Skip to main content

szl-ouroboros

base-python-ci continuous frontier

SZL Holdings — Orbit v2

Control before action. Evidence after.

Product · Proof · GitHub organization

Kernel-twin repo for Muyu, the bounded-loop package (repository name szl-ouroboros). Not the archived TypeScript product szl-holdings/ouroboros. Not a model. No weights.

House name: Muyu. Shared graph: Ayllu Ñawpa Yachay (the community's knowledge of what came before).

Python backend source: this repository, szl-holdings/szl-ouroboros. Frontend and agentic-loop product source: szl-holdings/platform, package @workspace/ouroboros-loop. Web product surfaces in that repository are listed in PRODUCT-SURFACES.md. Hugging Face SZLHOLDINGS is the publish mirror, not the source. There is no Space SZLHOLDINGS/szl-ouroboros.

Hub mirror: kernels/SZLHOLDINGS/szl-ouroboros. Card: SZLHOLDINGS/szl-ouroboros.

Continuous frontier review

Ouroboros continuous frontier review runs every two hours. It binds one read-only model review to the exact current Second Brain frontier candidate set. The reviewed packet has ten fixed source identities: the eight repository sources and public arXiv/Crossref metadata captures. Git revisions remain 40-character SHA-1 object identifiers. Metadata captures carry the explicit metadata-capture-sha256 kind and must bind their provider, identifier, canonical URL, retained metadata digest, bounded response receipt, title, and exact content projection. They provide metadata for review only; they do not admit paper text or infer full-text rights, training authority, or independently verified truth.

If input preparation fails, the loop retains a SOURCE_PREPARATION_FAILED receipt with an unavailable source and no reviewer attempt. It does not require a reviewer timer that was never started. The final enforcement step remains red; receipt closure records the failure and never substitutes for a valid review.

The default provider chain is explicit:

  1. use the existing pinned Codex action when OPENAI_API_KEY or CODEX_API_KEY is configured;
  2. otherwise run the public, exact-revision SZLHOLDINGS/SZL-Khipu-1.5B-GGUF locally through a verified llama-cpp-python CPU wheel;
  3. fail closed if neither reviewer produces output that passes the independent schema, evidence, and authority validator.

Manual workflow_dispatch runs also offer an optional reviewer choice:

Choice Behavior
auto (default) Use the existing provider chain above.
local-gguf Disable Codex authority for this run and use only the existing exact pinned CPU Khipu lane.

The choice is honored only for workflow_dispatch; scheduled, push, and other events retain automatic selection. Invalid manual values fail before source preparation or either reviewer. The local choice suppresses the API credential from the selection step and skips the Codex action even if an API key exists in the repository. It does not change secrets, select an external model endpoint, or weaken any model, evidence, output, timing, or authority gate.

A secret-free outputs/reviewer-selection.json artifact records the event, controller revision, normalized request, selected lane, and effective Codex authority for that run. The existing loop receipt records Codex configured=false and attempted=false for explicit local runs, plus the actual local attempt and its outcome. Here configured describes authority enabled for this run; it does not claim that the repository has no secret. A selected lane is not evidence that the model completed. The native reviewer still must produce an admitted terminal result under the unchanged finalizer and enforcement gate.

The loop is:

Second Brain protected main
        ↓ exact revision + candidate SHA-256
schema and authority replay
        ↓
Codex OR exact public Khipu GGUF
        ↓ untrusted evidence-linked structured JSON
independent deterministic validation
        ↓
Ouroboros timing + termination + receipt closure
        ↓
90-day secret-free artifact

The keyless lane pins all model and runtime identity:

  • model repository: SZLHOLDINGS/SZL-Khipu-1.5B-GGUF;
  • model revision: 67d60ec577730747055491640cfb91fc4a4b5d25;
  • model file: SZL-Khipu-1.5B-Q4_K_M.gguf;
  • model bytes: 986047904;
  • model SHA-256: 13c1a1993063e1dff92f7413ccf48eaca6d48efc8801ae9af35961ae3396623a;
  • runtime wheel: llama-cpp-python 0.3.35 official manylinux CPU wheel;
  • wheel SHA-256: d172f3d3c8cdd194c3c47c71cb077ed6e61354a2d0f939ceeac0c8fd29999596;
  • temperature 0, seed 749, one reviewer attempt, bounded context and output.

No Hugging Face token is required to retrieve the public model. A local OpenAI-compatible endpoint can also be selected through OPEN_MODEL_BASE_URL; this supports self-hosted Ollama, vLLM, or a llama.cpp-compatible server without changing the receipt contract. OPEN_MODEL_API_KEY is optional and is never recorded.

The Codex lane remains pinned to openai/codex-action, Codex CLI 0.138.0, permission-profile: :read-only, safety-strategy: drop-sudo, and the same output schema. The open-weight lane does not weaken or replace the deterministic validator. Model-generated JSON is always treated as untrusted.

A reviewer can propose at most twelve evidence-linked recommendations across the approved Brain, Anatomy, A11oy, Formula, Forge, Nemo, and Ouroboros repositories. The keyless lane narrows this to one compact recommendation over six selected candidate excerpts, or an explicit no-action review. Selection prefers distinct sources by deterministic relevance score, and respects the exact requested candidate limit even when more sources are available. The execution receipt records the total source candidate count and explicitly limits the review scope to selected public excerpts. It does not claim a full-portfolio review.

The prompt describes a compact contract and the independent output validator enforces it:

Output field Maximum
Recommendations 1
Summary 120 characters
Title 80 characters
Rationale 240 characters
Risk 120 characters
Evidence candidate IDs 2
Validation steps 2, each 120 characters
Canonical complete review 1,800 UTF-8 bytes

The earlier packet limits responded to a 24-candidate truncated run, but the six-candidate run on 2026-10-05 also ended with OUTPUT_TRUNCATED. A generic JSON-object grammar allowed unbounded strings and collections despite the prompt's limits.

The local GGUF lane now uses a small, finite GBNF grammar for the exact digest and selected candidate IDs. It generates one compact ASCII JSON object with bounded strings, one recommendation, and at most two evidence IDs and validation steps. Text fields use printable ASCII without quotes or backslashes; model output remains subject to the unchanged independent validators, including evidence uniqueness and advisory-text checks. The full JSON Schema is not automatically compiled into a native grammar. Receipts retain the generation grammar version and SHA-256.

The workflow keeps its 1,800-token allowance, exact model/runtime pins, temperature, seed and one-attempt bound. Grammar acceptance and byte limits do not establish useful model output or a successful live attempt. Any non-stop completion still fails closed as OUTPUT_TRUNCATED or COMPLETION_NOT_STOPPED; an otherwise valid general-schema review that exceeds the compact contract fails as OUTPUT_BUDGET_CONTRACT. No partial output is repaired or admitted. A new exact-source live replay must retain an admitted terminal review and loop receipt before this lane is called operational.

Neither provider can edit files, use repository credentials, train weights, promote candidates, execute tools, merge pull requests, mutate providers, reveal secrets, or load the private Second Brain graph. The output is an advisory artifact, not accepted truth.

The finalizer measures the reviewer attempt window and total wall time through the active szl_ouroboros.build_loop_trace kernel. Each receipt must prove:

  • steps <= maxBudget;
  • a terminal loop exit;
  • receiptsInEqOut = true as a doctrine invariant;
  • exact Second Brain source and candidate-set digests;
  • valid candidate evidence IDs;
  • zero training, promotion, execution, merge, and provider-mutation authority;
  • exact model/runtime identity for the keyless lane;
  • independent validation before any terminal review state is accepted.

What this is NOT

  • Hub model.joblib is QUARANTINED executable serialization. Do not joblib.load it. GitHub source is the approved path.
  • Not the ouroboros TypeScript product runtime.
  • Not trained weights.
  • Not an autonomous merge or deployment agent.
  • No measured CUDA benches are claimed here.
  • Lambda is not upgraded by loop execution or model output.
  • No shared, leaked, scraped, or fabricated API key is used.

Load

Set SZL_OUROBOROS_HF_REVISION to the immutable first-class Kernel Hub commit from a verified publication of kernels/SZLHOLDINGS/szl-ouroboros. Use the kernels client version qualified with that publication. The GitHub source commit, model-type mirror commit, and Kernel Hub commit are separate identities. An observed head, a branch name, or a successful import does not qualify a release.

trust_remote_code=True permits execution of the selected repository's Python. Review that exact revision, its provenance and publication evidence before enabling it. The format check below only rejects missing or mutable revision inputs; it does not verify hashes, publisher authorization or compatibility. If that evidence is unavailable, stop the Hub load and use separately reviewed local source for development.

import os
import re

hf_revision = os.environ.get("SZL_OUROBOROS_HF_REVISION", "")
if re.fullmatch(r"[0-9a-f]{40}", hf_revision) is None:
    raise ValueError("A verified immutable Kernel Hub revision is required")

from kernels import get_kernel

ouroboros = get_kernel(
    "SZLHOLDINGS/szl-ouroboros",
    revision=hf_revision,
    trust_remote_code=True,
)
trace = ouroboros.build_loop_trace(
    [{
        "provider": "llama-cpp-python",
        "model": "SZLHOLDINGS/SZL-Khipu-1.5B-GGUF",
        "ok": True,
        "latency_ms": 1200,
        "node": "scheduled-frontier-review",
    }],
    wall_ms=1450,
    exit="converged",
    max_budget=1,
)
assert trace["withinBudget"] is True
assert trace["receiptsInEqOut"] is True

Doctrine v11. Lambda = Conjecture 1, advisory and never a theorem. Apache-2.0. Owner: Stephen Lutar / SZL Holdings.

Source-only development

Review torch-ext/szl_ouroboros/ at that immutable GitHub source revision, separately from any Hub release. With the source's dependencies already available, run from the reviewed checkout root:

import sys
from pathlib import Path

sys.path.insert(0, str(Path("torch-ext").resolve()))
import szl_ouroboros as local_kernel

This selects local Python source rather than calling the Hub loader. Importing local source also executes Python. This documentation check does not run that import, install dependencies, qualify a runtime or establish a Hub publication.

Timing admission

The local loop-tax kernel accepts finite, nonnegative int/float durations in milliseconds. Booleans, numeric strings, negative/non-finite durations and unrepresentable or overflowing totals raise ValueError before a timing trace is returned. The frontier receipt finalizer passes observations through without clamping them to zero; rejected timing cannot produce its output receipt.

Zero-duration windows remain valid. wall_ms=None still means an unavailable wall measurement. A finite wall shorter than the sum of attempt windows retains the existing wallLessThanModel inconsistency flag. This validates the numeric domain only; it does not authenticate the upstream clock or prove that supplied values were measured. Counterfactual savings and doctrine labels are unchanged.

License

Apache-2.0 — see LICENSE.

Metadata

Release files for szl-ouroboros 0.1.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for szl-ouroboros 0.1.0
File Size Uploaded Provenance Actions
szl_ouroboros-0.1.0.tar.gz 44.5 kB With provenance Details for szl_ouroboros-0.1.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for szl-ouroboros 0.1.0
File Interpreter ABI Platform Actions
szl_ouroboros-0.1.0-py3-none-any.whl Python 3 none any Details for szl_ouroboros-0.1.0-py3-none-any.whl

Total release size: 62.1 kB

Release files / szl_ouroboros-0.1.0.tar.gz

Download URL szl_ouroboros-0.1.0.tar.gz
Size 44.5 kB
Tags Source
SHA-256 checksum
How to use checksums
11be1523da54f1274e1b767cd3847607289388cb64337d70ccf3bd5c28982699
BLAKE2b-256 checksum
How to use checksums
c4792f057f8319b5362176a3ef57873d2abf84fce6c83443d72b39abc64fa222
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.

Transparency log

Release files / szl_ouroboros-0.1.0-py3-none-any.whl

Download URL szl_ouroboros-0.1.0-py3-none-any.whl
Size 17.7 kB
Tags Python 3
SHA-256 checksum
How to use checksums
ccd400874091b98749b8c9c5677166000606b6379db312fd7bfe32d78ed81912
BLAKE2b-256 checksum
How to use checksums
105adbf2dd671a170789a85f93f039d2c4fb4918f55fe6262914c3fd02542477
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.

Transparency log

Security for szl-ouroboros 0.1.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for szl_ouroboros-0.1.0.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.1.0 This release

2 release files Full provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page