Repository navigation
Releases
v1.2.2: Clarified baseline and consistent rounding
This release resolves the descriptive-rounding discrepancy and explains the polynomial constant-metric choice and proof margin. The paper retains Neurosymbolic in the title and has no reviewer line numbers.
The repeated polynomial median upper-bound statistic is now displayed as 4.25 in both the prose and Table III. Descriptive summaries are explicitly rounded to nearest; individual certificate upper bounds retain upward rounding, and certified lower guarantees retain downward rounding. All verification calculations still use saved full-precision values.
For the polynomial benchmark, P = diag(p,1) strictly satisfies the contraction inequality when p > B^2/(1-lambda)^2. At B = 2 and the exact decimal rate lambda = 2/5, the threshold is 100/9, explaining the simple integer choice p = 12. The release adds an exact rational derivation record and an independent boundary test. The raw proof margin 1/20 is chosen independently of the normalized learner margin; separate shifted endpoint checks prove it. The original endpoint certificates, decimal/binary64 rate checks, and all learning protocols are preserved.
A fresh fifteen-stage pipeline regenerates every experiment, verifies both constant-metric baselines, rebuilds the paper, and compiles the extracted arXiv source. The complete reproduction and Claude review package contain source and artifact hashes, numerical environment records, all seed outcomes, positive proof records, and stage logs. The Claude prompt now explicitly asks reviewers to examine the clarified rounding and baseline rationale.
The scientific conclusions and limitations are unchanged. The matched polynomial experiments did not demonstrate a certification-count advantage from normalization or margin coupling. Constant metrics certify the polynomial benchmark and smallest swing-model box, so those results do not require neural state dependence. The analytical baselines are post hoc fixed candidates, without optimized search or matched timing comparisons.
Downloads include the paper, arXiv source archive, complete runnable reproduction, Claude review package, both baseline proof JSON files, abstract and citation metadata, and SHA256SUMS. The paper and CITATION.cff cite this exact release. Earlier immutable releases, including v1.2.1, remain unchanged. No arXiv submission is performed by this release.
Scientific source freeze: a821d6c64441004a1ef63f8e1b89635d966aac13. The release tag also includes the regenerated artifacts; the full-run manifest records source stability throughout the pipeline.
Validation: all 15 pipeline stages and 164 tests passed. The fresh run took 1,945.5 seconds, with all 63 source hashes unchanged and 45 artifacts recorded. Independent comparison found all 16 existing scientific JSON outputs unchanged from v1.2.1 after explicitly excluding runtime metadata and the arithmetic-check timestamp; all 180 application candidates, outcomes, positive proof covers, and guarantees agree. The added polynomial-choice explanation passed independent exact matrix and threshold checks.
The paper has 30 pages, no reviewer line numbers, and an unchanged 1,900-character abstract. The 17-file arXiv archive passed an independent three-pass compilation with zero warnings; every page's text and decoded content stream matches the delivered PDF. Independent audits checked all 125 reproduction files and 136 review payloads.
All 30 rendered pages passed visual inspection, including the revised baseline paragraph, matching median displays, release citation, and final references.
v1.2.1: Exact polynomial constant-metric baseline and verified submission artifacts
This release adds an exact constant-metric certificate for the polynomial benchmark and fixes the swing-model notation to use C_P(k(B)) consistently. The paper retains Neurosymbolic in its title and has no reviewer line numbers.
For the polynomial benchmark on [-2,2]^2, the fixed matrix P = diag(12,1) certifies contraction at rate 0.4, with raw matrix margin 1/20 and normalized margin 1/260. Exact rational endpoint checks establish the result for both the decimal rate 2/5 and the actual stored binary64 rate. The full proof, metric bounds, and replay are included in polynomial_constant_baseline.json and the reproduction.
This is a post hoc analytical certificate, without training, an optimized search, or a matched timing comparison. It uses affine endpoint verification rather than the learned-metric grid oracle. Together with the existing swing-model constant metric, it shows that the polynomial benchmark and the smallest swing box do not require neural state dependence. The abstract, introduction, benchmark discussion, neural section, and conclusion now make that limitation explicit. The original negative empirical finding remains: the matched polynomial experiments did not demonstrate a certification-count advantage from normalization or margin coupling.
A fresh fifteen-stage pipeline reruns all original experiments and the physical application, replays both constant-metric baselines, generates the paper, and compiles the extracted arXiv source. The reproduction and Claude review package retain all source and artifact hashes, numerical environment details, seed outcomes, positive proof records, and stage logs. The Claude package includes a standalone integrity verifier and an updated independent-review prompt.
Downloads include the paper, tested arXiv source archive, complete runnable reproduction, Claude review package, both baseline proof JSON files, abstract metadata, citation metadata, and SHA256SUMS. The manuscript and CITATION.cff cite this exact release. Previous immutable releases remain unchanged. This release prepares materials for arXiv; it does not submit the paper.
Scientific source freeze: 2fd469c4ef532d2e22ab445900a1037e88310216. The release tag also includes the generated artifacts; the full-run manifest verifies that the scientific source hashes remained unchanged.
Validation: all 15 pipeline stages and 163 tests passed. The full run took 2,185.8 seconds, with all 63 source hashes unchanged. Independent comparison confirmed that all 16 existing scientific JSON outputs agree with v1.2.0 after explicitly excluding runtime fields and the arithmetic-check timestamp; all 180 application candidates, outcomes, proof covers, and guarantees are unchanged. The new polynomial proof was independently reconstructed with exact arithmetic.
The paper has 30 pages, no reviewer line numbers, and a 1,900-character abstract. The 17-file arXiv source passed a fresh three-pass compilation with zero warnings; all 30 page texts and decoded content streams match the delivered PDF. The reproduction and review-package audits checked 125 and 136 payload files, respectively. All 30 rendered pages were visually inspected.
v1.2.0: Physical application and verified constant-metric baseline
This release adds a physical two-machine swing-model application, a controlled comparison with an independently implemented adaptation of the Fitzsimmons–Liu PDI objective, and a verified analytical constant-metric baseline. The manuscript retains the title Neurosymbolic Synthesis of Contraction Metrics with Certified Verification and has no reviewer line numbers.
The analytical matrix P = [[1, 1/2], [1/2, 2]] certifies the smallest application box at the same requested rate and strict margin as the neural comparison. Directed MPFR enclosures and exact rational matrix checks cover both exact-decimal and stored binary64 targets. The candidate has rigorous whole-box violations on the two larger domains. These failures concern this fixed matrix and do not exclude other constant metrics or contraction on the interior invariant energy subsets. The baseline was added after the neural experiment and has no training seeds or matched timing claim.
The application retains all 180 paired neural runs. Its small-box primary counts are 20/20 for the CEGIS hinge learner, 18/20 for the adapted PDI objective, and 20/20 for the fixed-data hinge control. Larger primary boxes have no certificates. An explicitly exploratory verification-budget increase certifies two frozen middle-box CEGIS candidates without retraining. The evidence does not establish general superiority, a need for neural metrics on the smallest box, or an advantage over the original complete LyZNet/SMT workflow. The earlier polynomial ablation still reports no demonstrated certification-count benefit from normalization or margin coupling.
A fresh fourteen-stage reproduction regenerates the scientific results, verifies the analytical baseline and proof records, rebuilds the paper, and compiles the extracted arXiv source in a clean directory. The archives retain the numerical environment, source and artifact hashes, all seed outcomes, and complete stage logs. The independent-review package includes a standalone integrity/provenance verifier; checksums and passing tests do not substitute for mathematical review.
Assets:
neurosymbolic-paper-v1.2.0.pdf: current paper without reviewer line numbers.neurosymbolic-arxiv-v1.2.0.tar.gz: minimal, clean-build-tested arXiv source.neurosymbolic-reproduction-v1.2.0.zip: complete verified runnable snapshot.neurosymbolic-claude-deep-review-v1.2.0.zip: paper, source, reproduction, evidence, and independent review prompt.SHA256SUMS: checksums for the uploaded artifacts and metadata.
The manuscript and CITATION.cff cite this exact release. Previous immutable releases remain unchanged. This release does not submit the paper to arXiv.
Scientific source freeze: 5b67589cf4e8e9660aac05b5d2098d62e4b0100a. The release tag also records the generated outputs; the full-run manifest verifies that all scientific source hashes remained unchanged.
Validation: 153 tests and all 14 pipeline stages passed. The paper has 30 pages, no reviewer line numbers, a 1,883-character abstract, and zero LaTeX warnings. The extracted 16-file arXiv source passed a clean three-pass build. Independent package audit verified all 30 page content streams against that build.
Layout and navigation update with verified reproduction (v1.1.1)
This immutable release supplies a layout and navigation update to the verified
preprint Neurosymbolic Synthesis of Contraction Metrics with Certified
Verification. It preserves the scientific claims, abstract, formal mathematical
statements and proofs, numerical results, and original twelve-stage run evidence
from v1.1.0.
Changes
- Compacts the table layout so the adaptive-verification and fitted-surrogate
results no longer occupy a sparsely filled isolated float page. - Updates the manuscript and citation metadata to identify this version.
- Provides current paper, source, reproduction, and independent-review assets;
a default-branch navigation README directs readers to the versioned current
project while preserving the historical code on that branch. Its exact text
is included asrelease/DEFAULT_BRANCH_README.md.
No learning or experiment producer was rerun for this layout update. The build
revalidates saved scientific results and adaptive certificates, confirms that the
abstract and formal mathematics are unchanged from the immutable v1.1.0 source,
and tests compilation of the current paper and extracted arXiv sources.
Scientific evidence and provenance
neurosymbolic-reproduction-scientific-verified.zip remains the exact original
passing twelve-stage scientific snapshot, SHA256
c1cdea042cca4f5495a3ed025174570bb68e494df2695641af2e97d91437eb39.
Its run manifest, scientific programs, saved results, generated numbers and tables,
figures, and stage logs are preserved unchanged. The original archive's manuscript
predates the current editorial and layout updates.
editorial_v1.1.1.json records the current input/output hashes and the relation to
both the unchanged scientific snapshot and the preceding immutable v1.1.0 source.
The current reproduction ZIP contains runnable source and evidence, the final
unnumbered manuscript, the original snapshot, and a standalone integrity verifier.
The source archive contains only the thirteen flat TeX dependencies. The Claude
package includes the same current paper, code and evidence, with a neutral prompt
for independent scientific review.
Assets
neurosymbolic-paper-v1.1.1.pdf: authoritative paper without reviewer line numbers.neurosymbolic-arxiv-v1.1.1.tar.gz: minimal clean-tested arXiv source bundle.neurosymbolic-reproduction-v1.1.1.zip: complete current runnable reproduction.neurosymbolic-reproduction-scientific-verified.zip: unchanged original scientific run.neurosymbolic-claude-deep-review-v1.1.1.zip: independent-review package.- Abstract, citation metadata, editorial provenance, and checksum inventory.
SHA256SUMS lists the published asset hashes. The reproduction's
VERIFY_CURRENT_REPRODUCTION.py and the review package's VERIFY_PACKAGE.py
validate their extracted inventories and provenance without third-party Python
dependencies. These checks establish integrity, not mathematical correctness.
Earlier immutable releases, including v1.1.0, remain unchanged. The scientific
limitations are also unchanged: idealized rather than practical termination
assumptions, finite training/verification budgets, no matched external-method
performance comparison, and known-truth residual certification in the surrogate
experiment. This release prepares the paper for arXiv and does not submit it.
Citation: Victor M. Preciado, Neurosymbolic Synthesis of Contraction Metrics with
Certified Verification: Verified Reproduction and Submission Artifacts, v1.1.1,
GitHub release, 2026. https://github.com/vmpreciado/NeuroSymbolic/releases/tag/v1.1.1
Verified revised preprint and reproduction (v1.1.0)
This immutable release archives the current verified reproduction and the
unnumbered preprint Neurosymbolic Synthesis of Contraction Metrics with Certified
Verification. It incorporates the external referee revision, the Bullo/Jaffe/
Davydov literature comparisons, and the final novelty review.
Manuscript changes
- Distinguishes dual control-contraction-metric formulations from direct joint
controller-and-primal-metric learning. - Identifies failure-mechanism analysis, computable outward-rounded certificates,
and reproducible diagnostics as the principal contribution, with established
sampling and CEGIS arguments as building blocks. - Compares Davydov's partition-certificate training objective with the sampled
hinge loss and diagnostic-point augmentation used here. - Cites this release for the current paper and reproduction. The abstract retains
the matched experiments' lack of a demonstrated certification-count advantage
from normalization or margin coupling. The title retains Neurosymbolic, with
the neural and symbolic roles defined explicitly.
Scientific evidence and provenance
The verified scientific snapshot passed all twelve declared pipeline stages,
including controlled ablations, adaptive three-dimensional verification,
known-truth fitted-surrogate transfer, complete certificate reporting, and a
clean arXiv source compilation. Its exact ZIP is preserved unchanged as
neurosymbolic-reproduction-scientific-verified.zip, SHA256
c1cdea042cca4f5495a3ed025174570bb68e494df2695641af2e97d91437eb39.
The final manuscript is an editorial rebuild of those scientific outputs.
Saved results and adaptive certificates were revalidated, numerical inputs were
regenerated, and both the main and extracted arXiv sources compiled cleanly.
The final release does not claim that training or experiments were run again.
The original run manifest and stage logs remain byte-for-byte intact.
Assets
neurosymbolic-paper-v1.1.0.pdf: authoritative paper without reviewer line numbers.neurosymbolic-arxiv-v1.1.0.tar.gz: minimal tested TeX source bundle.neurosymbolic-reproduction-v1.1.0.zip: complete runnable current reproduction,
with the final paper, saved evidence, explicit editorial provenance, the original
scientific archive, and a standalone integrity verifier.neurosymbolic-reproduction-scientific-verified.zip: exact original scientific run.neurosymbolic-claude-deep-review-v1.1.0.zip: independent-review package with
current paper/source, code, evidence, and a neutral review prompt.- Abstract, citation metadata, editorial provenance, and checksum inventory.
SHA256SUMS lists the published asset hashes. The reproduction ZIP's
VERIFY_CURRENT_REPRODUCTION.py and the Claude package's VERIFY_PACKAGE.py
check their extracted contents without third-party dependencies. These checks
establish integrity and provenance, not mathematical correctness.
The main limitations remain explicit: conditional idealized termination, finite
training and verification budgets, no matched external-method performance
comparison, and known-truth residual certification in the surrogate experiment.
The earlier immutable v1.0.0 and v1.0.1 releases are unchanged. This release prepares
the paper for arXiv; it does not submit or post the paper there.
Citation: Victor M. Preciado, Neurosymbolic Synthesis of Contraction Metrics with
Certified Verification: Verified Reproduction and Submission Artifacts, v1.1.0,
GitHub release, 2026. https://github.com/vmpreciado/NeuroSymbolic/releases/tag/v1.1.0
Line-numbered preprint and tested arXiv source bundle (v1.0.1)
Verified reproduction snapshot and final review manuscript
Release v1.0.1 archives the verified scientific reproduction for
Neurosymbolic Synthesis of Contraction Certificates and the final
line-numbered preprint, with a tested arXiv source bundle. The immutable
v1.0.0 release remains unchanged.
Scientific snapshot
neurosymbolic-reproduction.zip is the exact, unchanged archive from the
passing eight-stage run. It includes the source, pinned dependency
specifications, tests, scientific results, all six figures, the manuscript at
that run, stage logs, and provenance. Its SHA256 is:
c67b4247793c591125dfd71c1d86776a21b95ba1a59d116126c51d29cd934070
Final manuscript
Use the separately attached nsp_contraction_paper_ieee.pdf for the current
preprint. It adds Fitzsimmons and Liu (2024) to the related-work discussion,
distinguishes their post-training SMT verification from the current synthesis
analysis, and cites this release. The README removes references to historical
files absent from the published tree. The manuscript uses single spacing and
retains continuous line numbers; only the double-spaced referee-mode option
was removed from the presentation settings.
The verified ZIP deliberately preserves its earlier manuscript byte-for-byte.
The attached final TeX, numbers.tex, and the tag's paper sources include the
editorial updates. editorial_rebuild.json and editorial_build.log document
the new build and verify that scientific source, data, figures, and original
full-run evidence were reused unchanged. The editorial build validated the
saved results, regenerated numerical macros, and completed two LaTeX passes
with no errors, unresolved references, or warnings. It did not rerun learning
or the scientific experiments.
Integrity and reproduction
neurosymbolic-arxiv-v1.0.1.tar.gz contains exactly eight manuscript inputs:
the main TeX source, generated numbers.tex, and six figure PDFs, all at the
top level. The bibliography is embedded in the TeX. The archive excludes the
compiled manuscript, logs, code, scientific data, and nested archives. The
clean extraction and two-pass PDFLaTeX check are documented in
arxiv_bundle.json. It is a prepared source submission, not an arXiv upload.
GitHub release immutability protects the published assets and version tag.
SHA256SUMS lists the attached asset checksums. The tag's README documents
both full scientific reproduction and a paper-only rebuild. Installing the
specified Python packages and a suitable TeX distribution is required.
Citation: Victor M. Preciado, Neurosymbolic Synthesis of Contraction
Certificates: Verified Reproduction Snapshot and Review Manuscript, version
v1.0.1, GitHub release, 2026.
https://github.com/vmpreciado/NeuroSymbolic/releases/tag/v1.0.1
Verified reproduction snapshot and final review manuscript (v1.0.0)
Verified reproduction snapshot and final review manuscript
Release v1.0.0 archives the verified scientific reproduction for
Neurosymbolic Synthesis of Contraction Certificates and the final
line-numbered review manuscript.
Scientific snapshot
neurosymbolic-reproduction.zip is the exact, unchanged archive from the
passing eight-stage run. It includes the source, pinned dependency
specifications, tests, scientific results, all six figures, the manuscript at
that run, stage logs, and provenance. Its SHA256 is:
c67b4247793c591125dfd71c1d86776a21b95ba1a59d116126c51d29cd934070
Final manuscript
Use the separately attached nsp_contraction_paper_ieee.pdf for the final
review draft. It corrects the abstract's scale-equivariance qualification,
restricts the margin comparison to stalled candidates, identifies the growth
statistic as the certified bound for the normalized test, and cites this
release. It retains continuous reviewer line numbers.
The verified ZIP deliberately preserves its earlier manuscript byte-for-byte.
The attached final TeX, numbers.tex, and the tag's paper sources include the
editorial updates. editorial_rebuild.json and editorial_build.log document
the new build and verify that scientific source, data, figures, and original
full-run evidence were reused unchanged. The editorial build validated the
saved results, regenerated numerical macros, and completed two LaTeX passes
with no errors, unresolved references, or warnings. It did not rerun learning
or the scientific experiments.
Integrity and reproduction
GitHub release immutability protects the published assets and version tag.
SHA256SUMS lists the attached asset checksums. The tag's README documents
both full scientific reproduction and a paper-only rebuild. Installing the
specified Python packages and a suitable TeX distribution is required.
Citation: Victor M. Preciado, Neurosymbolic Synthesis of Contraction
Certificates: Verified Reproduction Snapshot and Review Manuscript, version
v1.0.0, GitHub release, 2026.
https://github.com/vmpreciado/NeuroSymbolic/releases/tag/v1.0.0