OSAHR 0.2

This is the project: SyberLabs/OSAHR_Cell. sykosyber/osahr_project is a private Cursor workspace, not a second OSAHR. See CELL.md.

Exact open stochastic adaptive graph-rewrite over a typed directed hypergraph. Python 3.11+, no kernel runtime dependencies.

The kernel is osahr/. Experiments are confirmatory records on that kernel. Invariants: ARCHITECTURE.md. Licensed packets over Experiment 06: workbench/.

Install

python -m pip install -e .
python examples/adaptive_signal.py

Or python -m pip install osahr-0.2.1-py3-none-any.whl.

What's here

• Kernel: typed hypergraph rewriting, three stochastic schedulers, incremental matching, adaptive parameters, and replay. Incidence-constrained search: docs/INCIDENCE_MATCHING.md.
• Decision workbench: reviewable evidence packets from frozen experiments.
• GrokCell: prototype agent control plane.
• Ontology profile: bounded admission experiment against an ontology snapshot.

Minimal CTMC

from osahr import BoundaryState, Expr, Hypergraph, Model, PatternGraph, Rule, Runtime, Schema, StateAssignment, TemplateGraph

schema = Schema([], [])
graph = Hypergraph(schema)
rule_a = Rule("a", PatternGraph(()), TemplateGraph(()), Expr("p.a"), adaptation=(StateAssignment("memory.last", "a"),))
rule_b = Rule("b", PatternGraph(()), TemplateGraph(()), Expr("p.b"), adaptation=(StateAssignment("memory.last", "b"),))
model = Model(graph, BoundaryState(), (rule_a, rule_b), parameters={"a": 1.0, "b": 3.0}, memory={"last": None})
runtime = Runtime(model, root_seed=42)
print(runtime.step().event.post_time, runtime.memory["last"])

Schedulers: direct SSA, modified next-reaction, bounded thinning. Incremental matching is an optimization; the exhaustive matcher is the oracle. DPO-invalid embeddings are not stochastic channels.

python -m workbench decide workbench/scenarios/03-long-outage.json --out /tmp/osahr-packet
python -m workbench replay /tmp/osahr-packet/decision.json

python -m pip install -e ".[benchmark]" pytest
python -m pytest

Experiments

Every public claim is graded KNOWN, MEASURED, INFERRED, or PROPOSED. Confirmatory seeds are declared before execution. Artifacts are frozen with checksums.

These models explore mechanism. They are not calibrated deployments, not 3GPP-conformance simulators, and not replacements for ns-3, srsRAN, or a commercial RF twin.

Experiment	Question	Status
6G	Does routing by what a task is for preserve value under disruption?	Executed
01	Can an irregular-time recurrent model supply intensities OSAHR consumes as an exact piecewise-constant law?	Executed
02A	Does closing the neural feedback loop improve counterfactual fidelity?	Executed
02B	Can a mechanistically anchored residual improve a twin without sacrificing intervention fidelity?	Executed
03	Is trust in a learned residual a property of the query rather than of the model?	Reanalysis of frozen 02B artifacts
04	Does one calibrated trust coefficient survive across estimands at a shared horizon?	Executed
05	What can be claimed, not merely estimated, about a residual effect?	Formulation only; confirmatory not executed
06	Can a semantic vault and a deterministic controller gate rewrites without licensing unsupported claims?	Executed; seed 260826

Selected MEASURED results:

• Semantic routing pays when capacity must be triaged. 6G: +0.0369 timely goal utility (CI +0.0071 to +0.0677) and +0.0531 critical-task deadline success (CI +0.0142 to +0.0930) over a 15 s edge outage. No-outage control: 0.8778 vs 0.8783. The mechanism is conditional.
• Predictive trust ≠ intervention trust. 02B: residual trust improved factual hazard identification; intervention calibration still selected the mechanistic fallback (18/18 LOSO). One global scalar trust coefficient is not enough.
• Closing a neural loop is not automatically better. 02A: no-jump and frozen-open-loop ablations recovered oracle intervention effects better than the fully closed learned jump model in several regimes.
• Acting and claiming are separate licenses. 06 scenario 3: ensemble signs disagree while every arm's point estimate is negative. The controller could act; it was not licensed to report the sign as a directed effect.

research_directions/ is PROPOSED notes, not science. Do not cite it as MEASURED. GrokCell (grokcell/) is a prototype control-plane host, not a confirmatory record.

GrokCell repair path

The bounded repair path keeps next-action selection deterministic and uses Qwen only to propose candidate code and tests. GrokCell's isolated runner, operator-owned checks, host oracle, and admission gate control whether a revision is accepted. The Jev router and paid routing comparison were removed because the live fixture chain exposes only one productive repair action at a time. No provider-based repair result has been measured. The current Kev migration does not add a decision-model router here; the deterministic repair gate remains the owner of action and admission.

From grokcell/, run python -m grokcell.repair_experiment --check for the offline preflight. A live single-episode run requires Docker, a digest-pinned image, a Qwen credential, and an explicit budget. The preflight makes no provider calls or candidate executions.

The ontology benchmark passed transition-law checks but was substantially slower than plain direct simulation. See docs/VALIDATION_FRAMEWORK.md.

Architecture · Changelog · Repository guide