Research / innovation / publication

Measure the behavior.
Keep the failure.

Kingan Logic studies how AI-assisted decision systems preserve evidence, state, authority and human context across long workflows. Product validation and scientific efficacy remain separate: negative, mixed, interrupted and repaired results stay in the record rather than disappearing when a later run improves.

01 / CURRENT RESEARCH PROGRAM

Beyond accuracy in governed AI decisions.

The research program asks what happens before, during and after a model answer: what evidence was available, what the evaluator actually measured, what permission existed, what state survived handoff, and what a later repair is allowed to change.

8public Zenodo preprints
8published research records
1bounded spinout
A–Hcontinuing research map
Current research discipline

Product validation is not scientific efficacy.

Implementation progress, test coverage, cloud receipts and product behavior remain separate from claims about human or model benefit. A system can be technically real without the research question being settled.

Failure remains part of the result.

Null, mixed, interrupted and repaired studies stay in the chronology. Parser, runtime, scorer, evaluator and protocol failures are recorded separately from target behavior rather than being silently folded into a later success.

Transfer has to be tested.

Cross-lane recurrence can generate a research question, but shared architecture is not proof that the same mechanism or benefit transfers across multifamily, physical assets, logistics, disaster response or long-horizon human–AI work.

02 / METHODS

Research infrastructure is part of the result.

A model score is only interpretable if the experiment around it survives inspection. Current work uses frozen protocols, blind treatment mapping, exact artifact custody, explicit missing/replacement handling, paired or heldout comparisons where justified, human/machine calibration and claim ceilings that remain binding after a positive-looking result.

Blind evaluation

Freeze before grade

Treatment identity and scoring surfaces are separated before evaluation.

BLIND MAP / FROZEN PACKET / PROVENANCE
Failure-aware evaluation

Instrument failures stay visible

Parser, runtime, scorer, evaluator and protocol failures are separated from target behavior.

INSTRUMENT / TARGET BEHAVIOR / FAILURE OWNERSHIP
Permission boundary

Evidence is not permission

A detected signal or successful repair does not automatically authorize a downstream claim or consequential action.

EVIDENCE / PERMISSION / CLAIM CEILING
Exact custody

Versioned evidence stays attached

Versioned prompts, receipts, hashes, run identities and negative results remain attached to the result they support.

PROMPTS / RECEIPTS / HASHES / RUN IDENTITY
03 / PAPER PROGRAM

Ongoing research questions.

The eight public preprints are listed below. These continuing questions, methods and unreleased work remain separate from publication status; publishing a paper does not mean its broader research problem is solved.

PAPER / A

Permission Is Not Obligation / Move or Hold?

When evidence permits movement, what determines whether move, hold or another state placement is appropriate?

ACTIVE / METHODS
PAPER / B

When the Evaluation Pipeline Fails Before the Model

Failure ownership, repair and claim ceilings when the evaluation apparatus itself breaks before the target is measured.

PUBLIC PREPRINT + CONTINUING WORK
PAPER / C

Reasoning Is Not Free

Compute, latency, reasoning cost and human correction burden.

HOLD / EVIDENCE DEMAND
PAPER / D

Does Decision Governance Transfer?

Whether recurring evidence, permission and state architecture across different domains represents architecture reuse or measurable effect transfer.

STUDY DESIGN
PAPER / E

Evaluator validity and provenance

Evaluator validity, reproducibility versus validity, human/machine calibration and independence of evidence.

PUBLISHED WORK + CONTINUING STUDY
PAPER / F

Recovery responsibility and answer donation

Human reconstruction burden, retained state, provenance and long-horizon human–AI collaboration.

PUBLISHED WORK + CONTINUING STUDY
PAPER / G

Peripheral Capture / Useful Things Found Off the Main Road

Whether valuable adjacent evidence can materially improve the route without stealing the destination.

MECHANISM RECONSTRUCTION
PAPER / H

The Guardrail Can Be Wrong / Evidence-Bound Protective Gates

Cases where a protective HOLD or FAIL may be appropriate even though the system's diagnosis of why it stopped is wrong.

ONGOING EVIDENCE REVIEW
SPINOUT

Detection Is Not Remediation

Finding a defect does not establish the repair or grant permission to execute it.

BOUNDED SPINOUT
04 / PUBLIC LIBRARY

Papers, preprints & research notes.

Eight DOI-linked public preprints are listed below. Preprint publication does not mean peer-review acceptance or independent validation. Newer research and follow-on methods remain separately scoped.

Preprint

Answer-Key Burden: A State-Responsibility Framework for Long-Horizon Human–AI Collaboration

Proposes a state-responsibility framework for identifying human recovery work in long-horizon human–AI collaboration. The construct and proposed measurement approach are not established measures of human benefit.

2026-08-22Zenodo v0.1 • DOI 10.5281/zenodo.22062448human–AI collaborationstate responsibility
Open preprint →
Preprint

When the Evaluation Pipeline Fails Before the Model: Failure, Repair, and Claim Ceilings in Multi-Stage AI Evaluation

Examines failure ownership and repair when parts of an evaluation pipeline fail before the target is fully measured. Introduces a provisional method for preserving partial evidence without overstating conclusions.

2026-08-25Zenodo v1 • DOI 10.5281/zenodo.22091682AI evaluationfailure ownership
Open preprint →
Preprint

Evaluator Provenance in Multi-Stage AI Evaluation: Separating Repeatability, Agreement, Human Ratification, and Decision Authority

Separates repeatable evaluator outputs, agreement among evaluators, human ratification and the authority to make a decision. Publication of the preprint does not establish independent validation.

2026-08-31Zenodo v1 • DOI 10.5281/zenodo.22207383evaluator provenancedecision authority
Open preprint →
Preprint

When the Judge Becomes the Experiment

Studies evaluator behavior as a measurement variable, including the difference between reproducibility, agreement and validity. Bounded findings do not establish general model superiority.

2026-09-05Zenodo v1.0 • DOI 10.5281/zenodo.22340649evaluation validitymodel-as-judge
Open preprint →
Preprint

Scorer Defects After the Result

Investigates scoring defects discovered after an evaluation result and their implications for reporting, recomputation and the conclusions supported by the original run.

2026-09-05Zenodo v1.0 • DOI 10.5281/zenodo.22343213scorer integrityevaluation repair
Open preprint →
Preprint

The Perfect Score With a Missing Case

Examines how incomplete evaluation coverage can coexist with an apparently complete score, and why case accounting must be preserved before interpreting performance.

2026-09-06Zenodo v1.0 • DOI 10.5281/zenodo.22541029coverage integrityevaluation validity
Open preprint →
Preprint

When Translation Is Not Transition

Examines the difference between translating information into another representation and demonstrating an intended state transition. Measured findings remain scoped to their evaluated cases.

2026-09-06Zenodo v1.0 • DOI 10.5281/zenodo.22543624state transitionAI evaluation
Open preprint →
Preprint

When the Human Is the Hidden Answer Key

Develops a protocol for evaluating recovery responsibility and answer-donation contamination in long-horizon AI work. This is a proposed method, not evidence of reduced human burden.

2026-09Zenodo • DOI 10.5281/zenodo.22681331human–AI evaluationanswer donation
Open preprint →
05 / RESEARCH PROFILES

Research profiles.

Public scholarly identity and persistent researcher identifiers. Development accounts and private product repositories are intentionally not linked here.

OpenReview

Logan Davis

Active researcher profile and submission identity.

~Logan_Davis1
Open profile →
ORCID

Logan Davis

Persistent scholarly identifier.

0009-0006-8244-5610
Open profile →
Web of Science

ResearcherID

Researcher identifier linked to the scholarly identity stack.

QWD-8017-2026
Open profile →
Research contact

Replication beats a flattering summary.

Kingan Logic is open to bounded research collaboration, independent replication, evaluator/instrument studies and selectively scoped product research where the protocol, evidence boundary, contribution and claim ceiling can be stated clearly.