szl-holdings/szl-command-lab

Holographic command body. Energy MEASURED from RAPL/NVML. GPU train fail-closed. Not flagship.

TypeScript

0

63 commits

updated Oct 19, 2026

See the code

README


title: SZL Atlas — Governed AI Estate emoji: 🧭 colorFrom: gray colorTo: indigo sdk: docker app_port: 7860 pinned: true license: apache-2.0 short_description: Explore SZL systems, models, kernels, and evidence. tags:

  • governed-ai
  • ai-governance
  • provenance
  • inference
  • kernels
  • evidence
  • agentic-ai
  • szl-holdings

SZL Holdings — Orbit v2

SZL Command Lab

Explore SZL systems, models, kernels and datasets in one searchable place. Follow each item to its source, current observations and stated use limits.

Artifact: Public portfolio application

Stage: Exploration and bounded demonstrations

Explore Command Lab → · Build with the source → · Inspect exact source →

Use limits

  • Catalog membership and reachable URLs do not establish model quality or operational readiness.
  • The five-organ demonstration is synthetic; it grants no action authority.
  • Unavailable or partial provider observations remain visible and must not be interpreted as a complete inventory.
Technical documentation, original evidence and reproduction

SZL governed AI command fabric

SZL Atlas

Governed intelligence. Portable evidence.

The public explorer and source-bound Launchpad for SZL systems, models, kernels, datasets, and their verification boundaries.

Open the live Atlas · Enter A11oy · Inspect evidence · Audit the source

What the Atlas does

The Atlas turns the public SZLHOLDINGS estate into one searchable command surface. It recaptures the live, unauthenticated Hugging Face inventory and presents:

  • all publicly listed models, first-class kernels, datasets, and Spaces;
  • curated paths into A11oy, Killinchu, Lyte, Sentra, Terra, PURIQ Finance, PRISM Counsel, Living Anatomy, the Khipu runtime, and YARQA;
  • a source-controlled Launchpad whose registered links remain navigation metadata—not liveness, capability, authorization, or readiness claims;
  • model and evidence spotlights driven by provider metadata rather than hard-coded performance claims;
  • a bounded five-organ demonstration that fail-closes when an input, receipt chain, or evidence claim is compromised;
  • source revision, runtime reachability, downloads, likes, update dates, tags, and direct links to each artifact card.

Public routes

RoutePurposeEvidence boundary
GET /Responsive public AtlasPresentation is navigation, not proof of readiness
GET /launchpadSource-bound Atlas navigation registryA registered destination is not a liveness or capability certificate
GET /healthzRuntime and organ healthReachability does not establish capability or authorization
GET /api/catalogLive public Hub catalogPrivate assets are excluded; unavailable provider data is not inferred
GET /api/estateBounded probes of selected public SpacesHTTP 200 proves reachability only
GET /api/launchpadSource-controlled registered route metadataRegistration is not availability, freshness, or production readiness
GET /api/organs/integrityFive-organ governed-loop demonstrationHealthy output remains advisory and never self-authorizes
GET /api/energyRAPL/NVML runtime probeJoules are reported only when a readable counter exists
GET /api/yarqaBounded YARQA compartmentalization and receipt replaySynthetic input verifies integrity/reproducibility, not CFD correctness
GET /api/build-infoSource-binding metadataMissing runtime revision is labeled REVISION_UNAVAILABLE
GET /demosPython-rendered entry to fixed retrieval and receipt demonstrationsRead-only public fixtures; authority NONE
GET /demos/retrieval/{id}Fixed BM25 lexical queryNamed IDs only; no benchmark or model-quality claim
GET /demos/receipts/{id}Fixed offline receipt verificationIntegrity, signature, and scientific validity shown separately
GET /api/demo-adaptersVersioned adapter artifact registryExact pinned distribution, source revision, wheel hash, limits, authority NONE

Estate probe evidence

/api/estate retains the szl.atlas.estate/v1 envelope, with explicitly scoped surface observations:

  • Remote HTTP 200: honesty: MEASURED, evidence_scope: http_reachability, and reachable: true. This measures only the HTTP observation, never the response's claimed health, capability, authorization, or energy. JSON error bodies, malformed JSON, arrays, and HTML receive the same limited treatment.
  • Other HTTP statuses or transport failures: honesty: UNAVAILABLE, reachable: false, with the observed status (or null for no response).
  • The local five-organ demonstration: evidence_scope: synthetic_kernel, honesty: SIMULATED only for an integer count of five and an explicit blocked: false; an explicit block stays BLOCKED, and incomplete evidence stays UNAVAILABLE. http and reachable are null: no HTTP probe ran.

reachable_surfaces counts only remote HTTP 200 observations; simulated_surfaces counts successful local demonstrations. The legacy live_surfaces field is retained as zero because no probe establishes live capability. Consumers must use these scopes and must not translate legacy LIVE/REACHABLE labels into stronger claims. The Atlas presents these counts without a green readiness indicator. Existing cache bounds remain unchanged; the GET does not mint a signed receipt.

Bounded Python demonstrations

The Atlas has a prominent Try a demo entry. Its original /#loop anatomy URL remains available and expands reliably from a deep link. The new /demos views are rendered deterministically in Python with escaped data; the browser uses HTML/CSS and the existing presentation JavaScript. Fixed links work with the existing form-action 'none' CSP.

The retrieval example ranks five original, Apache-2.0 public corpus summaries with szl-retrieval-bench==0.3.1 BM25. The receipt example calls szl-guardrail-receipt==0.1.2 over committed intact, tampered, and unsigned synthetic records. Both wheels are SHA-256 pinned in requirements-public-demos.txt, and the registry records their exact source tag commits and artifact identities. The core dependency closure is empty for both packages. No signer, model, paid provider, upload, external effect, dynamic import, or visitor URL fetch is used by these adapters.

All receipt fixtures are unsigned. An intact hash and chain can be PASS while signature verification is SKIP; the package's verify_records boolean alone is never displayed as a signature pass. Scientific validity stays NOT_EVALUATED and authority stays NONE. The corpus and fixture provenance are recorded in demo_data/README.md.

Governing boundary

The model proposes. Independent policy decides. A human binds consequential action. The Atlas does not convert a formula, model response, download count, HTTP response, link registration, or signature into permission.

  • Lambda uniqueness remains Conjecture 1 — OPEN.
  • proven_trust remains false in the demonstration.
  • A signature can establish scoped integrity and origin; it does not establish accuracy, safety, performance, compliance, or authorization.
  • Public physical actuation in Killinchu remains SIMULATED unless an operator-owned system provides a separately governed effect path.
  • No production authorization, regulatory approval, customer adoption, benchmark superiority, revenue, funding, or investment outcome is claimed.

Architecture

Hugging Face public APIs ─┐
selected runtime probes ─┼─> catalog + estate state ─> responsive Atlas
five-organ kernel ───────┤
RAPL / NVML probe ───────┤
source route registry ───┘                         └─> source-bound Launchpad

signal → proposal → policy → bounded action → receipt → verification

The Docker Space intentionally uses a compact standard-library Python server. gateway.py composes the permanent /launchpad and /api/launchpad routes onto the existing server.Handler; all other Atlas routes continue through the original handler. Dockerfile publishes the exact server.py, gateway.py, Python demo adapters and renderer, fixed demo data, space/index.html, space/launchpad.html, and source-bound holographic CSS and JavaScript closure. It installs YARQA from its pinned source archive and installs the two demo wheels by exact hashes.

Provider mutation is owned by the protected central reusable publisher in szl-holdings/.github. The thin caller in .github/workflows/hf-sync.yml pins that publisher to an immutable commit, deploys only the Dockerfile-derived file set, requires the exact default-branch tip, binds SZL_GIT_SHA, and verifies same-origin smoke routes before success. Only protected Command Lab main can update the retained SZLHOLDINGS/szl-command-lab Space. This repository retains the canonical source and delegation contract; the Hugging Face Space is the generated runtime mirror.

Run locally

python gateway.py
# open http://127.0.0.1:7860
# open http://127.0.0.1:7860/launchpad

Optional environment variables:

HOST=0.0.0.0
PORT=7860
SZL_GIT_SHA=<exact-40-character-source-revision>

Verification

Energy counter evidence

atlas_energy.py is the single energy implementation used by the Atlas server and the python/energy.py and space/energy.py compatibility entrypoints. The Dockerfile-derived canonical publication includes this module; the existing single-writer workflow also watches its changes.

A missing second read, changed counter identity, decreasing counter, negative value, overflow, NaN or infinity leaves energy UNAVAILABLE with null joules. A valid unchanged counter is a measured zero; it is not confused with a missing sample. RAPL intervals retain the same resolved counter file and NVML intervals the same device UUID. Wrap/reset correction is not inferred. A post-run snapshot cannot substitute for a complete interval around the callable.

The legacy energy_j and inference_energy_j names retain the observed shared counter interval for compatibility. They are not isolated inference energy; the payload and interface state that attribution limit. Ordinary probes do not import torch or invoke a driver CLI; full hardware inventory remains a separate explicit call. No model is loaded by this change.

tests/test_energy_counter_validity.py exercises synthetic counter faults, identity changes, true zeroes, serialization, entrypoint ownership, publication membership and the visible attribution boundary. These are software contracts, not measured hardware energy, CUDA benchmarks or model-quality evidence.

Responsive Atlas

The Atlas owns its navigation and layout. The shared holographic helper retains viewport and accessibility adaptation, while its optional decorative rail and ambient layer are disabled on this native surface. The hero uses the original SZL orbital mark as inline vector geometry inside normal document flow. The mark, caption and governance labels have intrinsic dimensions and never depend on absolutely positioned text. Container queries use the Space's own available width, including an embedded Hugging Face frame and CSS zoom.

The first paths are Explore the work, Build with SZL, and Try a demo. The bounded demo is disclosed on demand, requires a user click to run, and labels successful organ checks as demo results. Model-namespace counts include kernel repositories; kernel listings are not additional trained models.

The atlas-responsive-browser PR job starts the shipped Python gateway and runs scripts/atlas-responsive.test.cjs in Chromium with synthetic provider data. It checks 320, 375, 744, 768, 1024, 1440 and 1920px widths, 200% and 400% CSS zoom, and 320/744px frames inside a 1440px host, plus a 375px coarse-touch case. It verifies that requested CSS zoom was applied, measures the hero box and every visible content boundary with the demo open, then exercises navigation, Escape, catalog search, and 44px touch targets. Screenshots and JSON measurements are retained as a run artifact. This tests rendering; it does not certify live provider data or production capability.

To run the same matrix locally, install the pinned Python demo requirements shown below, the repository's Playwright dependency, and Chromium. Start HOST=127.0.0.1 PORT=7868 python gateway.py, then run:

ATLAS_EVIDENCE_DIR=/tmp/atlas-browser-evidence node --test scripts/atlas-responsive.test.cjs

ATLAS_TEST_ORIGIN may point to another loopback port. The test rejects public origins and intercepts provider API requests with explicit fixtures.

Install the test tool and hash-pinned public demo distributions before starting the gateway or running the suite:

python -m pip install pytest
python -m pip install --require-hashes --only-binary=:all: -r requirements-public-demos.txt
python -m py_compile server.py gateway.py demo_adapters.py visitor_views.py
python - <<'PY'
import gateway
import server

assert server.selftest()["ok"] is True
assert server.build_info()["surface"] == "SZL Atlas"
launchpad = gateway.launchpad_payload()
assert launchpad["schema"] == "szl.atlas.launchpad/v1"
assert launchpad["state"] == "REGISTERED_NAVIGATION"
assert launchpad["public_effectors"] == []
print("atlas + launchpad self-test: PASS")
PY
python -m pytest -q tests
node --test scripts/atlas-estate-ui.test.mjs scripts/atlas-estate-consumer.test.mjs

The dependency-free estate consumer tests use Node 24, matching the public experience CI gate; they execute the shipped static renderer and typed parser with synthetic fixtures, not a browser or a remote production deployment.

The canonical source is szl-holdings/szl-command-lab. The Hugging Face Space is a published runtime mirror and must be checked against its exact source-binding evidence before a consequential deployment decision.


Understand · build · verify

Apache-2.0 · Control before action · Evidence after

apache-2-0
governed-ai
holographic
huggingface
lambda-gate
receipts
spaces
szl-holdings

szl-holdings/szl-command-lab

Holographic command body. Energy MEASURED from RAPL/NVML. GPU train fail-closed. Not flagship.

TypeScript

0

63 commits

updated Oct 19, 2026

See the code

README


title: SZL Atlas — Governed AI Estate emoji: 🧭 colorFrom: gray colorTo: indigo sdk: docker app_port: 7860 pinned: true license: apache-2.0 short_description: Explore SZL systems, models, kernels, and evidence. tags:

  • governed-ai
  • ai-governance
  • provenance
  • inference
  • kernels
  • evidence
  • agentic-ai
  • szl-holdings

SZL Holdings — Orbit v2

SZL Command Lab

Explore SZL systems, models, kernels and datasets in one searchable place. Follow each item to its source, current observations and stated use limits.

Artifact: Public portfolio application

Stage: Exploration and bounded demonstrations

Explore Command Lab → · Build with the source → · Inspect exact source →

Use limits

  • Catalog membership and reachable URLs do not establish model quality or operational readiness.
  • The five-organ demonstration is synthetic; it grants no action authority.
  • Unavailable or partial provider observations remain visible and must not be interpreted as a complete inventory.
Technical documentation, original evidence and reproduction

SZL governed AI command fabric

SZL Atlas

Governed intelligence. Portable evidence.

The public explorer and source-bound Launchpad for SZL systems, models, kernels, datasets, and their verification boundaries.

Open the live Atlas · Enter A11oy · Inspect evidence · Audit the source

What the Atlas does

The Atlas turns the public SZLHOLDINGS estate into one searchable command surface. It recaptures the live, unauthenticated Hugging Face inventory and presents:

  • all publicly listed models, first-class kernels, datasets, and Spaces;
  • curated paths into A11oy, Killinchu, Lyte, Sentra, Terra, PURIQ Finance, PRISM Counsel, Living Anatomy, the Khipu runtime, and YARQA;
  • a source-controlled Launchpad whose registered links remain navigation metadata—not liveness, capability, authorization, or readiness claims;
  • model and evidence spotlights driven by provider metadata rather than hard-coded performance claims;
  • a bounded five-organ demonstration that fail-closes when an input, receipt chain, or evidence claim is compromised;
  • source revision, runtime reachability, downloads, likes, update dates, tags, and direct links to each artifact card.

Public routes

RoutePurposeEvidence boundary
GET /Responsive public AtlasPresentation is navigation, not proof of readiness
GET /launchpadSource-bound Atlas navigation registryA registered destination is not a liveness or capability certificate
GET /healthzRuntime and organ healthReachability does not establish capability or authorization
GET /api/catalogLive public Hub catalogPrivate assets are excluded; unavailable provider data is not inferred
GET /api/estateBounded probes of selected public SpacesHTTP 200 proves reachability only
GET /api/launchpadSource-controlled registered route metadataRegistration is not availability, freshness, or production readiness
GET /api/organs/integrityFive-organ governed-loop demonstrationHealthy output remains advisory and never self-authorizes
GET /api/energyRAPL/NVML runtime probeJoules are reported only when a readable counter exists
GET /api/yarqaBounded YARQA compartmentalization and receipt replaySynthetic input verifies integrity/reproducibility, not CFD correctness
GET /api/build-infoSource-binding metadataMissing runtime revision is labeled REVISION_UNAVAILABLE
GET /demosPython-rendered entry to fixed retrieval and receipt demonstrationsRead-only public fixtures; authority NONE
GET /demos/retrieval/{id}Fixed BM25 lexical queryNamed IDs only; no benchmark or model-quality claim
GET /demos/receipts/{id}Fixed offline receipt verificationIntegrity, signature, and scientific validity shown separately
GET /api/demo-adaptersVersioned adapter artifact registryExact pinned distribution, source revision, wheel hash, limits, authority NONE

Estate probe evidence

/api/estate retains the szl.atlas.estate/v1 envelope, with explicitly scoped surface observations:

  • Remote HTTP 200: honesty: MEASURED, evidence_scope: http_reachability, and reachable: true. This measures only the HTTP observation, never the response's claimed health, capability, authorization, or energy. JSON error bodies, malformed JSON, arrays, and HTML receive the same limited treatment.
  • Other HTTP statuses or transport failures: honesty: UNAVAILABLE, reachable: false, with the observed status (or null for no response).
  • The local five-organ demonstration: evidence_scope: synthetic_kernel, honesty: SIMULATED only for an integer count of five and an explicit blocked: false; an explicit block stays BLOCKED, and incomplete evidence stays UNAVAILABLE. http and reachable are null: no HTTP probe ran.

reachable_surfaces counts only remote HTTP 200 observations; simulated_surfaces counts successful local demonstrations. The legacy live_surfaces field is retained as zero because no probe establishes live capability. Consumers must use these scopes and must not translate legacy LIVE/REACHABLE labels into stronger claims. The Atlas presents these counts without a green readiness indicator. Existing cache bounds remain unchanged; the GET does not mint a signed receipt.

Bounded Python demonstrations

The Atlas has a prominent Try a demo entry. Its original /#loop anatomy URL remains available and expands reliably from a deep link. The new /demos views are rendered deterministically in Python with escaped data; the browser uses HTML/CSS and the existing presentation JavaScript. Fixed links work with the existing form-action 'none' CSP.

The retrieval example ranks five original, Apache-2.0 public corpus summaries with szl-retrieval-bench==0.3.1 BM25. The receipt example calls szl-guardrail-receipt==0.1.2 over committed intact, tampered, and unsigned synthetic records. Both wheels are SHA-256 pinned in requirements-public-demos.txt, and the registry records their exact source tag commits and artifact identities. The core dependency closure is empty for both packages. No signer, model, paid provider, upload, external effect, dynamic import, or visitor URL fetch is used by these adapters.

All receipt fixtures are unsigned. An intact hash and chain can be PASS while signature verification is SKIP; the package's verify_records boolean alone is never displayed as a signature pass. Scientific validity stays NOT_EVALUATED and authority stays NONE. The corpus and fixture provenance are recorded in demo_data/README.md.

Governing boundary

The model proposes. Independent policy decides. A human binds consequential action. The Atlas does not convert a formula, model response, download count, HTTP response, link registration, or signature into permission.

  • Lambda uniqueness remains Conjecture 1 — OPEN.
  • proven_trust remains false in the demonstration.
  • A signature can establish scoped integrity and origin; it does not establish accuracy, safety, performance, compliance, or authorization.
  • Public physical actuation in Killinchu remains SIMULATED unless an operator-owned system provides a separately governed effect path.
  • No production authorization, regulatory approval, customer adoption, benchmark superiority, revenue, funding, or investment outcome is claimed.

Architecture

Hugging Face public APIs ─┐
selected runtime probes ─┼─> catalog + estate state ─> responsive Atlas
five-organ kernel ───────┤
RAPL / NVML probe ───────┤
source route registry ───┘                         └─> source-bound Launchpad

signal → proposal → policy → bounded action → receipt → verification

The Docker Space intentionally uses a compact standard-library Python server. gateway.py composes the permanent /launchpad and /api/launchpad routes onto the existing server.Handler; all other Atlas routes continue through the original handler. Dockerfile publishes the exact server.py, gateway.py, Python demo adapters and renderer, fixed demo data, space/index.html, space/launchpad.html, and source-bound holographic CSS and JavaScript closure. It installs YARQA from its pinned source archive and installs the two demo wheels by exact hashes.

Provider mutation is owned by the protected central reusable publisher in szl-holdings/.github. The thin caller in .github/workflows/hf-sync.yml pins that publisher to an immutable commit, deploys only the Dockerfile-derived file set, requires the exact default-branch tip, binds SZL_GIT_SHA, and verifies same-origin smoke routes before success. Only protected Command Lab main can update the retained SZLHOLDINGS/szl-command-lab Space. This repository retains the canonical source and delegation contract; the Hugging Face Space is the generated runtime mirror.

Run locally

python gateway.py
# open http://127.0.0.1:7860
# open http://127.0.0.1:7860/launchpad

Optional environment variables:

HOST=0.0.0.0
PORT=7860
SZL_GIT_SHA=<exact-40-character-source-revision>

Verification

Energy counter evidence

atlas_energy.py is the single energy implementation used by the Atlas server and the python/energy.py and space/energy.py compatibility entrypoints. The Dockerfile-derived canonical publication includes this module; the existing single-writer workflow also watches its changes.

A missing second read, changed counter identity, decreasing counter, negative value, overflow, NaN or infinity leaves energy UNAVAILABLE with null joules. A valid unchanged counter is a measured zero; it is not confused with a missing sample. RAPL intervals retain the same resolved counter file and NVML intervals the same device UUID. Wrap/reset correction is not inferred. A post-run snapshot cannot substitute for a complete interval around the callable.

The legacy energy_j and inference_energy_j names retain the observed shared counter interval for compatibility. They are not isolated inference energy; the payload and interface state that attribution limit. Ordinary probes do not import torch or invoke a driver CLI; full hardware inventory remains a separate explicit call. No model is loaded by this change.

tests/test_energy_counter_validity.py exercises synthetic counter faults, identity changes, true zeroes, serialization, entrypoint ownership, publication membership and the visible attribution boundary. These are software contracts, not measured hardware energy, CUDA benchmarks or model-quality evidence.

Responsive Atlas

The Atlas owns its navigation and layout. The shared holographic helper retains viewport and accessibility adaptation, while its optional decorative rail and ambient layer are disabled on this native surface. The hero uses the original SZL orbital mark as inline vector geometry inside normal document flow. The mark, caption and governance labels have intrinsic dimensions and never depend on absolutely positioned text. Container queries use the Space's own available width, including an embedded Hugging Face frame and CSS zoom.

The first paths are Explore the work, Build with SZL, and Try a demo. The bounded demo is disclosed on demand, requires a user click to run, and labels successful organ checks as demo results. Model-namespace counts include kernel repositories; kernel listings are not additional trained models.

The atlas-responsive-browser PR job starts the shipped Python gateway and runs scripts/atlas-responsive.test.cjs in Chromium with synthetic provider data. It checks 320, 375, 744, 768, 1024, 1440 and 1920px widths, 200% and 400% CSS zoom, and 320/744px frames inside a 1440px host, plus a 375px coarse-touch case. It verifies that requested CSS zoom was applied, measures the hero box and every visible content boundary with the demo open, then exercises navigation, Escape, catalog search, and 44px touch targets. Screenshots and JSON measurements are retained as a run artifact. This tests rendering; it does not certify live provider data or production capability.

To run the same matrix locally, install the pinned Python demo requirements shown below, the repository's Playwright dependency, and Chromium. Start HOST=127.0.0.1 PORT=7868 python gateway.py, then run:

ATLAS_EVIDENCE_DIR=/tmp/atlas-browser-evidence node --test scripts/atlas-responsive.test.cjs

ATLAS_TEST_ORIGIN may point to another loopback port. The test rejects public origins and intercepts provider API requests with explicit fixtures.

Install the test tool and hash-pinned public demo distributions before starting the gateway or running the suite:

python -m pip install pytest
python -m pip install --require-hashes --only-binary=:all: -r requirements-public-demos.txt
python -m py_compile server.py gateway.py demo_adapters.py visitor_views.py
python - <<'PY'
import gateway
import server

assert server.selftest()["ok"] is True
assert server.build_info()["surface"] == "SZL Atlas"
launchpad = gateway.launchpad_payload()
assert launchpad["schema"] == "szl.atlas.launchpad/v1"
assert launchpad["state"] == "REGISTERED_NAVIGATION"
assert launchpad["public_effectors"] == []
print("atlas + launchpad self-test: PASS")
PY
python -m pytest -q tests
node --test scripts/atlas-estate-ui.test.mjs scripts/atlas-estate-consumer.test.mjs

The dependency-free estate consumer tests use Node 24, matching the public experience CI gate; they execute the shipped static renderer and typed parser with synthetic fixtures, not a browser or a remote production deployment.

The canonical source is szl-holdings/szl-command-lab. The Hugging Face Space is a published runtime mirror and must be checked against its exact source-binding evidence before a consequential deployment decision.


Understand · build · verify

Apache-2.0 · Control before action · Evidence after

apache-2-0
governed-ai
holographic
huggingface
lambda-gate
receipts
spaces
szl-holdings

Project health

Overall

6.6/10

Maintained

0/10

Code review

0/10

Contributors

6/10
OpenSSF Scorecard