Skip to content

About

Oscillink Safety Ops is an independent safety and risk-mitigation supervisor for AI-controlled industrial equipment, connecting machine intent, observed behavior, and safety-manager oversight.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

54 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Oscillink Safety Ops

Oscillink Safety Ops is an independent safety and risk-mitigation supervisor for AI-controlled industrial equipment, connecting machine intent, observed behavior, and safety-manager oversight.

Oscillink

The production AI runs the machine. Oscillink independently monitors whether commanded and observed behavior remain within the approved operating envelope and requests a protective response when they do not.

The repository implements the governed evidence and offline-evaluation plane plus a deterministic closed-record runtime supervisor, persistent intervention/recovery state, robot-cell replay, local simulated one-way protective-request records, and a generated read-only synthetic monitor. It does not connect to equipment or establish that a physical stop occurred.

Current status: experimental private product with implemented deterministic contracts. The code makes no legal, compliance, certification, safe-operation, work-authorization, deployment, practitioner-validation, or production-readiness claim.

Synthetic Safety Evidence Packet showing exact source identity, unresolved evidence, and fixed no-authority states

Project-authored synthetic demonstration. It is not facility, customer, legal, or operational evidence.

Make the evidence inspectable

Physical-intelligence plans and recordings do not carry the authority needed to establish what rules apply, whether a procedure is current, whether evidence is missing, or who reviewed a decision. Safety Ops keeps those questions explicit.

Given an identified asset, task, and bounded source set, the current contract surface can:

  • verify exact local source bytes, revisions, locations, and SHA-256 identities;
  • preserve asset, model, serial, site, task, role, and observation context;
  • separate untrusted extraction candidates from external human review;
  • expose stale, conflicting, missing, ambiguous, unreadable, and unsupported evidence;
  • retain corrections, retractions, supersession, and review lineage;
  • assemble a versioned Safety Evidence Packet; and
  • evaluate proposed plans and recorded episodes offline as cited evidence findings.

Unknown evidence stays unknown. Reviews do not silently carry onto changed bytes or revisions.

How it works

Oscillink Safety Ops governed evidence architecture, ending in offline review with no physical control path

Source bytes, extraction candidates, normalization, review decisions, and evaluation outcomes stay separate. Every transformation retains exact identity. No arrow reaches equipment control.

Product state

Implemented and tested locally In validation Not claimed
Physical Intelligence Evidence Envelope Broader official-publication parser coverage Legal correctness or regulatory applicability
Safety Evidence Packet v1 Qualified-practitioner workflow fit Compliance determination or certification
Exact-byte and source-revision verification Sanitized real-work example bundles Safe operation or work authorization
Official-source regulatory reconciliation contracts Additional read-only operational adapters Production readiness or customer outcomes
Federal Register and LSA source-change lineage Independent hidden-evaluation scoring Incident prevention or commercial traction
Metadata-only licensed-standard registry Hosted CI for local-only maturation commits Equipment, PLC, interlock, or robot control
Read-only operational JSONL evidence Legal, engineering, and OT-owner review Automatic conflict resolution or policy promotion
Offline plan and episode evaluation Public release and publication audit Autonomous approval of model-generated constraints
Deterministic schemas, fixtures, and hidden-test protocol
36-case exact-byte synthetic robot-cell benchmark Physical stopping or field effectiveness
Dependency-free read-only static evidence monitor A runtime operator/control interface

Tests establish deterministic software behavior only. They do not establish that a source applies, that an interpretation is legally correct, or that an operation is safe.

Quickstart

Python 3.11 and uv are required. All demonstration inputs are project-authored and synthetic.

uv sync --locked --dev
PYTHONPATH= uv run python scripts/verify.py
PYTHONPATH= uv run safety-ops benchmark verify --root benchmark/robot_cell_v1
PYTHONPATH= uv run python scripts/verify_demo.py demo
PYTHONPATH= uv run safety-ops episode-evaluate \
  --packet tests/fixtures/synthetic_press/safety-evidence-packet-v1.json \
  --episode tests/fixtures/synthetic_press/episode.json \
  --envelope tests/fixtures/synthetic_press/episode-envelope.json \
  --root tests/fixtures/synthetic_press

Open demo/index.html directly to inspect all 36 generated synthetic scenarios without a server. The interface can only select a scenario and inspect evidence. It has no command, reset, rearm, acknowledgment, stop, form-submit, runtime-machine, or external-network surface. Read SAFETY_MANAGER_DEMO.md before interpreting the results.

The episode report binds exact packet and payload identities and emits closed evidence states. Its fixed top-level output remains:

evaluation_state       evidence_findings_only
compliance_state       no_conclusion
operational_authority  none

See the step-by-step synthetic demonstration for source-envelope, plan, episode, and operational-evidence commands.

Evidence and authority boundaries

Safety Ops may emit cited evidence states such as:

  • matched
  • missing_evidence
  • asset_mismatch
  • revision_stale
  • source_conflict
  • ambiguous
  • unreadable
  • unsupported_interpretation
  • requires_authorized_review

It cannot emit or grant:

  • a compliance conclusion or legal opinion;
  • certification, a work permit, or lockout/tagout authorization;
  • approval to operate;
  • an automatic resolution among regulations, standards, manuals, SOPs, or plans;
  • safety PLC, interlock, emergency-stop, controller, robot, machine, or actuator changes; or
  • automatic review carry-forward after source bytes, identity, configuration, or revision changes.

Safety Ops is not a safety-rated control system or a replacement for qualified EHS, legal, maintenance, integration, or safety-engineering review.

Evidence domains

Regulatory evidence

The current reconciliation contract preserves exact GovInfo annual CFR, dated eCFR, Federal Register, and GovInfo LSA identities. Exact text equality proves deterministic equality only. Accepted source-change bundles establish reviewed source lineage, not legal meaning or compliance.

Licensed standards

The current registry stores official metadata and explicit rights states only. It contains no licensed standard text or derived requirements. Full-text processing stays blocked until lawful access and compatible storage and processing rights are confirmed.

Operational and episode evidence

Operational adapters and episode evaluators are offline and read-only. They preserve source identity, quality, ordering, calibration, missing fields, and parser warnings without filling records or sending commands back to source systems.

Read the technical overview for the full implemented contract surface.

Who this is for

The current audience order is:

  1. Safety, compliance, and assurance leaders defining evidence and review boundaries.
  2. Physical-intelligence builders and integrators preparing systems for real-world operations.
  3. Operations, maintenance, and OT owners evaluating read-only evidence fit and authority limits.
  4. Researchers and evaluators studying provenance, source reconciliation, and offline assurance.

The repository is not currently positioned as a self-service compliance product for operators or a replacement for existing CMMS, checklist, document-management, or control systems.

Verification status

The canonical verifier checks text hygiene, generated schemas, the OSHA source catalog, synthetic fixtures, the 36-case exact-byte benchmark, byte-identical static-demo generation and no-control/ no-network interface rules, Ruff, formatting, strict mypy, package builds, and the full test suite.

Exact-SHA verification has passed on Windows and an independent Linux Buildbox for the current implemented feature baseline. Hosted CI has not evaluated local-only maturation commits because they have not been pushed. This is reproducibility evidence, not release, safety, compliance, applicability, or practitioner-validation evidence.

See Assurance status and limitations for the current implementation, planned runtime-supervision boundary, deployment dependencies, and claims that remain unsupported.

Documentation

License

Oscillink Safety Ops is licensed under the Apache License 2.0. The license applies to the project-authored source and documentation unless a file states otherwise. It does not grant rights to third-party standards, manuals, customer procedures, datasets, model weights, services, or trademarks.

About

Oscillink Safety Ops is an independent safety and risk-mitigation supervisor for AI-controlled industrial equipment, connecting machine intent, observed behavior, and safety-manager oversight.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages