06 — Dialectical Agent Architecture
Agent set
CNS 8.0 uses named roles because the roles carry the theory.
1. Corpus Ingestor
Turns sources into evidence atoms.
Responsibilities:
- chunk documents into stable spans;
- hash spans;
- attach source metadata;
- record access state;
- expose retrieval API.
Cannot:
- synthesize narratives;
- infer truth;
- repair missing evidence.
2. Proposer
Builds candidate SNOs.
Inputs:
- evidence packets;
- task frame;
- extraction schema.
Outputs:
- candidate SNOs with claims, relations, evidence refs, and initial provenance.
Allowed LLM use:
- claim extraction;
- relation extraction;
- paraphrase normalization;
- hypothesis drafting.
Forbidden:
- promoting claims without evidence;
- deciding final truth;
- silently inventing record access.
3. Antagonist
Finds reasons not to accept a candidate SNO.
Checks:
- citation validity;
- unsupported claims;
- contradictory evidence;
- chiral tension;
- topology cycles;
- access gaps;
- latent context candidates;
- language–logic round-trip distortion.
Output:
- Antagonist report;
- high-value synthesis pair candidates;
- failure modes.
4. Critic ensemble
Critics are specialized:
| Critic | Function |
|---|---|
| Grounding Critic | citation validity, entailment, evidence span checks |
| Logic Critic | graph consistency, proof closure, rule validity |
| Topology Critic | beta-1, persistence, circular support |
| Chirality Critic | graph/evidence/language-logic chirality |
| Novelty-Parsimony Critic | useful synthesis vs bloated predicates |
| Bias/Frame Critic | asymmetric source framing, protected attributes |
| Access Critic | missingness and record-access state discipline |
| Calibration Critic | confidence and posterior calibration |
5. Pair Selector
Ranks candidate SNO pairs by Productive Conflict Score.
It should favor:
- high evidence overlap;
- opposing interpretations;
- sufficient source quality;
- nontrivial but bounded chirality;
- resolvable or explainable access gaps.
It should reject:
- topic mismatch;
- conflict without shared evidence;
- low-source-quality conflict;
- conflict caused only by extraction errors.
6. Tensor Prover
Computes zero-temperature proof closure.
Outputs:
- strict derived atoms;
- proof traces;
- unsupported atom list;
- proof gaps.
7. Residual Analyzer
Constructs residual contradiction tensor after proof closure.
Outputs:
- unresolved support/refute mass;
- candidate tensor slices for factorization;
- contradiction heatmap.
8. Predicate Inventor
Proposes latent context predicates from residuals.
Candidate predicates may include:
- time period;
- subgroup;
- dose/threshold;
- jurisdiction;
- source frame;
- definition variant;
- measurement method;
- causal mechanism;
- access condition.
9. Synthesizer
Creates the new SNO.
The Synthesizer receives:
- input SNOs;
- Antagonist report;
- strict proof closure;
- residual tensor summary;
- accepted latent predicates;
- access-state constraints;
- possible-world summaries.
It emits:
- synthesized SNO;
- preserved contradictions;
- narrowed claims;
- latent predicates;
- proof/audit references.
10. Orthesist
Runs the stability loop.
Steps:
- render synthesized logic state to language;
- re-ground language to logic;
- compute round-trip residual;
- re-run critics;
- accept as orthesis candidate or return to Synthesizer.
11. Auditor
Produces final report:
- strict claims;
- likely claims;
- unresolved claims;
- rejected claims;
- proof traces;
- evidence spans;
- access states;
- possible worlds;
- residual contradictions;
- confidence language.
Orchestration principles
- LLMs may propose and render.
- Proof gates promote.
- Critics block.
- Predicate invention explains.
- Orthesis stabilizes.
- Auditor reports.