Most AI systems are built for answer production. Ask a question, retrieve some documents, generate a response. If the retrieved material looks authoritative, the answer often sounds authoritative too.
Incomplete, contradictory, or controlled evidence requires a stronger method. The harder problem is deciding what can be said when the available record is only part of the real record.
Grounded Chiral Tensor Synthesis (GCTS) is the current research direction of the Chiral Narrative Synthesis project. It is an evidence-first architecture for ranking likely truth under limited, contradictory, and adversarial information.
The short version:
GCTS ranks claims by building structured possible worlds, scoring them against evidence and access conditions, and reporting calibrated likely-truth rankings with explicit uncertainty, proof support, and record contingencies.
The Prior-Art Boundary
The individual ingredients are crowded. Automated fact verification, truth discovery, citation-grounded generation, provenance systems, probabilistic logic, possible-world databases, legal argumentation, missing-data theory, and benchmark-leakage controls all have substantial prior art.
The GCTS research question concerns the integration. Can those ingredients be assembled into one architecture where missing records, controlled access, record-generation duties, contradiction residuals, strict proof, likely-truth posterior mass, and runtime oracle-boundary controls all remain visible in the output?
The proposed contribution lives in the typed record-access layer connected to possible-world ranking and audit-ready explanation.
The Failure Mode: Evidence Has Multiple Access States
Many systems treat the evidence layer as a bucket. If a document is retrieved, it is available. A missing retrieval result gets treated as absence.
Real disputes have a broader record surface.
A record may exist but be sealed. It may be expected under ordinary procedure but never produced. It may be controlled by the actor whose conduct is being evaluated. It may have been destroyed. It may never have been generated. It may be available but non-supportive. Or it may affirmatively refute the claim.
Those are different epistemic states.
Conflating them produces bad reasoning. “No evidence was found” becomes “the claim is false.” “The record was not produced” becomes “there is nothing to see.” “A source denies the claim” becomes “the claim is unsupported,” even when that source controls the decisive records.
GCTS makes record access a first-class object. A claim can be plausible but record-contingent. A claim can be likely but low-confidence because decisive records are inaccessible. A claim can be rejected because expected evidence affirmatively negates it. Those statuses should not collapse into one generic “unsupported.”
The Record-Access Layer
A record-access state asks concrete questions:
- What record would matter?
- Who would own or control it?
- What duty, policy, role, or instrumentation would generate it?
- Should the event have been observable?
- Was the record requested?
- Was it produced, refused, partially produced, contradicted, sealed, destroyed, produced late, or unavailable?
- How confident is the system in that access-state classification?
The key rule is conservative: absence affects ranking only through generation duty, expected observability, access path, control, production state, and source or institutional incentives.
Strict Proof And Likely Truth Are Separate Outputs
GCTS separates strict proof from likely truth.
A strict claim requires resolvable evidence and a proof trace. It must survive zero-temperature rule closure. When that proof path is unavailable, the claim stays outside strict proof unless an explicit rule marks it false.
Likely truth is different. A claim may receive high posterior mass across structured possible worlds even when strict proof is unavailable. That ranking must come from explicit evidence, access-state assumptions, source reliability, contradiction structure, and calibration.
GCTS emits three separate quantities:
P(c | E,A,I): likely-truth posterior mass across structured worlds;P0(c | E): strict proof support from resolvable evidence and proof traces;Conf(c): confidence after uncertainty decomposition.
A competent system must be able to say:
- this is strictly proven;
- this is likely without strict proof;
- this is plausible but depends on a missing record;
- this is conflicted;
- this is unsupported;
- this is rejected.
Each output has different operational meaning.
Possible Worlds Instead Of One Forced Answer
When evidence is contradictory, a single synthesis can hide the strongest alternative or smooth over the missing record that would decide the case.
GCTS maintains a distribution over possible worlds.
Each world contains accepted facts, assumptions, latent context predicates, proof traces, access hypotheses, and missingness hypotheses. Worlds are scored against the available evidence and penalized for contradiction, unsupported complexity, access mismatch, weak grounding, and excessive assumptions.
The output becomes a ranked set of alternatives:
- world A explains the evidence with low contradiction but depends on a sealed record;
- world B fits the produced documents but requires a narrow time interval;
- world C is simpler but leaves a major access-state mismatch unresolved.
The output should support ranked judgments: “X has the highest posterior mass given the current record; Y remains live because record R is inaccessible.”
Chirality As Residual Mismatch
The word “chiral” points to mismatch.
A narrative can be fluent, persuasive, and semantically coherent while still failing when translated into structured evidence, proof, and access states. GCTS measures that failure as chirality: the distortion between language, logic/proof structure, available evidence, and expected-but-missing records.
High chirality means the story has tension that deserves inspection. It may be a genuine contradiction. It may be a hidden context variable. It may be a missing record. It may be source framing. It may be an unsupported synthesis that sounds better than it is.
The system treats chirality as diagnostic. Productive conflict can have high chirality. The engine measures the residual, decomposes it, and assigns it to evidence gaps, hidden context, access limits, or explicit uncertainty.
The Oracle Boundary
The most dangerous version of this system would secretly ask an oracle for the answer.
That oracle might be a hidden benchmark label, a human reviewer at runtime, or an LLM prompted to decide truth. GCTS forbids that pattern.
Labels and expert judgments may be used for training, calibration, evaluation, and error review. They may not bypass runtime evidence closure and world ranking. A deployable run must be able to explain how each promoted claim came from evidence, access states, rules, worlds, proof traces, and calibrated parameters.
LLMs may extract, normalize, and render. They may not supply runtime truth mass.
Why This Matters
The obvious applications are scientific disagreement, intelligence analysis, investigative reporting, legal review, compliance, institutional accountability, and high-stakes organizational decisions. The underlying problem is broader: modern information environments are full of partial records.
Ordinary RAG systems are useful when the answer is in the retrieved documents. They are weaker when the decisive fact is in a record that is unavailable, withheld, sealed, destroyed, or never generated. They are weaker still when the absence of that record is itself part of the evidence.
GCTS targets that harder setting. Its output should show what the evidence supports, where support fails, which worlds remain possible, and which record would change the analysis.
Current Research Hub
- CNS 7.1 / GCTS: Grounded Chiral Tensor Synthesis
- Prior-Art Boundary
- Record-Access Ontology
- Worked Example
- References
Older CNS 2.0 material remains available as prior work, especially for the history of Structured Narrative Objects, chirality, and dialectical synthesis. The current framework moves beyond that stage by making likely-truth ranking, possible worlds, access modeling, and the oracle boundary central.
Selected References
- Thorne et al., FEVER: a Large-scale Dataset for Fact Extraction and VERification.
- Wadden et al., Fact or Fiction: Verifying Scientific Claims.
- Schlichtkrull et al., AVeriTeC: A Dataset for Real-world Claim Verification with Evidence from the Web.
- Min et al., FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation.
- Li et al., A Survey on Truth Discovery.
- W3C, PROV-O: The PROV Ontology.
- Bach et al., Hinge-Loss Markov Random Fields and Probabilistic Soft Logic.
- Rubin, Inference and Missing Data.
- Cornell LII, Federal Rule of Civil Procedure 37.