CNS 7.1 / GCTS: Grounded Chiral Tensor Synthesis

This is the current research hub for CNS 7.1 / GCTS: Grounded Chiral Tensor Synthesis.

GCTS changes the center of gravity from “synthesize conflicting narratives” to rank likely truth under limited, contradictory, and adversarial evidence. It ranks truth through structured possible worlds, evidence, access conditions, and calibrated uncertainty.

The core move:

CNS should build a distribution over structured possible worlds, quantify the mismatch between language, logic, evidence, and access states, and emit ranked, confidence-calibrated likely-truth hypotheses with explicit evidence, record dependencies, and uncertainty.

What Changed From CNS 2.0

The earlier CNS 2.0 work introduced Structured Narrative Objects, chirality, evidential entanglement, critic pipelines, and generative synthesis. GCTS keeps the useful intuition that productive disagreement has structure, but replaces several loose parts with stricter machinery:

CNS 2.0 emphasis	GCTS upgrade
Structured Narrative Objects	Evidence atoms, claims, access states, and possible worlds
Critic score	Separate strict proof, posterior probability, and confidence
Chiral pair synthesis	Chirality plus contradiction residuals across graph, tensor, and access layers
LLM-centered synthesis	LLMs extract and render; structured evidence ranks truth
Evidence overlap	Access-aware missingness, source control, and record-generation duty
Truth-like trust score	Likely-truth ranking with oracle-boundary controls

Design Commitments

Likely truth is the target. Claims are ranked by calibrated posterior mass across possible worlds, not by direct LLM confidence.
Strict proof is separate. A strict claim requires zero-temperature closure, resolvable evidence, and a proof trace.
Access states are first-class. Missing, unavailable, sealed, withheld, destroyed, and not-generated records are distinct epistemic states.
No runtime oracle. Labels and expert judgments may calibrate the system offline, but runtime ranking must come from evidence, access states, rules, and calibrated model parameters.
Multiverse views are first-class. The output is a ranked set of possible worlds before any single answer is selected.
Every claim gets a status. The report distinguishes proven, probable, plausible, record_contingent, conflicted, unsupported, and rejected.

Reading Path

Theory for the formal objects, chirality, worlds, and confidence decomposition.
Architecture for the runtime pipeline.
Oracle Boundary for which inputs are allowed to influence runtime truth ranking.
Experiments for the falsifiable test plan.
MVP Build for the implementation path.
Adversarial Evidence for access modeling and missing-record discipline.
Glossary for canonical terms.

Source Status

This section is adapted from the CNS 7.1 / GCTS research docset generated on May 13, 2026. It presents a buildable research proposal before a completed implementation.