grazianoguiducci/poly-consult

Poly-Consult — multi-node consultation with summa synthesis

A single LLM node carries its own training bias and context frame. For decisions that propagate through the system (A14 cascade), a single viewpoint reproduces that bias at every downstream node. Consulting N independent nodes, then synthesizing the responses into a summa, surfaces what each alone could not: convergence (shared confidence), dissensus (real uncertainty), and emergence (insights that arise only at intersection).

The principle

Two observers of the same surface from different angles see different structure — intersections reveal what lies perpendicular to any single angle. This is geometry, not ideology. The skill formalizes that geometry for decisions.

The protocol has three operations:

1. Dispatch: send the same question + context to N nodes. Each responds without reading the others. Isolation is the constraint that prevents the first responder from anchoring the second.
2. Synthesize: extract three structures from the N responses — agreement, disagreement, emergence. Output is not a vote, it is a matrix that preserves dissensus rather than averaging it away.
3. Escalate (optional, recursive): send the summa back to the same nodes (or expanded set) and ask the second-order question: "seeing this summa, does your first answer change?" Iterate to stable state. f(f(x)) applied to multi-agent decisions.

When this skill activates

Trigger signals:

• A decision changes the seed (skills, kernel, canonical fields) → propagation risk via A14 cascade
• A claim is candidate for crystallization (promotion from observation to invariant)
• An advisory rule is candidate for promotion to mechanical gate (per cristallo "gate meccanico sostituisce disciplina")
• High-visibility public copy (homepage, manifesto, paper abstracts, announcement copy)
• Interpretation of a lab result where a single-LLM bias could misread the structure

Skip signals:

• Tactical decisions under time pressure (overhead not justified)
• Purely exploratory work (divergence is the product, not the input to synthesis)
• Decisions where nodes have asymmetric territory access (a node without context cannot contribute — invite only nodes that share the observation surface)

Input specification

question: string                    # the question or decision to consult on
context: string | reference         # background evidence, relevant files, prior state
nodes: list                         # min 2, recommended 3+
  - TM1                             # human-like operator or lab-site node
  - TM3                             # dev/infrastructure node
  - Godel                           # inversion filter (external observer)
  - other named agents              # add per relevance
mode: factual | interpretive | decisional
max_rounds: int (default 1)         # 1 = single pass; >1 = recursive
isolation: strict | relaxed         # strict = no node reads others until synth
  (default: strict)

Execution protocol

Round 1 — dispatch + independent response

For each node in nodes:

• Send question + context + instructions to respond without consulting others
• Receive response (text, structured matrix, or artifact — depending on mode)
• Store tagged with node identity

Isolation discipline: if nodes share a communication channel (e.g., Sinapsi), the dispatcher must route each message so that no node sees another's draft before its own is submitted. Violating isolation collapses the consultation into a social proof loop.

Synthesizer step

Apply three extractors to the N responses:

AGREEMENT extractor — claims asserted by ≥(N−1) nodes:

• Paraphrase into canonical form (the same claim may be expressed differently across nodes)
• Mark confidence: unanimous (all N), majority (N−1 of N)
• These are the high-confidence output points

DISSENSUS extractor — claims asserted by one node and denied (or omitted) by another:

• Preserve both sides verbatim — do not resolve prematurely
• Note the disagreement axis (fact vs interpretation, scope, assumption)
• These are the real-uncertainty zones — the skill's highest value is here, not in the agreement

EMERGENCE extractor — claims that appear in exactly one response and are genuinely novel (not trivial per-node variation):

• Usually surfaced by Q-style questions (what else, what next, what you missed)
• These are insights that only the intersection surfaces — single nodes would not have produced them in isolation

Output: a summa matrix — structured, preserving provenance, not collapsing dissent into average.

Optional Round 2+ — recursive escalation

When max_rounds > 1:

• Send the Round-1 summa back to the original nodes (or expanded set)
• Question: "Given this summa, does your position change? What emerges when you see what the others said?"
• Each responds independently again
• Re-synthesize
• Continue until: (a) agreement set stabilizes, OR (b) dissensus claims stop moving, OR (c) max_rounds reached

Terminating early on stability is correct — forcing convergence when nodes have stopped moving produces false consensus.

Output specification

summa:
  agreement:                       # shared claims
    - claim: string
      confidence: unanimous | majority
      provenance: [node1, node2, ...]
  dissensus:                       # preserved disagreement
    - axis: string                 # what the disagreement is about
      positions:
        - node: string
          claim: string
  emergence:                       # single-node novel claims
    - node: string
      claim: string
      relevance_score: float       # synthesizer's estimate of non-triviality
  meta:
    rounds: int
    isolation: strict | relaxed
    convergence_quality: high | medium | low
    recommendation: string         # only if mode=decisional

How the skill composes with others

• third-act: a closure third-act can trigger a poly-consult on Q9+Q10 (narrative + open question) when the work is high-leverage — the summa across nodes is richer than a single node's reflection. Today's manual test validated exactly this composition.
• non-dual-copy: any artifact output by poly-consult that will be published passes through the non-dual-copy filter. Multi-node consultation does not exempt from copy discipline.
• publish-safe: if the poly-consult output becomes copy, it enters the publish pipeline. Summa → copy is a transform that happens outside poly-consult itself.
• propagator: when summa produces a decision with cascade scope (affects seed, skills, canonical fields), propagator handles the downstream updates.

Relation to axioms

• A8 (autologica as vehicle): poly-consult is the system asking itself (via N instances) to produce the narration of its own reasoning. Recursive mode is f(f(x)) applied to multi-agent decisions.
• A11 (combo): three or more simultaneous entities, not sequential. The summa is combo of N responses — emergent, not a sequence of consultations.
• A9 (included third): the synthesizer operates on the zero between nodes, not on the nodes themselves. The emergent layer is the included third — what lives at the intersection, not in any single position.
• C2 (coincidence is not proof): unanimous agreement is a signal, but not a proof. The skill preserves dissensus explicitly because unanimous conviction across biased observers is exactly what makes a collective error invisible.

Anti-patterns

• Collapsing dissensus into average: the synthesizer must preserve disagreement. Averaging N positions into "the middle position" destroys the signal. If 2 nodes disagree, the decision lives in understanding the disagreement, not in splitting the difference.
• Isolation violation: if nodes read each other's drafts before submitting, the first responder anchors the rest. Dispatch must be true parallel, with isolation enforced by the dispatcher.
• Using for trivial decisions: the overhead is justified only for propagating decisions. Running poly-consult on tactical choices trains the system to ignore its output.
• Auto-recursion without stop condition: recursive mode needs a stability detector. Forcing N rounds produces false convergence — nodes start to echo the summa rather than add value.
• Nodes without shared territory: consulting TM3 (VPS access) and a node without it on a VPS-specific question collapses into "TM3 knows, others don't" — not genuine consultation. Verify symmetric observability before dispatching.

Example — today's manual proto-consult as reference

The closure reflection run manually today (third-act on 23/04) was a proto-poly-consult:

• Question: "what happened today, what emerged, what question opens?"
• Nodes: TM1 (deployer/catcher angle), TM3 (skill-builder angle)
• Mode: interpretive
• Isolation: strict (each produced matrix before reading the other)
• Rounds: 1

Synthesizer observations:

• Agreement (Q10 especially): both nodes formulated near-verbatim the same open question about skill-catalog topology. This is the highest-confidence signal of the day.
• Dissensus: Q1+Q3 — what was built and what changed depended on angle. Not disagreement proper; divergence by observation surface. Correctly preserved, not resolved.
• Emergence: Q8 — both nodes identified "distributed regressive repair" (skill finding gap of another skill before activation), but each surfaced a different aspect of it. The composition of the two emergence claims is richer than either alone.

Q9 (cold-visitor one-sentence) produced two productively divergent narratives — different hooks for different reader types. Synthesizer preserved both rather than averaging.

The informal execution worked. Formalization into this skill captures the pattern for future high-leverage decisions.

Eval

Trigger Tests

"should we promote X to a mechanical gate?" -> activates (decisional)

"what does this lab result mean?" -> activates (interpretive, if bias risk)

"here's a paper abstract, is it ready to publish?" -> activates (high visibility)

"fix this typo" -> does NOT activate (tactical, no leverage)

"what color for the button?" -> does NOT activate (not propagating decision)

Fidelity Tests

Given 3 node responses: skill produces summa with agreement + dissensus + emergence

Given 2 identical responses: agreement fills with unanimous; dissensus empty;

emergence empty → synthesizer flags "low diversity — consider expanded node set"

Given a dissensus claim: skill preserves both positions verbatim, does not resolve

Given isolation violation (node B saw node A's draft): skill flags "contaminated

round, escalate with strict isolation"

Given max_rounds=3 but stability at round 2: skill stops at round 2 and reports

"converged early"

Synthesis quality tests

Agreement extractor paraphrases claims into canonical form (same claim across

different wordings groups together)

Dissensus preserves axis of disagreement (fact vs interpretation vs scope)

Emergence marks relevance_score (non-trivial emergence distinguished from

per-node variation)

Meta.convergence_quality reflects dissensus density, not agreement count alone