CRUCIBLE

CRUCIBLE is a contradiction-native symbolic runtime that can change its own ontology when ordinary state repair is not enough. It is the main execution engine behind the DENSN line of work: unresolved contradiction is treated as evidence that the representation itself may need to change.

What This Is

• a CPU-first research runtime for exact contradiction accounting, ontology revision, verifier-backed structural admission, and replayable artifact generation
• a system with two real structure-learning paths:
  • Pathway A for compression of coherent low-conflict structure
  • Pathway B for invention of new latent abstractions when recurrent contradiction persists
• a benchmarked and audited implementation, not a paper sketch

What This Is Not

• not a general-purpose language model or agent framework
• not a production distributed system
• not a raw-log ingestion system
• not a general scientific discovery engine
• not a GPU runtime on this machine

Release Status

• research runtime
• CPU-only
• benchmark-bounded
• public package version is alpha: 0.1.0a1
• source-available under PolyForm Strict 1.0.0, not OSI open source
• public evidence is governed under artifacts/current_evidence/
• real separate-host distributed proof is still deferred

Commercial or broader implementation rights are not granted by the default public license. See COMMERCIAL-LICENSING.md.

Hard Boundaries

These limits are part of the public release contract, not footnotes:

• CPU-only release: the public implementation and evidence are CPU-only on this machine; there is no GPU claim
• distributed boundary: distributed execution code exists, but the public release does not ship distributed execution as governed current evidence and it does not prove separate-host deployment
• external blind benchmark boundary: the blind external invariant assessment depends on a frozen symbolic projection over external incident traces; it is not raw-log-native ingestion
• scientific boundary: the scientific family is a bounded secondary benchmark wedge with evidence-bearing behavior; it is not a general scientific discovery system
• release maturity boundary: this is an alpha research runtime, not a production system claim

Claims Supported By Current Evidence

These claims are supported by the governed current-evidence corpus, not by private notes or stale local outputs.

• exact contradiction accounting, exact DeltaPsi, and exact admission/rejection decisions
• persistent ontology state with checkpoint/resume, lineage replay, and retirement accounting
• real Pathway B invention on the formal wedge
• real Pathway A compression on the formal wedge
• cross-task train-authority propagation with negative-transfer blocking
• verifier-backed admission on the formal wedge
• scientific benchmark behavior as a bounded secondary evidence-bearing wedge
• native CPU production path with explicit reference-mode fallback only when requested

Not Claimed

• no claim of physically separate multi-host proof
• no claim of raw external log ingestion without a frozen symbolic projection step
• no claim of broad real-world generalization beyond the benchmark families in this repo
• no claim that the scientific benchmark family is a general discovery engine
• no claim of commercial production readiness

Current Evidence Envelope

The governed public evidence corpus is tracked by artifacts/current_evidence/current_evidence_manifest.json. That corpus is intentionally small. It currently ships exactly six representative bundles:

• xor
• pathway_a_compression
• wedge_protocol
• joint_two_train_shared_latent
• cross_task_negative_transfer
• scientific_discovery

Each public current-evidence bundle is a curated release extract, not the full local working artifact warehouse. The shipped files are the reviewer-facing records listed in What The Public Evidence Bundle Contains below.

The separate public assessment surface lives under artifacts/assessments, indexed by assessment_manifest.json, and including the blind external invariant assessment. Non-current phase history and bulky working archives are intentionally not part of the public repo surface.

Install

python -m pip install -e .[dev]
crucible-build-kernels --report artifacts/reports/kernel_build_report.json

CRUCIBLE production runs default to native_execution_policy: require. On a fresh install, build the native CPU extension before running benchmark CLIs. The report path above is a local generated diagnostic, not a committed evidence artifact.

Smoke Test

pytest -q tests/integration/test_public_release_surface.py tests/integration/test_current_evidence_corpus.py
crucible-run-benchmark --benchmark xor --config configs/cpu_reference.yaml --output artifacts/demo_xor

That gives you one fast correctness pass plus a minimal benchmark run with the normal artifact bundle.

Reproducing Key Results

Formal protocol wedge:

crucible-run-benchmark --benchmark wedge_protocol --config configs/wedge_protocol.yaml --output artifacts/demo_protocol

Joint multi-task train-authority benchmark:

crucible-run-benchmark --benchmark joint_two_train_shared_latent --config configs/joint_multitask.yaml --output artifacts/demo_joint_multitask

Scientific benchmark family:

crucible-run-benchmark --benchmark scientific_discovery --config configs/scientific_discovery.yaml --output artifacts/demo_scientific

Blind external invariant assessment:

crucible-run-external-assessment --output artifacts/assessments/demo_external_assessment

The external assessment is intentionally outside the governed current-evidence corpus. It is a harder evaluation path that uses a frozen symbolic projection over an external incident corpus.

Public CLI Surface

The intentionally public console entrypoints are:

• crucible-build-kernels
• crucible-run-cycle
• crucible-run-benchmark
• crucible-run-ablation
• crucible-run-profile
• crucible-run-worker
• crucible-run-external-assessment

Maintainer-only corpus refresh and release tooling are intentionally not part of the public console-script surface.

What The Public Evidence Bundle Contains

The runtime can emit much richer bundles than the public repo ships. The public release keeps a curated reviewer-facing subset for each representative run:

• final_proof_bundle.json
• artifact_audit.json
• config_snapshot.json
• environment_manifest.json
• runtime_trace.jsonl
• candidate_lifecycle.json
• verifier_results.json
• verifier_agreements.json
• lineage_history.json
• lineage_replay.json
• retirement_history.json
• semantic_history.json
• reuse_outcomes.json
• cross_task_manifest.json
• cross_task_candidates.json
• authority_audit.json

The full runtime contract is still documented in docs/telemetry_contract.md. The public repo simply does not ship every optional or bulky artifact the runtime can produce.

Repository Layout

• src/crucible: runtime, ontology state, dynamics, TSL, verifiers, and CLIs
• configs: reproducible benchmark and runtime configurations
• tests: unit, integration, property, perf, and golden checks
• artifacts/current_evidence: governed public evidence surface
• artifacts/assessments: public assessment packs outside the live claim corpus
• artifacts/assessments/assessment_manifest.json: public assessment index
• artifacts/reports: location for locally regenerated auxiliary reports
• docs: public technical documentation

Further Reading

• docs/README.md
• docs/architecture.md
• docs/benchmark_contract.md
• docs/telemetry_contract.md
• docs/math.md

Governance

Public repository governance files:

• LICENSE
• CHANGELOG.md
• CODEOWNERS
• CONTRIBUTING.md
• SECURITY.md
• CODE_OF_CONDUCT.md
• SUPPORT.md
• CITATION.cff