NSP
ENFR
Menu

Research / evidence record

Claims should carry their evidence with them.

The research record distinguishes public preprints, working papers, engineering traces and open questions.

Follow the three-paper arcOpen the evidence index

Three-paper arc

One question widens across three papers.

The sequence moves from authorization to act, through accumulation across attempts, toward structural cognition across domains. Each step retains its own evidence status.

  1. Paper IPublic preprint

    Think Before You Act

    Research question
    Should an agent act on what it remembers?
    Proposed contribution
    Separate recall from authorization to act through explicit understanding, assumptions, confidence and verification.
    From pre-action discipline to persistent cognition
  2. Paper IIPublic preprint

    Programmable Emergence via Cognitive Engine

    Research question
    Can structured state change how research matures across attempts?
    Proposed contribution
    Organize persistent cognition into state representation, attention allocation and self-regulation.
    From one research domain to structural transfer
  3. Paper IIIWorking paper

    Structure Over Similarity

    Research question
    Can one scaffold take different cognitive shapes across domains?
    Proposed contribution
    Propose a four-layer functional separation and examine mathematics, software and music under different truth conditions.
    Working-paper boundary

Individual paper records

Method and limits stay beside the abstract.

Formal titles remain in English in both languages. Status, dates and links come from one verified fact source rather than from localized copy.

PaperI
Public preprint

Think Before You Act: A Cognitive Gating Protocol for Reliable AI Agents

Status
Public preprint
Date
March 30, 2026
Authors
Zhang, S.Y.
Version DOI
10.5281/zenodo.19334052
Original language
English
Version
v1
Access
Open
Licence
Creative Commons Attribution 4.0 International

Abstract

The paper presents NSP as an external cognitive-gating layer around memory and agent action, then reports selected retrieval, gating and roleplay evaluations.

Key contribution

It separates the ability to recall context from the decision to act on that context, making interpretation and pre-action verification explicit.

Method

Architecture proposal; three selected LongMemEval-s categories; an author-created 150-item Cognitive Gating Benchmark; and a roleplay A/B evaluation.

Evidence scope

Public preprint evidence from the reported configurations, models and judges. It is not an independently replicated assessment.

PaperII
Public preprint

Programmable Emergence via Cognitive Engine: Longitudinal Evidence of AI Research Maturation

Status
Public preprint
Date
April 2, 2026
Authors
Zhang, S.Y.
Version DOI
10.5281/zenodo.19392279
Original language
English
Version
v1
Access
Open
Licence
Creative Commons Attribution 4.0 International

Abstract

The paper describes seven external cognitive mechanisms and examines whether structured state accumulates across long-running mathematical research.

Key contribution

It frames longitudinal research maturation as an observable interaction between state representation, attention allocation and self-regulation.

Method

A 261-exploration Hadamard campaign; an 80-exploration full-system ablation across four problems; a cross-domain activation audit; and a 47-exploration Williamson campaign.

Evidence scope

Public preprint evidence from one system instance around a frozen language model, with primary empirical coverage in mathematical research.

PaperIII
Working paper

Structure Over Similarity: Structural Cognition Across Mathematics, Software, and Music

Status
Working paper
Date
Not publicly available
Authors
Zhang, S.Y.
Version DOI
Not publicly available
Original language
English
Version
Not publicly available
Access
Private working copy
Licence
Not publicly available

Abstract

The working manuscript examines structure-first cognition across mathematics, software and music through a shared four-layer functional frame.

Key contribution

It distinguishes a structural substrate, cognitive scaffold, domain encoding and outward expression, while keeping each domain's truth conditions explicit.

Method

Cross-domain system inventory; geometric analysis; a music quality-improvement arc; an axiom-transfer matrix; and a qualitative three-domain cognitive-profile comparison.

Evidence scope

Private working-paper evidence. There is no verified public record, publication date or download, and the three domains use non-equivalent evidence regimes.

No public file or DOI verified

Evidence index

Highlighted values, with their scope attached.

This is a reading index into the preprints, not an independent re-analysis. Positive, negative and null results are retained together.

  1. E-01Think Before You Act
    Reported value94.5%

    Selected LongMemEval-s result

    94.5% across multi-session, temporal and knowledge-update categories.

    Sample / observations
    344
    Method
    Official LongMemEval evaluator with GPT-4o-mini as judge.
    Boundary
    Three of six categories; LLM-as-judge; no direct comparison with six-category averages.
    Source
    Paper I, Section 5.1
  2. E-02Think Before You Act
    Reported value-28 pp

    Capability-matching boundary

    The evaluated Haiku condition moved from 98% baseline to 70% with the protocol.

    Sample / observations
    50
    Method
    Cognitive Gating Benchmark gate-accuracy subset.
    Boundary
    One model condition; shows that more scaffolding is not universally beneficial.
    Source
    Paper I, Sections 5.2 and 6.4
  3. E-03Programmable Emergence via Cognitive Engine
    Reported value+22.5 pp

    Full-system ablation

    Aggregate success was 85.0% with the cognitive system and 62.5% when accumulation was ablated.

    Sample / observations
    80
    Method
    Balanced full-system ablation; CMH p = 0.021.
    Boundary
    10 runs per condition per problem; full-system rather than component-level ablation; no individual problem remained significant after multiplicity correction.
    Source
    Paper II, Sections 5.7 and 6
  4. E-04Programmable Emergence via Cognitive Engine
    Reported value+81%

    Constraint extraction

    Reported structural constraints per exploration increased from 2.0 to 3.6.

    Sample / observations
    80
    Method
    Balanced ablation subset; Mann-Whitney U p < 0.001.
    Boundary
    Extracted constraints were not independently validated for mathematical correctness.
    Source
    Paper II, Section 6.3
  5. E-05Programmable Emergence via Cognitive Engine
    Reported value0 pp

    Reported null result

    The Ising 2D comparison was 90% in both full and ablated conditions.

    Sample / observations
    20
    Method
    Per-problem result within the controlled ablation.
    Boundary
    Small per-problem sample; interpreted as a ceiling effect in the paper.
    Source
    Paper II, Section 6.2

Benchmark and experiment summaries

The record includes improvement, degradation and no effect.

The studies are most useful when the operating condition is visible. The page therefore avoids collapsing unlike evaluations into one headline number.

Paper I

Gating and retrieval

Selected retrieval and gate decisions improved in the reported Sonnet configurations.

The evaluated Haiku condition degraded under the same protocol, and the retrieval result covers only selected categories.

Public preprint
Paper II

Accumulation and transfer

The full-system ablation reported higher aggregate success and structural-constraint extraction.

The Ising comparison was null; primary evidence remains one mathematical research system.

Public preprint
Paper III

Structural cognition across domains

The manuscript compares how one scaffold is encoded and evaluated in mathematics, coding and music.

Working paper only; evidence types are asymmetric and several comparisons remain qualitative.

Working paper

Engineering evidence

Operational traces are labelled as operational traces.

Engineering records can show how a safeguard behaves in practice. They do not substitute for controlled or independent evaluation.

Engineering record

A real failure became a system safeguard.

  1. A producer generated a faulty mathematical enumeration.
  2. An independent auditor blocked it.
  3. A later round weakened the same finding.
  4. NSP detected relabelling without correction.
  5. The incident became a persistent deterministic guard.
Evidence source
internal-system-trace
Boundary
This is an internal engineering record, not an independent peer-reviewed evaluation.

Limitations

The boundary is part of the result.

These conditions recur across the current research record and should travel with downstream claims.

  1. 01

    Preprints, not peer review

    Paper I and II are open public preprints. Paper III is a private working manuscript.

  2. 02

    One implementation

    No second backend has reproduced the reported cognitive architecture or cross-domain profile.

  3. 03

    Verifier dependence

    LLM judges, parsers and unverified extracted constraints can transmit error into reported outcomes.

  4. 04

    Uneven evidence regimes

    Mathematics uses campaigns, coding uses operational traces and music includes subjective listener verdicts.

  5. 05

    Model and task fit

    Protocol complexity can help, do nothing or harm depending on model capability and problem structure.

Open questions

What the current record does not settle.

  1. Will the architecture replicate on a second backend and a different model family?
  2. Which individual mechanism drives the ablation effect?
  3. How should extracted constraints and LLM-judged outcomes be independently verified?
  4. Do the music findings survive blind multi-listener evaluation and production-grade rendering?
  5. When does the benefit of persistent cognition exceed its integration and protocol cost?

Citation information

Cite the version that was actually reviewed.

Version DOIs identify the exact public records used by this site. Concept DOIs remain available on Zenodo for all-version citation.

  1. I
    Suggested citation

    Zhang, S. Y. (2026). Think Before You Act: A Cognitive Gating Protocol for Reliable AI Agents. Zenodo. https://doi.org/10.5281/zenodo.19334052

    Version record

    10.5281/zenodo.19334052

  2. II
    Suggested citation

    Zhang, S. Y. (2026). Programmable Emergence via Cognitive Engine: Longitudinal Evidence of AI Research Maturation. Zenodo. https://doi.org/10.5281/zenodo.19392279

    Version record

    10.5281/zenodo.19392279

  3. III
    Suggested citation

    Zhang, S. Y. (n.d.). Structure Over Similarity: Structural Cognition Across Mathematics, Software, and Music. Working paper.

    Version record

    Not publicly available

Downloads / DOI

Official files where they exist.

Paper I and II point directly to their verified Zenodo records and open PDFs. Paper III stays metadata-only until a public record is confirmed.

Paper II

Programmable Emergence via Cognitive Engine

10.5281/zenodo.19392279

Licence: Creative Commons Attribution 4.0 International

Open official recordOpen PDF
Record verified: July 15, 2026
Paper III

Structure Over Similarity

No public file or DOI verified