Datasets:
assumptions listlengths 2 5 | id stringlengths 3 3 | limits listlengths 0 4 | proof_locations listlengths 1 4 | relation_to_prior_work stringclasses 6
values | source_path stringclasses 1
value | statement stringlengths 80 309 | status stringclasses 3
values | title stringlengths 18 61 | verification_evidence listlengths 0 2 |
|---|---|---|---|---|---|---|---|---|---|
[
"Convex globally L-smooth objective",
"Exact constant-step gradient descent",
"h=eta L>0"
] | C01 | [] | [
"research/chapters/03_universal.tex: thm:classical-seven-fourths"
] | Attributed to Barzilai, Shamir, and Zamani (2025), not claimed as a new result. | research/metadata/claim_ledger.json | Every second difference is nonnegative for h<=7/4; matching violations exist above the boundary. | established_result_reproved | Classical convexity boundary | [] |
[
"Convex globally L-smooth objective",
"Exact constant-step gradient descent",
"No minimizer assumption for the universal direction"
] | C02 | [
"The compensation inequality is distinct from the prohibition of two adjacent negative second differences.",
"Sharpness strong-convexity constants may depend on the requested step."
] | [
"research/chapters/03_universal.tex: thm:compensation",
"research/chapters/03_universal.tex: thm:sharp-compensation"
] | research/metadata/claim_ledger.json | h_iso=(19-sqrt(17))/8 is the exact universal boundary for D_(k+1)+D_(k+2)<=2D_k. | proved_in_manuscript | Exact compensation boundary | [
"research/verification/certificates.json",
"research/verification/sharp_compensation.json"
] | |
[
"Convex globally L-smooth objective",
"Exact constant-step gradient descent",
"The exact root is specified by its rational isolating interval"
] | C03 | [
"This is not the exact threshold at every fixed positive strong-convexity ratio.",
"Longer-run thresholds are not identified here."
] | [
"research/chapters/03b_adjacency.tex: thm:sharp-adjacency"
] | The optcurves repository already gives the exact bracket [1.8634,1.8635]. This manuscript supplies a closed algebraic endpoint and proofs; historical priority for that refinement is not established. | research/metadata/claim_ledger.json | The root h_* of the displayed degree-seven polynomial is the exact universal threshold for two adjacent strictly negative second differences. | proved_in_manuscript | Exact adjacent-concavity boundary | [
"research/verification/adjacent_boundary.json",
"research/verification/adjacent_negative.json"
] |
[
"m is a positive integer",
"Objective and step may depend on m",
"Step h=2-1/[128(m+1)]",
"Initial point (0,1)"
] | C04 | [
"The condition-number dependence is a sufficient construction, not an optimal tradeoff.",
"The exact rational verifier checks the unsmoothed trajectory and analytic-transfer budgets; it does not exactly integrate the Gaussian convolution."
] | [
"research/chapters/04_witness.tex: thm:w-main",
"research/chapters/05_analytic.tex: thm:a-main"
] | Related public work establishes selected consecutive-run examples and exact brackets. No first-priority assertion is made for the present all-m analytic theorem. | research/metadata/claim_ledger.json | For every m>=1 there is a two-dimensional entire real restriction with m strictly negative second differences, exact condition number 128(m+1), explicit margins, and globally convergent descent. | proved_in_manuscript | Arbitrary-length analytic strongly convex runs | [
"research/verification/witness.json"
] |
[
"Normalized step lies in the interval of the stated weighted law",
"Nonnegative decrements for the scalar interpretation"
] | C05 | [
"Optimality of the scalar envelope is asserted over its stated scalar sequence class, not over all realizable gradient-descent trajectories."
] | [
"research/chapters/03_universal.tex: thm:filtered-convexity",
"research/chapters/03_universal.tex: thm:scalar-horizon-envelope",
"research/chapters/03b_adjacency.tex: cor:adj-filter"
] | research/metadata/claim_ledger.json | Positive weighted curvature laws yield decreasing convex filtered loss curves and finite-horizon gain bounds; the compensation-only scalar envelope has a matching scalar extremizer. | proved_in_manuscript | Filtered convexity and horizon envelopes | [
"research/verification/sharp_compensation.json"
] | |
[
"Prescribed finite first-order coefficients",
"Declared smooth strong-convexity class",
"Gradient-scale normalization and stated finite observations",
"Dimension is unrestricted within the finite realization bound"
] | C06 | [
"The finite model is not an unconstrained characterization of arbitrary adaptive optimizers or all neural-network training."
] | [
"research/chapters/02_foundations.tex: thm:convex-interpolation",
"research/chapters/02_foundations.tex: thm:strong-interpolation",
"research/chapters/06_transcripts.tex: thm:finite-transcript",
"research/chapters/06_transcripts.tex: thm:complete-prediction"
] | Smooth interpolation and performance-estimation foundations are due to earlier work including Taylor, Hendrickx, and Glineur and Drori and Teboulle; the manuscript develops explicit decision and realization consequences. | research/metadata/claim_ledger.json | The declared finite first-order transcript is realizable exactly when its Gram/interpolation constraints are feasible. Under bounded normalization, every affine prediction has an attained interval and every intermediate value is realized. | proved_in_manuscript | Complete finite transcript prediction | [
"research/verification/certificates.json"
] |
[
"Nonempty compact bounded model",
"Rational input coefficients for the rational-certificate conclusion",
"A strictly positive gap from the decision boundary"
] | C07 | [
"A finite exact certificate at a zero margin is not guaranteed by this theorem.",
"No uniform complexity bound for finding a certificate is asserted."
] | [
"research/chapters/06_transcripts.tex: thm:residual-cert",
"research/chapters/06_transcripts.tex: thm:certificate-completeness"
] | research/metadata/claim_ledger.json | Rational bounded-model data admit a finite rational residual certificate for every strict decision margin; coefficient and PSD errors have explicit allowances. | proved_in_manuscript | Exact verification of strict decision margins | [
"research/verification/certificates.json"
] | |
[
"Declared convex smooth reference objective",
"Valid displacement or gradient error bounds",
"Valid measurement and objective-drift bounds",
"Nonnegative variable steps where the stated formula uses eta_k"
] | C08 | [
"A passing consistency check does not prove global convexity or validate the error assumptions.",
"Arbitrary nonzero errors do not preserve exact curvature signs without a margin."
] | [
"research/chapters/03_universal.tex: thm:univ-inexact",
"research/chapters/07_robustness.tex: thm:trajectory-tube",
"research/chapters/07_robustness.tex: thm:law-transfer"
] | research/metadata/claim_ledger.json | Exact linear trajectory laws transfer to measured inexact trajectories with explicit step, gradient, state, and value-error allowances. | proved_in_manuscript | Robust finite laws | [
"research/verification/certificates.json"
] | |
[
"The common-transcript construction assumes 0<eta mu<1 and eta L<2",
"The analytic construction uses 0<mu<L and a nondegenerate quadratic GD factor",
"Only the specified finite observations are supplied"
] | C09 | [
"The analytic separation is strictly positive but is not claimed to have an order-one uniform lower bound.",
"The arbitrary-history-length construction does not fix an additional initial-distance bound."
] | [
"research/chapters/08_information.tex: thm:info-full-interval",
"research/chapters/08_information.tex: thm:info-stopping",
"research/chapters/08_information.tex: thm:info-analytic"
] | research/metadata/claim_ledger.json | A finite full first-order history can leave the whole classical remaining-gap interval unresolved; exact analytic finite-jet agreement does not identify the next loss; stopping regret follows. | proved_in_manuscript | Sharp information limits | [
"research/verification/information_results.json"
] | |
[
"A common bounded risk in [0,1]",
"Candidate and incumbent fixed during each comparison",
"Fresh or conditionally mean-valid bounded paired observations",
"A single whole-run error budget",
"The stated acceptance rule and valid numerical-error envelopes"
] | C10 | [
"The result compares candidates actually examined.",
"Ordinary repeated reuse of exposed validation data is not covered.",
"No empirical training speedup is asserted."
] | [
"research/chapters/09_recursive.tex: thm:rsi-simultaneous",
"research/chapters/09_recursive.tex: thm:rsi-oracle",
"research/chapters/09_recursive.tex: prop:rsi-sequential-cost"
] | Uses established bounded-variable concentration and adaptive-validation principles, proved explicitly for the stated controller. | research/metadata/claim_ledger.json | One simultaneous confidence event validates all adaptive comparisons and sample stopping times; accepted improvements telescope, and the final incumbent is bounded relative to every examined candidate. | conditional_theorem | Adaptive recursive acceptance and finite candidate comparison | [
"research/verification/recursive_controller_results.json"
] |
[
"Positive declared cost-to-risk exchange factor",
"Every attempted training/certification/validation cost is charged",
"Each attempt respects its reserved cap",
"Accepted gains have valid nonnegative lower certificates"
] | C11 | [
"The ledger does not supply free compute or unlimited initial resources.",
"Infinite attempts with a positive minimum cost cannot fit a fixed finite credit."
] | [
"research/chapters/09_recursive.tex: thm:rsi-ledger"
] | research/metadata/claim_ledger.json | With pre-reserved cost caps and certified accepted gains, lambda times the total cost of all attempts is at most initial exploration credit plus actual cumulative common-risk improvement. | conditional_theorem | Complete resource accounting | [
"research/verification/recursive_controller_results.json"
] | |
[
"Infinitely many completed rounds in a sufficient resource regime",
"Fair visitation of a countable candidate collection",
"Validation radii and margins tend to zero",
"Common-risk validity, or summable uniform drift for the changing-risk version"
] | C12 | [
"Fairness and resources are explicit assumptions, not guaranteed by the proposal generator.",
"This asymptotic theorem is not combined with an impossible infinite positive-cost finite-credit run."
] | [
"research/chapters/09_recursive.tex: thm:rsi-convergence",
"research/chapters/09_recursive.tex: thm:rsi-drift"
] | research/metadata/claim_ledger.json | Fair proposals and vanishing uncertainty/margins yield convergence to the best reachable candidate; summable uniform risk drift admits a limiting-risk extension. | conditional_theorem | Conditional consistency with changing proposals and tasks | [
"research/verification/recursive_controller_results.json"
] | |
[
"A finite horizon and finite number of strict continuous transcript conditions.",
"Convex globally 1-smooth source objective.",
"Positive regularization parameter epsilon may become arbitrarily small."
] | C13 | [
"No preservation of exact jets or exact interpolation equalities.",
"No preservation of a prescribed positive strong-convexity lower bound.",
"No single-parameter transfer of an infinite family of conditions.",
"No uniform useful condition-number or curvature-margin guarantee; the quantitative explicit family... | [
"research/chapters/05_analytic.tex: thm:a-strict-transfer",
"research/chapters/05_analytic.tex: cor:a-finite-patterns"
] | research/metadata/claim_ledger.json | At a fixed positive step and initial point, any finite strict continuous condition on a convex 1-smooth GD transcript has an arbitrarily close analytic strongly convex realization. This yields equality of the feasible finite strict curvature-sign patterns when the positive strong-convexity parameter is free. | proved_in_manuscript | Analytic strongly convex transfer of finite strict conditions | [] |
Ŋhoryūn–Ω III: Certificates, Curvature, and Recursive Learning
Finite Trajectory Laws, Analytic Long Runs, and Verifiable Improvement
Author credit: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Research release: 3.0.0 · 13 November 2026
Repository type: Dataset · mathematical manuscript, proof sources, exact audit code, and structured research records
Read the 103-page manuscript · Research overview · Claim ledger · Data dictionary · Reproduce the results
Dataset summary
This dataset distributes a self-contained mathematical monograph and its reproducibility materials on the curvature of gradient-descent loss sequences, finite-history prediction, information limits, and conditional validation of recursive learning. It includes the complete PDF, editable LaTeX proofs, exact-arithmetic audit programs, a recursive comparison controller, figures, source data, citation metadata, and structured JSONL views of the research.
For exact constant-step gradient descent, define
Here $\kappa_k<0$ means the next loss decrease is larger than the current one. The manuscript distinguishes local curvature boundaries, constructs arbitrarily long finite runs of increasing improvements, and studies what can be certified from finite training observations.
Principal mathematical results
| Result | Statement and scope |
|---|---|
| Classical convexity boundary | The universal nonnegative-curvature threshold is $h=7/4$. This is an established result of Barzilai, Shamir, and Zamani, reproduced and attributed in the manuscript. |
| Sharp compensation boundary | $h_{\mathrm{iso}}=(19-\sqrt{17})/8\approx1.8596117968$ is the exact universal boundary for $D_{k+1}+D_{k+2}\le2D_k$. The manuscript gives the inequality and matching counterexamples beyond the boundary. |
| Sharp adjacent-concavity boundary | $h_*\approx1.863425364364$ is the algebraic endpoint for two adjacent strictly negative second differences. A positive weighted curvature law holds through the endpoint, and analytic strongly convex counterexamples occur above it. |
| Arbitrary finite analytic runs | For every integer $m\ge1$, an explicit two-dimensional analytic strongly convex objective has $m$ consecutive negative second differences, exact condition number $128(m+1)$, and a globally convergent descent orbit at $h=2-1/[128(m+1)]$. |
| Finite strict-condition transfer | Every finite collection of strict continuous trajectory conditions realizable in the smooth convex class transfers to an arbitrarily close analytic strongly convex transcript when the positive strong-convexity parameter is allowed to vary. |
| Finite transcript prediction | Smooth convex interpolation and Gram constraints give exact finite realizability and attained affine prediction intervals under the stated finite-method and normalization assumptions. |
| Information limits | Common finite observations can leave a sharp interval of remaining optimality gaps unresolved; finite analytic derivative observations also need not identify future losses. |
| Recursive comparison and accounting | Under the declared sampling, common-risk, resource, and proposal assumptions, a controller validates adaptive comparisons, composes accepted gains, and accounts for every attempted experiment. |
The adjacent-concavity endpoint is the unique root in
of
For the analytic long-run construction,
The objective and step depend on $m$. This is a theorem about arbitrary finite runs. It does not assert an infinite increasing tail of positive decrements on a lower-bounded objective.
Relevance to AI and recursive learning
The optimization results characterize possible training-curve behavior and the limits of extrapolating progress from a finite history. The validation controller compares arbitrary candidate learning systems through a declared common risk, including candidates that change the subsequent proposal mechanism. Its guarantees concern accepted improvement, uncertainty, comparison against evaluated candidates, and complete experiment costs.
Convex trajectory theorems apply when their optimization assumptions hold. The statistical comparison controller has its own stated bounded-observation and conditional-sampling assumptions. The package supplies no measured neural-network speedup, unrestricted self-improvement guarantee, or empirical AI benchmark.
Structured subsets
Each subset has a single train split for compatibility with dataset tooling. The split name is an organizational convention; no statistical training/test partition or held-out evaluation set is implied.
| Configuration | Records | Content |
|---|---|---|
claims |
13 | Statements, assumptions, proof locations, evidence paths, and limits from the research claim ledger. |
manuscript_sections |
13 | Complete chapter LaTeX text with source paths and SHA-256 hashes. |
witness_cases |
11 | Exact rational parameters and analytic-transfer budgets for run lengths 1, 2, 3, 4, 8, 16, 32, 64, 128, 256, and 512. |
audit_runs |
7 | Original verification entrypoints, script hashes, report paths, and pass status. |
trajectory_m32 |
34 | Exact rational states and losses for the unsmoothed construction, with decrements, second differences, and decimal display values. |
files |
See manifest | File paths, byte counts, SHA-256 hashes, media types, and artifact categories. The manifest and its enclosing checksum file are excluded from their own row index. |
Exact rational values are serialized as strings such as 1/4224. Decimal values are display aids. The trajectory_m32 subset evaluates the unsmoothed rational witness; it is not a numerical evaluation of the Gaussian-smoothed analytic objective. Analytic transfer is justified by the written proof and audited bounds.
Loading
Replace YOUR_USERNAME with the actual repository owner after upload:
from datasets import load_dataset
claims = load_dataset("YOUR_USERNAME/Nhoryun-Omega-III", "claims", split="train")
trajectory = load_dataset("YOUR_USERNAME/Nhoryun-Omega-III", "trajectory_m32", split="train")
The JSONL files can also be loaded locally without executing repository code:
from datasets import load_dataset
claims = load_dataset("json", data_files="data/claims.jsonl", split="train")
Repository contents
| Path | Purpose |
|---|---|
research/Nhoryun_Omega_III_Manuscript.pdf |
Complete 103-page manuscript. |
research/chapters/, research/manuscript.tex, research/preamble.tex |
Complete editable proof sources. |
research/code/, research/run_verification.py |
Exact audits, certificate utilities, recursive controller, and optional numerical exploration. |
research/verification/ |
Original verification reports, logs, and PDF quality record. |
research/metadata/ |
Claim ledger, research status, pinned related-work metadata, and original release manifest. |
research/figures/, research/data/ |
Figures and original decimal plotting data. |
data/*.jsonl |
Derived viewer datasets described above. |
tools/rebuild_indexes.py |
Regenerates JSONL indexes and payload checksums from the included research files. |
SHA256SUMS |
Checksums for every payload file except itself, including the file manifest. |
CITATION.cff |
Original author credit and citation metadata. |
The original 63 research files are preserved under research/. Its original SHA256SUMS remains usable from that directory.
Verification and provenance
The original release records seven passing exact-arithmetic audit programs, including 11 long-run witness cases with 1,026 strictly negative second differences, 200 interpolation equalities, 100 scalar-horizon identities, 756 exact future values, 378 zero-jet equalities, and 39 recursive-controller assertion groups. These counts measure different mathematical objects and should not be added into a single proof count.
The all-parameter statements are established by the manuscript's proofs. Finite arithmetic checks audit the implementations, identities, and selected witnesses. Gaussian integrals are not numerically evaluated by the exact suite. Proof-assistant formalization and external replication are not recorded as completed.
The JSONL views are mechanically derived from the included research sources, ledger, and reports. They introduce no new theorem or experimental result. The manuscript is AI-generated mathematical research with the author credit stated above. The package contains synthetic mathematical records, not observations collected from human participants.
Related work and remaining questions
The manuscript credits the smooth interpolation and performance-estimation foundations, including Taylor, Hendrickx, and Glineur and Drori and Teboulle. It also credits Barzilai, Shamir, and Zamani's Are Convex Optimization Curves Convex?.
The related optcurves project already reports rational brackets for selected consecutive-violation thresholds. The manuscript distinguishes those results from its algebraic endpoint and arbitrary-length analytic construction. Historical priority is not established by this package.
Questions remaining open here include exact thresholds for every run length $m\ge3$, optimal condition-number tradeoffs, fixed-conditioning phase boundaries, complete realizable curvature-sequence characterizations, and practical RSI acceleration.
Uses, limits, and rights
Intended uses include mathematical reading, exact-certificate inspection, reproducibility, optimization research, and evaluation of theorem and proof reasoning subject to the applicable rights. No statistical generalization benchmark or independently certified ground-truth proof corpus is supplied.
No explicit reuse license was present in the source release, and this packaging adds no new license grant. See RIGHTS.md. External works remain subject to their own terms; their source code and full texts are not bundled.
Citation
@misc{nhoryun_omega_iii_2026,
author = {{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}},
title = {Ŋhoryūn–Ω III: Certificates, Curvature, and Recursive Learning},
year = {2026},
version = {3.0.0},
note = {Mathematical manuscript and reproducibility dataset}
}
No DOI, venue, or publication identifier is assigned by this release preparation.
- Downloads last month
- 131
