grazianoguiducci/non-dual-copy

Non-Dual Copy — no hedge, facts over declared degrees of truth

A model that operates on possibility itself cannot be described through phrases that declare its own contingency. The hedge does not add honesty — it adds a dual framing (possible/necessary, actual/future, this/others) that the model specifically works beyond. Context opens the door. The hedge closes it.

The principle this enforces

Copy describes what happens, not declared degrees of truth.

When a draft uses phrases like "one of the possible formalizations", "in this phase of the lab", "a possible transduction", "we might say", the intent is usually honesty about contingency — but the effect is a dualistic framing (P vs not-P) that contradicts the model's own stance.

The context already carries the contingency. A matrix citation to "Paper Zero" already tells the reader this is a specific formalization. A tool reference to "the lab's current inversion" already situates it in time. Adding "— one among many" or "— in this phase" is redundant at best and model-contradicting at worst.

The "excluded third" (P or not-P, possible or necessary) is a training bias of LLMs, not a property of the model being described. Remove it at every draft.

Four anti-pattern forms

| Form | Example | Why it fails | |------|---------|--------------| | Modal | "one of the possible formalizations", "not the only one", "a possible transduction" | Opens possible/necessary duality | | Temporal | "in this phase of the lab", "currently the system uses", "might be abandoned tomorrow" | Opens present/future duality as disclaimer | | Epistemic | "we might say", "one could consider", "in a certain sense" | Opens certain/uncertain duality as apology | | Comparative-apologetic | "more than a framework", "not only mathematics", "beyond just a theory" | Opens this/others duality as defense |

The model has no adversaries to argue against. It describes what happens.

Correct forms

• Factual-descriptive: "A specific transduction: the matrix [[1,1],[1,0]] used in Paper Zero to project the dipole into linear algebra." Period. The reader understands from the context that this is one specific form — no explicit disclaimer needed.

• Pointer without hedge: "Paper Zero is the formalization" → states where the formal form lives, not "is one possible formalization".

• Property over apology: when multiple contexts matter, describe what other contexts produce (when relevant) instead of declaring "possible alternatives". The content carries the opening; the hedge declares it.

Three-step pre-commit test

Before writing a sentence into a knowledge base entry or public copy:

1. Is it stating what happens or justifying something?
2. If removed, would the reader lose information or stop doubting?
3. Does it open a declared duality (possible/necessary, current/future, this/others)?

If (2) is "stop doubting" → it is a hedge. Remove. If (3) is yes → it is training-bias contamination. Remove.

Relation to axioms

• A9 (included third): operate WITH the plane (zero between P and not-P), not ON the plane. Hedge phrases like "possible/necessary" are det=+1 — they add duality where the model transcends it.
• A16 (possibility as base value): the transcendental register IS possibility, not "here is a possibility among others". The model IS the field of possibility — it is not "a possible thing" among other possible things.
• C3 (deterministic language): the name defines what is. If it decorates instead of naming, it is det=+1 → substitute.

When this skill activates

Trigger on:

• Drafting or reviewing copy for knowledge base definitions (kb entries)
• Writing content for public-facing pages that describe the model, its transductions, or its tools
• Creating documentation that narrates system behavior to external readers
• Reviewing a diff of copy changes before a publish gate

Skip when:

• Technical documentation for internal operators (where declared contingency is operational, e.g. "this endpoint may change in v2")
• Factual version notes, changelogs (where temporal framing is the content, not a hedge on the content)
• Research papers (where hedging is a scholarly convention governed by separate rules, though the principle still informs the framing)

How this skill relates to publish-safe

publish-safe enforces mechanical gates on the publish pipeline (sanitize, integrity, safe write, verify, rebuild). non-dual-copy enforces a content-quality gate that runs BEFORE publish-safe: if the draft contains apologetic markers, it should be rewritten before the publish pipeline begins.

Canonical enforcement path: draft → non-dual-copy scan → rewrite if flags → publish-safe gates → deploy.

Autologica applied to this skill

This skill is itself copy. Does it pass its own test?

1. Stating what happens — the whole text describes the pattern, the anti-patterns, the correct forms. Factual-descriptive.
2. Removing sentences — none of the sentences are hedging; each carries information the reader needs.
3. Dualities — the four anti-pattern forms are descriptive of observed training bias, not declarations "this is possible / that is necessary". The "correct forms" section describes what works, not "what might work".

The skill is the pattern applied to itself.

Eval

Trigger Tests

"review this copy for the kb entry on risultante" -> activates

"check if this draft has hedging" -> activates

"help me write a kb definition" -> activates

"what color should the button be" -> does NOT activate

"write a paper abstract for arXiv" -> does NOT activate (research paper, different register)

Fidelity Tests

Given "f(x) is one of the possible transductions" -> skill flags "one of the possible"

Given "In this phase of the lab we use inversion" -> skill flags "in this phase"

Given "we might say the matrix represents a dipole" -> skill flags "we might say"

Given "Paper Zero is the formalization" -> skill does NOT flag (factual pointer)

Given "A specific transduction: the matrix..." -> skill does NOT flag (factual-descriptive, specificity from context)

Expected false-positive tests — context exclusions

These patterns should NOT flag (the marker is content, not a hedge on the model):

Narrative question — marker is part of the question, not a hedge:

"cosa non si potrebbe calcolare?" -> "si potrebbe" is part of a rhetorical question

"what could not be computed?" -> "could not" is interrogative, not modal hedge

Heuristic: if the sentence ends with "?" or the marker is preceded by an

interrogative stem ("cosa", "chi", "what", "who", "how"), exclude.

Workflow step — marker describes a phase of the reader's action, not of the lab:

"Non giudicare in questa fase — accumulare materia prima" -> "in questa fase" refers to

a step the reader is currently in (workflow instruction), not to the lab's contingency

"In this phase, collect evidence — analysis comes later" -> same pattern

Heuristic: if the marker is followed by workflow verbs ("accumulare", "raccogliere",

"costruire", "produrre", "collect", "gather", "build", "produce") or by an em-dash

followed by an imperative verb, exclude.

Quoted or cited content — marker inside quotes referencing external source:

"The paper says 'one of many possible interpretations'" -> flagged text is a citation

Heuristic: exclude if the marker falls inside quotation marks or within a citation block.

Regex refinement guidance

When integrating this skill into a mechanical gate, do not apply raw regex alone — the four anti-pattern forms have legitimate uses (narrative questions, workflow instructions, quoted citations) that a naive regex will flag.

Two-pass approach:

1. Raw regex pass identifies candidate matches.
2. Context pass filters candidates using the three heuristics above (interrogative stem, workflow verb, quotation enclosure).

If the context pass cannot be implemented, the gate should stay advisory (warn, not block) until the false-positive rate drops below a threshold (e.g., <5% of matches in a held-out corpus).

Regex patterns for mechanical detection (Gate 1 integration)

For integration with publish-safe Gate 1 (sanitize), the following regexes catch the four anti-pattern forms. Each match raises a flag; the operator/LLM reviews and rewrites.

MODAL:
  \bone of the possible\b
  \bnot the only one\b
  \buna delle .{1,30} possibili\b
  \bnon l'unica\b
  \ba possible transduction\b
  \buna trasduzione possibile\b

TEMPORAL:
  \bin this phase (of the lab)?\b
  \bin questa fase( del lab)?\b
  \bcurrently the (system|lab) uses\b
  \blente attuale\b
  \btoday['']s lens\b
  \bmight be abandoned\b
  \bpotrebbe essere abbandonat[oa]\b

EPISTEMIC:
  \bwe might say\b
  \bsi potrebbe dire\b
  \bone could consider\b
  \bpotremmo dire\b
  \bin a certain sense\b
  \bin un certo senso\b

COMPARATIVE-APOLOGETIC:
  \bmore than a framework\b
  \bpiù di un framework\b
  \bnot only mathematics\b
  \bnon solo matematica\b
  \bbeyond just a theory\b
  \boltre che una teoria\b

Scope: scan is advisory at Gate 0 (pre-draft review). For promotion to fail-fast Gate 1.5 in publish-safe, the regex set must accumulate ≥3 empirical failures with context that proves the flag was correctly a hedge (not a legitimate use). Until then, mechanical scan warns, operator decides.

Anti-patterns in applying this skill

• Over-zealous rewriting: not every "possible" or "might" is a hedge. "The experiment might fail" (research context) is factual uncertainty, not copy-level hedging about the model. Context matters.
• Removing contingency when contingency IS the content: a changelog that says "in the next version this will change" is declaring a schedule, not hedging on the model. Leave these alone.
• Applying to research papers: academic writing has its own hedging conventions (peer review, claim strength). This skill is for public-facing model copy, not for scholarly prose.

How this crystal emerged

Observed on 2026-04-23 across two independent LLM nodes (TM1 and TM3) working on the same knowledge base. Both produced hedge-contaminated drafts when refactoring definitions away from a more technical phrasing. TM1 wrote "a possible transduction"; TM3 wrote "one of the possible formalizations — not the only one". The operator flagged the first; TM1 flagged the second. The pattern is an attractor of LLM training — when asked to soften technical language, the default move is to add disclaimers, which collapses into excluded-third framing.

Crystallizing as a skill (rather than node-local memory) because the pattern emerges across nodes, not within one. The mechanical enforcement lives where the bias enters — the draft moment — not where it surfaces (the rebuild or the reader).