is a thing to notice."

**Ruling**: nulls are primarily *observed near-misses*, not synthesized
distributions. A mapping that almost works, minus a piece, is a
first-class object — the missing piece is the datum.
**Consequence**: every partial/failed correspondence is retained into the
control population (§53 upgraded: negative results are not just product,
they are the *reference field* against which interestingness is
measured). Synthetic nulls (rewiring, permutation) become supplements for
when the observed near-miss field is thin, not the foundation.

## 5. Retrieval

> "retrievable in my head means what information passed through that
> left footprints."

**Ruling**: retrieval is trace-following, not lookup. What has passed
through the system leaves footprints; retrieval traverses them.
**Consequence**: the instrument gets (at least) two retrieval modes,
deliberately different in kind — trace-based (footprint field,
history-dependent, riverbed-shaped: this is delta/riverbed's machinery
re-entering as a *component* under external observation) and
fingerprint-based (WL/spectral, ahistorical). **Their disagreement is
data**: what trace retrieval finds that fingerprints miss measures the
instrument's history; the reverse measures structure the instrument
hasn't lived through. The §64 basis worry resolves: the trace field is
explicitly subjective state, never claimed as a neutral index.

## The unification (engineer's observation, for the record)

Answers 2, 3 and 5 are one answer: **dents, taste, footprints — the
instrument's history of exposure is simultaneously its memory, its
weighting, and its retrieval structure.** That is riverbed's thesis
("weight = navigational salience given accumulated overlap") stated three
times in three vocabularies, which is itself a small invariance-under-
rephrasing datum about the researcher's model. The design accepts it:
this instrument is not a neutral observer but a *subject with auditable
history* — observer-dependence (§40) embraced at day one, accountable
through provenance, perturbation, critics, and the near-miss field.