Skip to content

The claim is the only place meaning lives, and it is the one field no mechanism reads #45

Open

Description

@junixlabs

Not a bug. The weakest point in the design, written down so it stops being implicit.

What the code actually does with a claim

claim is read in five places, and in every one it is an opaque string:

explore.py:251     printed into HTML
engine.py:983      a["claim"] == b["claim"]        contradiction check
engine.py:1048     {a["claim"] for a in group}     contradiction check
engine.py:1055     a["claim"] != b["claim"]        contradiction check
engine.py:1069     "value": winner["claim"]        Effective Reality

Nothing reads it as a statement. The contradiction check treats different string as
different claim, so two assertions saying the same thing in different words register
as a contradiction, and two that genuinely contradict each other are indistinguishable
from two that merely differ. That is #40 seen from the other side, and it explains why
that check is both quadratic and uninformative.

What confidence is actually measuring

Confidence is DERIVED, which INV-001 requires. But trace what it derives from:

  • artifact staleness — blob, normalised blob, and since init reads no source file; decay asks git which symbol changed #38 git log -L. Real,
    verifiable, and entirely about addresses.
  • supports — supports | refutes | inconclusive, taken as a parameter in
    observe() and written straight to the log. The only semantic content in the whole
    chain, authored by the observer, checked by nothing.

So a confirmed claim means: the addresses this claim cites still resolve, they have
not moved since somebody looked, and that somebody typed supports.
It does not mean
the sentence is true, and nothing in the system is in a position to.

Effective Reality

resolve() sets "value": winner["claim"] — an authored sentence. INV-002 forbids an
authored Reality document, and the letter is kept: nobody writes a Reality file, and the
selection of which assertion wins is genuinely derived from context, lifecycle and
time. But the content of Effective Reality is 100% authored string. What is derived
is which sentence gets shown, never anything about the world.

Why this is the weakest point rather than one more gap

It is the root of the others. #42 (staleness, never wrongness) is a consequence: nothing
detects a wrong claim because nothing reads any claim. The rubber-stamp problem that
INV-024 fights is a consequence: the one bit that carries meaning is authored, so the
gate can only make stamping inconvenient, never detectable. #40 is a consequence.

The tiers doctrine's integrity is formal rather than actual. The rule "nothing
derived is authored" holds at every step, and the derivation still bottoms out in an
authored judgement that no mechanism can check.

What follows, honestly

Not a fix — this is close to irreducible for any system that does not execute or
understand the code. What it changes is what may be claimed:

  • confirmed should be described as address freshness plus an observer's judgement,
    everywhere it is presented, including explore's legend.
  • The value of the model is the graph and the vocabulary, which are authored and
    make no verification claim, plus erosion detection, which is genuinely mechanical.
    Both survive this criticism intact.
  • Any future feature justified by "the model knows X is true" is justified by nothing.

Related: #40, #42, #43.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions