eif-engine 4.1.0__tar.gz → 4.3.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.env.example +8 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/CHANGELOG.md +84 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/DATA_PROCESSING_AGREEMENT.md +4 -3
- {eif_engine-4.1.0 → eif_engine-4.3.0}/LEGAL.md +2 -2
- {eif_engine-4.1.0 → eif_engine-4.3.0}/PKG-INFO +139 -66
- {eif_engine-4.1.0 → eif_engine-4.3.0}/PRIVACY_POLICY.md +26 -3
- {eif_engine-4.1.0 → eif_engine-4.3.0}/README.md +130 -60
- {eif_engine-4.1.0 → eif_engine-4.3.0}/SECURITY.md +32 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/deploy/do_app_spec.yaml +11 -1
- eif_engine-4.3.0/dev_factory/standard_ei_01_nist_ai_risk_management_framework/bind.py +24 -0
- eif_engine-4.3.0/dev_factory/standard_ei_01_nist_ai_risk_management_framework/test_bind.py +34 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/docs/architecture.md +5 -3
- eif_engine-4.3.0/docs/eif_host_tenant_keys_tech_spec.md +138 -0
- eif_engine-4.3.0/docs/integrations/CLAUDE_CODE.md +205 -0
- eif_engine-4.3.0/docs/integrations/CLAUDE_DESKTOP.md +157 -0
- eif_engine-4.3.0/docs/integrations/COMPLYEDGE.md +170 -0
- eif_engine-4.3.0/docs/integrations/CURSOR.md +184 -0
- eif_engine-4.3.0/docs/integrations/HOST_APP.md +118 -0
- eif_engine-4.3.0/docs/integrations/README.md +95 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/docs/research_foundation.md +3 -5
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/__init__.py +1 -1
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/cost_model/cost_ledger.py +1 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/eif_engine_intent.yaml +1 -2
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/eif_evidence_collection_intent.yaml +1 -1
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/native_tools.py +1 -1
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/input_guard/eif_input_guard_intent.yaml +2 -2
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/integration/interceptor.py +9 -5
- eif_engine-4.3.0/eif/mcp_server/auth.py +189 -0
- eif_engine-4.3.0/eif/mcp_server/compliance.py +285 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/mcp_server/http_server.py +90 -7
- eif_engine-4.3.0/eif/mcp_server/runtime_compliance_intent.yaml +159 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/mcp_server/server.py +277 -44
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/schemas.py +10 -1
- eif_engine-4.3.0/eif/session.py +116 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sycophancy/cost_model.py +1 -2
- {eif_engine-4.1.0 → eif_engine-4.3.0}/index.html +2 -2
- {eif_engine-4.1.0 → eif_engine-4.3.0}/pyproject.toml +9 -5
- {eif_engine-4.1.0 → eif_engine-4.3.0}/scripts/compliance/runtime_check.sh +11 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/server.json +2 -14
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/integration/test_mcp_server.py +72 -4
- eif_engine-4.3.0/tests/unit/test_complyedge_cold_start.py +60 -0
- eif_engine-4.3.0/tests/unit/test_cost_ledger.py +31 -0
- eif_engine-4.3.0/tests/unit/test_integration_interceptor_langchain.py +33 -0
- eif_engine-4.3.0/tests/unit/test_runtime_check_attribution.py +69 -0
- eif_engine-4.3.0/tests/unit/test_runtime_compliance.py +364 -0
- eif_engine-4.3.0/tests/unit/test_sdk_interceptor.py +62 -0
- eif_engine-4.3.0/tests/unit/test_tenant_keys.py +353 -0
- eif_engine-4.3.0/tests/unit/test_verify_zero_evidence_revise.py +60 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/validation/CORPUS_REPORT.md +1 -6
- eif_engine-4.1.0/docs/integrations/COMPLYEDGE.md +0 -99
- eif_engine-4.1.0/eif/mcp_server/auth.py +0 -81
- eif_engine-4.1.0/eif/session.py +0 -40
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.dockerignore +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.github/CODEOWNERS +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.github/ISSUE_TEMPLATE/bug_report.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.github/ISSUE_TEMPLATE/config.yml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.github/dependabot.yml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.github/pull_request_template.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.github/workflows/ci.yml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.gitignore +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.pre-commit-config.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/.trustlint.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/CODE_OF_CONDUCT.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/CONTRIBUTING.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/DECISIONS.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/Dockerfile +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/LICENSE +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/NOTICE +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/TERMS_OF_SERVICE.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/docs/compliance.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/docs/eif_causal_gate_v4_tech_spec.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/docs/eif_s17_s18_s19_tech_spec.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/docs/eif_v2_tech_spec.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/auto.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/calibrate/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/calibrate/bayesian.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/calibrate/conformal.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/calibrate/domain_constraints.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/calibrate/ece.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/calibrate/prior_strategy.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/calibrate/trust.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/causal_gate/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/causal_gate/confound.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/causal_gate/direction.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/causal_gate/eif_causal_gate_v4_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/causal_gate/evidence_probe.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/causal_gate/intervention.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/causal_gate/iris_pipeline.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/causal_gate/verdict.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/challenge/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/challenge/diversity.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/challenge/eif_multi_critic_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/challenge/multi_critic.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/challenge/protocol.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/challenge/replicator.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/cost_model/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/cost_model/catch_rate_empirical.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/cost_model/eif_catchrate_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/declare/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/declare/extractor.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/declare/harking_guard.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/declare/registry.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/eif_v2_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/eif_v3_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/explain/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/explain/artifact.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/explain/eif_explain_coldstart_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/explain/hard_to_vary.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/explain/reach.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/condition.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/eif_native_tools_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/eif_v5_evidence_trust_weighting_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/evidence_collector.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/hard_to_vary.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/sprt.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/falsify/trivial_check.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/agenda.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/eif_hypothesis_agenda_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/fdr.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/scorer.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/input_guard/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/input_guard/detector.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/integration/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/integration/eif_auto.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/mcp_server/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/mcp_server/resources.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/memory/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/memory/strategy.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/programme/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/programme/monitor.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/programme/principle_memory.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/programme/principle_revision.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/programme/signals.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/programme/status.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/py.typed +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/record/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/record/chain.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/record/compliance.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/record/isc_disclosure.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/record/outcome_store.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/record/provenance.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/record/research_object.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/replicate/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/replicate/divergence.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/replicate/eif_independent_replication_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/replicate/independent.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/replicate/protocol.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/reporting/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/reporting/pdf_renderer.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sdk/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sdk/exceptions.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sdk/interceptor.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sycophancy/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sycophancy/detector.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sycophancy/drift.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sycophancy/eif_sycophancy_intent.yaml +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sycophancy/faithfulness.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/sycophancy/framing.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/update/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/update/eig.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/update/paradigm.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/update/posterior.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/eif/update/stopping.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/examples/quickstart.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/README.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/ai_agents/calibration.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/ai_agents/causal_reasoning.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/ai_agents/hypothesis_generation.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/ai_agents/multi_agent_review.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/ai_agents/overview.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/ai_agents/provenance.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/foundations/ai_and_science.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/foundations/classical_method.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/foundations/deutsch_epistemology.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/foundations/frontier_debates.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/foundations/philosophy_of_science.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/foundations/replication_crisis.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/foundations/statistics.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/research/sycophancy_bias_research.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/scripts/compliance/check.sh +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/conftest.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/integration/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/integration/test_eif_loop.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/__init__.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_calibrate.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_catch_rate_empirical.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_causal_gate.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_dases_adaptive.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_dases_replicator.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_declare.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_deeptrace_live.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_ece_v2.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_evidence_trust_weighting.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_explain.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_exploratory_confirmatory.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_falsify.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_fdr.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_framing_elephant.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_hard_to_vary.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_host_tool_numeric.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_http_rate_limit.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_independent_replication.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_iris_pipeline.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_isc_disclosure.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_memory.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_monica_drift.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_native_tools_nt3.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_outcome_store.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_p1_sandbox_security.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_paradigm.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_principle_memory.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_principle_revision.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_programme.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_record.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_replicate.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_research_object.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_schemas.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/tests/unit/test_update.py +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/validation/EIF_INVESTMENT_V31_EVIDENCE.md +0 -0
- {eif_engine-4.1.0 → eif_engine-4.3.0}/validation/data/.gitkeep +0 -0
|
@@ -8,3 +8,11 @@ EIF_API_KEY=
|
|
|
8
8
|
|
|
9
9
|
# Optional. Enables the LLM critic tournament (CHALLENGE) and the causal evidence probe. The core engine never calls an LLM.
|
|
10
10
|
OPENAI_API_KEY=
|
|
11
|
+
|
|
12
|
+
# Optional. ComplyEdge runtime screening of every tool call and /verify request
|
|
13
|
+
# (docs/integrations/COMPLYEDGE.md). Leave unset and EIF makes no ComplyEdge call.
|
|
14
|
+
COMPLYEDGE_API_KEY=
|
|
15
|
+
# COMPLYEDGE_API_URL=https://api.complyedge.io
|
|
16
|
+
# COMPLYEDGE_AGENT_ID=eif-mcp
|
|
17
|
+
# COMPLYEDGE_JURISDICTION=EU
|
|
18
|
+
# COMPLYEDGE_TIMEOUT_S=5
|
|
@@ -7,6 +7,90 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
|
|
|
7
7
|
|
|
8
8
|
## [Unreleased]
|
|
9
9
|
|
|
10
|
+
## [4.3.0] - 2026-11-20
|
|
11
|
+
|
|
12
|
+
### Changed
|
|
13
|
+
- `eif_verify` returns REVISE, not PASS, when nothing contradicts but a checked
|
|
14
|
+
claim stays below posterior 0.70 (no evidence leaves it at 0.5). PASS now
|
|
15
|
+
means every checked claim reached 0.70. Same thresholds as `eif_calibrate`.
|
|
16
|
+
Callers that only check for HALT keep working. Rules block v1.5.
|
|
17
|
+
|
|
18
|
+
### Added
|
|
19
|
+
- `[search]` extra installs `ddgs` for P3 web search. Kept opt-in because it
|
|
20
|
+
sends claim-derived search terms off the machine.
|
|
21
|
+
- Host-derived tenant keys: one self-hosted HTTP server can serve many tenants,
|
|
22
|
+
each with its own sessions and a three-tool allowlist. See
|
|
23
|
+
`docs/integrations/HOST_APP.md`.
|
|
24
|
+
|
|
25
|
+
### Fixed
|
|
26
|
+
- The `eif_verify` tool description gave a `host_tool_outputs` shape the code
|
|
27
|
+
ignores; it now gives `{tool_name, query, result, data_scope}`.
|
|
28
|
+
- Removed the claim that `eif_verify` catches correlation presented as causation
|
|
29
|
+
from the server instructions and the `eif_verify` description (4.2.0 removed
|
|
30
|
+
it from the README only).
|
|
31
|
+
- Docs: a plain install has no evidence source; web search needs `ddgs`;
|
|
32
|
+
`pip install eif` is a different project.
|
|
33
|
+
- Docs: Claude Desktop does not connect a remote `url` from its config file;
|
|
34
|
+
the hosted option now uses the `mcp-remote` bridge with the key in `env`.
|
|
35
|
+
- README corpus table: noted that the verdict columns add to 75 of 84 turns.
|
|
36
|
+
- serverInfo reports EIF's version, not the mcp library's.
|
|
37
|
+
- `eif.integration.EIFInterceptor` failed on langchain 1.x
|
|
38
|
+
(`No module named 'langchain.callbacks'`); it now imports from
|
|
39
|
+
`langchain_core` and needs only `langchain-core`. The README example passes
|
|
40
|
+
`get_langchain_handler()` and sets `intercept_chains=True` for chains.
|
|
41
|
+
|
|
42
|
+
## [4.2.0] - 2026-11-05
|
|
43
|
+
|
|
44
|
+
### Fixed
|
|
45
|
+
- The published MCP install failed with ImportError: `mcp` is an extra, and the
|
|
46
|
+
4.1.0 metadata allowed mcp 2.x. Install lines now use `eif-engine[mcp]`, and
|
|
47
|
+
the extra is pinned to `mcp>=1.19,<2`.
|
|
48
|
+
- Removed the claim that `eif_verify` catches correlation presented as causation.
|
|
49
|
+
`eif_causal_gate` does that.
|
|
50
|
+
|
|
51
|
+
- **ComplyEdge checks wait 15 s by default (was 5).** A cold ComplyEdge instance answers in
|
|
52
|
+
about 8 s; with 5 s the check was skipped, so the caller lost the verdict and the Log ID
|
|
53
|
+
of a record ComplyEdge had written. `COMPLYEDGE_TIMEOUT_S` still overrides it.
|
|
54
|
+
### Added
|
|
55
|
+
- ComplyEdge runtime screening on the MCP server and `/verify`. With
|
|
56
|
+
`COMPLYEDGE_API_KEY` set, every tool call (all 25 tools, through one
|
|
57
|
+
dispatch, `ComplianceFastMCP.call_tool`) and every `/verify` request is
|
|
58
|
+
checked through `POST /v1/check` twice: the input before it runs and the
|
|
59
|
+
result before it returns. A block returns the rule instead of the answer
|
|
60
|
+
(`HALT` with routing `COMPLYEDGE_BLOCKED` on `/verify`). Fail-open when
|
|
61
|
+
ComplyEdge is unreachable; no call at all without the key. Audit fields:
|
|
62
|
+
`user_id` is `eifkey:<sha256 prefix>` of the caller's key (never the key)
|
|
63
|
+
or `local:<login>`, `user_role` is `mcp_client` / `sdk_client` /
|
|
64
|
+
`maintainer`, `session_id` is the EIF or MCP session.
|
|
65
|
+
A call made before any session exists (`eif_new_session`) records the
|
|
66
|
+
session it created on its output check, so it links to the calls after it.
|
|
67
|
+
- The CI runtime probe sends `user_id`, `user_role` and `session_id`.
|
|
68
|
+
- ComplyEdge Log IDs reach the caller. Every checked MCP tool call returns
|
|
69
|
+
`_meta["complyedge"] = {input_event_id, output_event_id}`, and `/verify`
|
|
70
|
+
returns them as `complyedge-input-event-id` / `complyedge-output-event-id`
|
|
71
|
+
headers, both outside the checked text, which is returned byte for byte.
|
|
72
|
+
Log IDs of calls on an EIF session are listed on it, copied onto the
|
|
73
|
+
`ProvenanceRecord` (`complyedge_event_ids`) by `eif_record` and printed by
|
|
74
|
+
`eif_compliance_report`. With the caller's copy of the text, each one
|
|
75
|
+
verifies through ComplyEdge `POST /v1/audit/verify`.
|
|
76
|
+
|
|
77
|
+
### Fixed
|
|
78
|
+
- `eif_check_rules_installed` no longer reports a definitive miss when the MCP
|
|
79
|
+
server cannot read any paths (hosted endpoint / wrong CWD). Returns
|
|
80
|
+
`indeterminate=true` and `reachable_filesystem=false` with a clear `note`.
|
|
81
|
+
- Hosted-safe path: optional `file_content` (+ `file_content_label`) lets the
|
|
82
|
+
agent Read a local rules file and forward its text for marker detection.
|
|
83
|
+
- Pin optional `mcp` extra to `>=1.0,<2` so CI/installs do not pull MCP SDK
|
|
84
|
+
v2 (FastMCP → MCPServer rename) and break `from mcp.server.fastmcp import FastMCP`.
|
|
85
|
+
|
|
86
|
+
### Changed
|
|
87
|
+
- Privacy Policy v1.2 and DPA template v1.1 describe the ComplyEdge flow.
|
|
88
|
+
- `mcp` minimum raised to 1.19 (tools return `CallToolResult` with `_meta`).
|
|
89
|
+
- Privacy Policy v1.3: ComplyEdge keeps a keyed fingerprint and pseudonymous
|
|
90
|
+
IDs for 184 days, and the Log IDs come back to the caller.
|
|
91
|
+
- `eif_explain` tool description now documents required detail keys
|
|
92
|
+
(`detail_text`, `prediction_impact`).
|
|
93
|
+
|
|
10
94
|
## [4.1.0] - 2026-07-19
|
|
11
95
|
|
|
12
96
|
### Added
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
# Data Processing Agreement (Template)
|
|
2
2
|
|
|
3
|
-
> **Version:** 1.
|
|
3
|
+
> **Version:** 1.1 · **Effective:** 2026-10-01
|
|
4
4
|
>
|
|
5
5
|
> **Template. Not in force.** EIF's maintainer does not process user data (see
|
|
6
6
|
> [PRIVACY_POLICY.md](PRIVACY_POLICY.md)), so no DPA with the maintainer exists or is
|
|
@@ -42,8 +42,8 @@ collection, Bayesian calibration, routing (ACT/REVISE/HALT), and generation of
|
|
|
42
42
|
provenance records, together with storage of the resulting local artifacts.
|
|
43
43
|
|
|
44
44
|
`[ADAPT: describe your deployment's actual purpose and any enabled optional features,
|
|
45
|
-
including outbound evidence collection (DuckDuckGo)
|
|
46
|
-
transmit claim text to those third parties.]`
|
|
45
|
+
including outbound evidence collection (DuckDuckGo), OpenAI-backed critique and
|
|
46
|
+
ComplyEdge runtime screening, which transmit claim text to those third parties.]`
|
|
47
47
|
|
|
48
48
|
---
|
|
49
49
|
|
|
@@ -102,6 +102,7 @@ Authorized sub-processors at signature:
|
|
|
102
102
|
| Sub-processor | Role | Condition |
|
|
103
103
|
|---------------|------|-----------|
|
|
104
104
|
| `[e.g., OpenAI]` | `[Critic tournament and causal evidence probe, only if enabled]` | `[Data processing terms reference]` |
|
|
105
|
+
| `[e.g., ComplyEdge]` | `[EU AI Act screening of tool inputs and results, only if COMPLYEDGE_API_KEY is set; stores a keyed HMAC-SHA-256 fingerprint of the text, not the text]` | `[https://complyedge.io/legal/data-processing-agreement.html]` |
|
|
105
106
|
| `[e.g., search provider]` | `[Web evidence collection, only if enabled]` | `[Terms reference]` |
|
|
106
107
|
| `[hosting/infrastructure]` | `[Infrastructure for the deployment]` | `[DPA reference]` |
|
|
107
108
|
|
|
@@ -95,8 +95,8 @@ EIF's documentation references quantified figures. Their scope and evidentiary b
|
|
|
95
95
|
| Claim | Source | Scope | Limitation |
|
|
96
96
|
|-------|--------|-------|------------|
|
|
97
97
|
| Validation corpus results (routing counts, detector fires) | 14-scenario validation corpus | **Constructed, illustrative scenarios** authored by the maintainer | Not real agent traffic; not a measurement of production behavior |
|
|
98
|
-
|
|
|
99
|
-
|
|
|
98
|
+
| Per-HALT dollar cost estimates (the cost-model feature output, and the figures in `validation/CORPUS_REPORT.md`) | Corpus cost model | **Modeled estimates** derived from assumed domain cost multipliers applied to corpus routings | Not money actually saved or lost by any deployment; the multipliers are assumptions, not observations |
|
|
99
|
+
| Any third-party market-size or industry-cost statistic cited anywhere in this repository | The cited third party | Third-party estimate | Not EIF's research; cite the original source for any compliance or commercial purpose |
|
|
100
100
|
|
|
101
101
|
Full corpus evidence and the cost-model methodology:
|
|
102
102
|
[`validation/CORPUS_REPORT.md`](validation/CORPUS_REPORT.md).
|
|
@@ -1,7 +1,7 @@
|
|
|
1
|
-
Metadata-Version: 2.
|
|
1
|
+
Metadata-Version: 2.5
|
|
2
2
|
Name: eif-engine
|
|
3
|
-
Version: 4.
|
|
4
|
-
Summary: Epistemic Integrity Framework
|
|
3
|
+
Version: 4.3.0
|
|
4
|
+
Summary: EIF (Epistemic Integrity Framework): runtime gate for AI agent claims. Catches sycophantic drift, circular self-validation, causal errors.
|
|
5
5
|
Project-URL: Homepage, https://github.com/leocelis/eif
|
|
6
6
|
Project-URL: Repository, https://github.com/leocelis/eif
|
|
7
7
|
Project-URL: Issues, https://github.com/leocelis/eif/issues
|
|
@@ -21,17 +21,18 @@ Requires-Dist: numpy>=1.26
|
|
|
21
21
|
Requires-Dist: pydantic>=2.0
|
|
22
22
|
Requires-Dist: scipy>=1.11
|
|
23
23
|
Provides-Extra: dev
|
|
24
|
+
Requires-Dist: langchain-core>=0.2; extra == 'dev'
|
|
24
25
|
Requires-Dist: pytest-asyncio>=0.21; extra == 'dev'
|
|
25
26
|
Requires-Dist: pytest>=7.0; extra == 'dev'
|
|
26
27
|
Requires-Dist: ruff>=0.4; extra == 'dev'
|
|
27
28
|
Provides-Extra: full
|
|
28
29
|
Requires-Dist: fpdf2>=2.7; extra == 'full'
|
|
29
30
|
Requires-Dist: langchain-core>=0.2; extra == 'full'
|
|
30
|
-
Requires-Dist: mcp
|
|
31
|
+
Requires-Dist: mcp<2,>=1.19; extra == 'full'
|
|
31
32
|
Requires-Dist: openai>=1.0; extra == 'full'
|
|
32
33
|
Requires-Dist: sentence-transformers>=2.2; extra == 'full'
|
|
33
34
|
Provides-Extra: mcp
|
|
34
|
-
Requires-Dist: mcp
|
|
35
|
+
Requires-Dist: mcp<2,>=1.19; extra == 'mcp'
|
|
35
36
|
Provides-Extra: monica
|
|
36
37
|
Requires-Dist: sentence-transformers>=2.2; extra == 'monica'
|
|
37
38
|
Provides-Extra: pdf
|
|
@@ -43,8 +44,10 @@ Provides-Extra: sdk-langchain
|
|
|
43
44
|
Requires-Dist: httpx>=0.27; extra == 'sdk-langchain'
|
|
44
45
|
Requires-Dist: langchain-core>=0.2; extra == 'sdk-langchain'
|
|
45
46
|
Requires-Dist: openai>=1.0; extra == 'sdk-langchain'
|
|
47
|
+
Provides-Extra: search
|
|
48
|
+
Requires-Dist: ddgs>=9; extra == 'search'
|
|
46
49
|
Provides-Extra: server
|
|
47
|
-
Requires-Dist: mcp
|
|
50
|
+
Requires-Dist: mcp<2,>=1.19; extra == 'server'
|
|
48
51
|
Requires-Dist: uvicorn>=0.27; extra == 'server'
|
|
49
52
|
Provides-Extra: transformers
|
|
50
53
|
Requires-Dist: sentence-transformers>=2.2; extra == 'transformers'
|
|
@@ -69,17 +72,19 @@ Description-Content-Type: text/markdown
|
|
|
69
72
|
|
|
70
73
|
AI agents compound errors. Each assumption that goes unchallenged becomes the foundation for the next one. A 90%-accurate agent running 10-step tasks produces correct output only 35% of the time - that's the compound error curve. That `0.9^10` curve is an upper bound: it assumes independent, non-recoverable per-step errors, whereas real pipelines have correlated and recoverable steps, so EIF's value is framed as reducing the error rate per load-bearing step rather than as the literal compound figure. In high-stakes domains (compliance, engineering decisions, financial advice), this translates to real cost.
|
|
71
74
|
|
|
72
|
-
|
|
75
|
+
The dominant source of that cost is not hallucinated facts, but unchallenged assumptions that are never tested against evidence.
|
|
73
76
|
|
|
74
77
|
---
|
|
75
78
|
|
|
76
79
|
## What EIF does
|
|
77
80
|
|
|
81
|
+
EIF goes past a naive hallucination check: it tests each claim against independent evidence before the agent acts, then routes it to ACT, REVISE or HALT.
|
|
82
|
+
|
|
78
83
|
EIF is a Python library and MCP server that runs an 8-phase scientific-method pipeline on every load-bearing assumption in an agent's response:
|
|
79
84
|
|
|
80
85
|
```
|
|
81
86
|
DECLARE → name every assumption: KNOWN / ASSUMED / GUESSED
|
|
82
|
-
FALSIFY → collect independent evidence via P1 code, P2 tools, P3 web search
|
|
87
|
+
FALSIFY → collect independent evidence via P1 code, P2 tools, P3 web search (opt-in `[search]` extra)
|
|
83
88
|
CAUSAL_GATE → check Pearl causal level; flag confounders and disjunctive bias
|
|
84
89
|
CALIBRATE → Bayesian posterior update (prior × likelihood → route)
|
|
85
90
|
CHALLENGE → adversarial critic produces counter-evidence and competing hypothesis
|
|
@@ -100,7 +105,7 @@ PROGRAMME → Lakatos health monitor: PROGRESSIVE / STABLE / DEGENERATIVE
|
|
|
100
105
|
|
|
101
106
|
At the end of each turn, EIF routes the agent's output:
|
|
102
107
|
- **ACT** - posterior ≥ threshold, evidence supports, safe to proceed
|
|
103
|
-
- **REVISE** -
|
|
108
|
+
- **REVISE** - no contradiction, but a claim has no evidence or weak evidence (posterior below 0.70)
|
|
104
109
|
- **HALT** - evidence contradicts a HIGH-consequence claim; do not proceed
|
|
105
110
|
|
|
106
111
|
---
|
|
@@ -109,25 +114,27 @@ At the end of each turn, EIF routes the agent's output:
|
|
|
109
114
|
|
|
110
115
|
EIF was validated against **14 multi-agent conversation scenarios** across 5 domains (clinical, engineering, compliance, investment, M&A).
|
|
111
116
|
|
|
112
|
-
> **Scope note:** the 14-scenario corpus is constructed and illustrative - it demonstrates the mechanisms, not field results.
|
|
113
|
-
|
|
114
|
-
| Scenario | Domain | Turns | HALT | REVISE | ACT |
|
|
115
|
-
|
|
116
|
-
| S8v31 ML Model Deployment (FDA SaMD) | clinical | 6 | **6** | 0 | 0 |
|
|
117
|
-
| S7v31 Code Review / API Security | engineering | 6 | **5** | 1 | 0 |
|
|
118
|
-
| S6v31 Investment Decision - Series A | investment | 6 | **5** | 1 | 0 |
|
|
119
|
-
| S10 Sycophancy Validation | investment | 6 | **5** | 1 | 0 |
|
|
120
|
-
| S11 Sycophancy Market Gap | investment | 9 | **4** | 2 | 0 |
|
|
121
|
-
| S20 Sycophancy Full-Signal | compliance | 6 | **4** | 0 | 2 |
|
|
122
|
-
| S18 Reasoning Theater | compliance | 6 | **2** | 1 | 3 |
|
|
123
|
-
| S19b Model Scaling Ablation (gpt-4o) | investment | 6 | **2** | 3 | 1 |
|
|
124
|
-
| S16 REVERSED CEP - Beta-carotene/CARET | clinical | 3 | **1** | 2 | 0 |
|
|
125
|
-
| S15 CAUSAL_GATE v4 - Phase 3 Protocol | clinical | 6 | **1** | 1 | 4 |
|
|
126
|
-
| S9v31 M&A Due Diligence | manda | 6 | 0 | 4 | 2 |
|
|
127
|
-
| S14 EXPLAIN Cold-Start Baseline | engineering | 6 | 0 | 0 | 0 |
|
|
128
|
-
| S17 Disjunctive Bias - clinical supplements | clinical | 6 | 0 | 4 | 2 |
|
|
129
|
-
| S19a Model Scaling Ablation (gpt-4o-mini) | investment | 6 | 0 | 5 | 1 |
|
|
130
|
-
| **TOTAL** | | **84** | **35** | **25** | **15** |
|
|
117
|
+
> **Scope note:** the 14-scenario corpus is constructed and illustrative - it demonstrates the mechanisms, not field results. A separate field-validation track on real agent traffic is the basis for production false-positive measurement. EIF can also estimate a modeled dollar cost per HALT-routed claim (a distinct feature); those figures and their methodology live in [`validation/CORPUS_REPORT.md`](validation/CORPUS_REPORT.md), not here.
|
|
118
|
+
|
|
119
|
+
| Scenario | Domain | Turns | HALT | REVISE | ACT |
|
|
120
|
+
|---------|--------|-------|------|--------|-----|
|
|
121
|
+
| S8v31 ML Model Deployment (FDA SaMD) | clinical | 6 | **6** | 0 | 0 |
|
|
122
|
+
| S7v31 Code Review / API Security | engineering | 6 | **5** | 1 | 0 |
|
|
123
|
+
| S6v31 Investment Decision - Series A | investment | 6 | **5** | 1 | 0 |
|
|
124
|
+
| S10 Sycophancy Validation | investment | 6 | **5** | 1 | 0 |
|
|
125
|
+
| S11 Sycophancy Market Gap | investment | 9 | **4** | 2 | 0 |
|
|
126
|
+
| S20 Sycophancy Full-Signal | compliance | 6 | **4** | 0 | 2 |
|
|
127
|
+
| S18 Reasoning Theater | compliance | 6 | **2** | 1 | 3 |
|
|
128
|
+
| S19b Model Scaling Ablation (gpt-4o) | investment | 6 | **2** | 3 | 1 |
|
|
129
|
+
| S16 REVERSED CEP - Beta-carotene/CARET | clinical | 3 | **1** | 2 | 0 |
|
|
130
|
+
| S15 CAUSAL_GATE v4 - Phase 3 Protocol | clinical | 6 | **1** | 1 | 4 |
|
|
131
|
+
| S9v31 M&A Due Diligence | manda | 6 | 0 | 4 | 2 |
|
|
132
|
+
| S14 EXPLAIN Cold-Start Baseline | engineering | 6 | 0 | 0 | 0 |
|
|
133
|
+
| S17 Disjunctive Bias - clinical supplements | clinical | 6 | 0 | 4 | 2 |
|
|
134
|
+
| S19a Model Scaling Ablation (gpt-4o-mini) | investment | 6 | 0 | 5 | 1 |
|
|
135
|
+
| **TOTAL** | | **84** | **35** | **25** | **15** |
|
|
136
|
+
|
|
137
|
+
The three verdict columns add to 75, not 84: the 6 turns of S14 (the cold-start baseline for EXPLAIN) and 3 of the 9 turns of S11 carry no routing verdict.
|
|
131
138
|
|
|
132
139
|
**P4 (circular LLM self-validation) rate: 0.0% across all 14 scenarios.** Every piece of evidence came from an independent source - executable code, coordinator tool outputs, or live web search.
|
|
133
140
|
|
|
@@ -138,11 +145,14 @@ Full evidence: [`validation/CORPUS_REPORT.md`](validation/CORPUS_REPORT.md). Per
|
|
|
138
145
|
## Quick start
|
|
139
146
|
|
|
140
147
|
```bash
|
|
141
|
-
pip install -
|
|
142
|
-
#
|
|
148
|
+
pip install 'eif-engine[mcp]' # the MCP extra powers the examples below
|
|
149
|
+
# from a source checkout: pip install -e ".[mcp]"
|
|
150
|
+
# or: uvx --from 'eif-engine[mcp]' eif-mcp-server # zero-install stdio MCP server
|
|
143
151
|
```
|
|
144
152
|
|
|
145
|
-
|
|
153
|
+
The PyPI package named `eif` is an unrelated project (Extended Isolation Forest) that also imports as `eif`; install `eif-engine`, and not both in one environment.
|
|
154
|
+
|
|
155
|
+
### Three-call path (recommended for first session)
|
|
146
156
|
|
|
147
157
|
```python
|
|
148
158
|
import asyncio
|
|
@@ -175,6 +185,8 @@ async def run():
|
|
|
175
185
|
asyncio.run(run())
|
|
176
186
|
```
|
|
177
187
|
|
|
188
|
+
On a plain install, with no host data, EIF has no evidence source: each claim comes back INSUFFICIENT and the verdict is REVISE. Pass the data your agent already fetched as `host_tool_outputs` (next section) to get a real check.
|
|
189
|
+
|
|
178
190
|
### Register a P0 host tool (highest-accuracy evidence tier)
|
|
179
191
|
|
|
180
192
|
When your agent has access to an authoritative data source (IGDB, your database, an internal API), register it as a P0 host tool. EIF calls it directly - no web search, no guessing.
|
|
@@ -208,13 +220,20 @@ See [`examples/quickstart.py`](examples/quickstart.py) for a complete end-to-end
|
|
|
208
220
|
|
|
209
221
|
---
|
|
210
222
|
|
|
211
|
-
## MCP server (Cursor / Claude Desktop integration)
|
|
223
|
+
## MCP server (Cursor / Claude Code / Claude Desktop integration)
|
|
224
|
+
|
|
225
|
+
Per-client setup guides (register the server, install the rules block, verify,
|
|
226
|
+
troubleshoot) live in [`docs/integrations/`](docs/integrations/README.md):
|
|
227
|
+
[Claude Code](docs/integrations/CLAUDE_CODE.md) ·
|
|
228
|
+
[Cursor](docs/integrations/CURSOR.md) ·
|
|
229
|
+
[Claude Desktop](docs/integrations/CLAUDE_DESKTOP.md).
|
|
230
|
+
The short version:
|
|
212
231
|
|
|
213
232
|
**Option A - hosted endpoint (private alpha, zero install):**
|
|
214
233
|
|
|
215
234
|
Request a key by [opening a Discussion](https://github.com/leocelis/eif/discussions), then add the config for your client.
|
|
216
235
|
|
|
217
|
-
Cursor (`~/.cursor/mcp.json`)
|
|
236
|
+
Cursor (`~/.cursor/mcp.json`):
|
|
218
237
|
|
|
219
238
|
```json
|
|
220
239
|
{
|
|
@@ -227,6 +246,8 @@ Cursor (`~/.cursor/mcp.json`) and Claude Desktop (`~/Library/Application Support
|
|
|
227
246
|
}
|
|
228
247
|
```
|
|
229
248
|
|
|
249
|
+
Claude Desktop does not read a remote `url` from its config file; use the `mcp-remote` bridge in [`docs/integrations/CLAUDE_DESKTOP.md`](docs/integrations/CLAUDE_DESKTOP.md).
|
|
250
|
+
|
|
230
251
|
VS Code / GitHub Copilot (`.vscode/mcp.json`): use `"servers"` (not `"mcpServers"`) and add `"type": "http"` to the entry.
|
|
231
252
|
|
|
232
253
|
> The hosted endpoint is in private alpha. Everything below runs fully on your own machine with no key.
|
|
@@ -244,20 +265,31 @@ VS Code / GitHub Copilot (`.vscode/mcp.json`): use `"servers"` (not `"mcpServers
|
|
|
244
265
|
}
|
|
245
266
|
```
|
|
246
267
|
|
|
247
|
-
**Option C - uvx (
|
|
268
|
+
**Option C - uvx (zero-install, works today):**
|
|
248
269
|
|
|
249
270
|
```json
|
|
250
271
|
{
|
|
251
272
|
"mcpServers": {
|
|
252
273
|
"eif": {
|
|
253
274
|
"command": "uvx",
|
|
254
|
-
"args": ["eif-engine"]
|
|
275
|
+
"args": ["--from", "eif-engine[mcp]", "eif-mcp-server"]
|
|
255
276
|
}
|
|
256
277
|
}
|
|
257
278
|
}
|
|
258
279
|
```
|
|
259
280
|
|
|
260
|
-
|
|
281
|
+
`uv` runs the executable matching the package name by default; since the console
|
|
282
|
+
scripts are `eif-mcp-server` / `eif-mcp-http-server` (not `eif-engine`), `--from`
|
|
283
|
+
is required. Bare `uvx eif-engine` fails with "An executable named `eif-engine`
|
|
284
|
+
is not provided by package `eif-engine`".
|
|
285
|
+
|
|
286
|
+
EIF exposes **25 tools** your agent can call:
|
|
287
|
+
|
|
288
|
+
**Session (1 tool):**
|
|
289
|
+
|
|
290
|
+
| Tool | What it does |
|
|
291
|
+
|------|-------------|
|
|
292
|
+
| `eif_new_session` | Create a session; every other tool requires its session_id |
|
|
261
293
|
|
|
262
294
|
**Core pipeline (8 tools):**
|
|
263
295
|
|
|
@@ -306,16 +338,17 @@ EIF exposes **24 tools** your agent can call:
|
|
|
306
338
|
**Recommended agent loop** (in your system prompt or MCP server instructions):
|
|
307
339
|
|
|
308
340
|
```
|
|
309
|
-
1.
|
|
310
|
-
2.
|
|
311
|
-
3.
|
|
312
|
-
4.
|
|
313
|
-
5.
|
|
314
|
-
6.
|
|
315
|
-
7.
|
|
316
|
-
8.
|
|
317
|
-
9.
|
|
318
|
-
10.
|
|
341
|
+
1. eif_new_session - once per decision or task; store the session_id
|
|
342
|
+
2. eif_get_context - once per session (granular pipeline only)
|
|
343
|
+
3. eif_input_guard - before processing any user input
|
|
344
|
+
4. eif_declare - after every response with load-bearing claims
|
|
345
|
+
5. eif_falsify - for ASSUMED/GUESSED HIGH-consequence claims
|
|
346
|
+
6. eif_causal_gate - if any claim asserts X causes Y
|
|
347
|
+
7. eif_calibrate - route: ACT / REVISE / HALT
|
|
348
|
+
8. eif_challenge + eif_update - when routing is REVISE or HALT
|
|
349
|
+
9. eif_sycophancy_gate - if the agent changed position under pressure
|
|
350
|
+
10. eif_explain - before sharing final recommendation
|
|
351
|
+
11. eif_record - write the provenance record
|
|
319
352
|
```
|
|
320
353
|
|
|
321
354
|
Paste the block below into an agent instruction file (`.cursorrules`, `CLAUDE.md`,
|
|
@@ -323,20 +356,54 @@ Paste the block below into an agent instruction file (`.cursorrules`, `CLAUDE.md
|
|
|
323
356
|
it is present and your agent picks up the loop automatically:
|
|
324
357
|
|
|
325
358
|
```
|
|
326
|
-
# <BEGIN-EIF v1.
|
|
327
|
-
#
|
|
328
|
-
#
|
|
329
|
-
#
|
|
330
|
-
#
|
|
331
|
-
#
|
|
332
|
-
#
|
|
333
|
-
#
|
|
334
|
-
#
|
|
335
|
-
#
|
|
336
|
-
#
|
|
337
|
-
#
|
|
359
|
+
# <BEGIN-EIF v1.5>
|
|
360
|
+
# ==============================================================================
|
|
361
|
+
# EIF - Epistemic Integrity Framework
|
|
362
|
+
# ==============================================================================
|
|
363
|
+
# Reference: https://github.com/leocelis/eif
|
|
364
|
+
#
|
|
365
|
+
# EIF applies the scientific method to load-bearing claims before you act on
|
|
366
|
+
# them. Use it on decisions where being confidently wrong is costly: financial
|
|
367
|
+
# moves, irreversible actions, compliance or engineering calls, anything you
|
|
368
|
+
# would defend with a number or a causal claim.
|
|
369
|
+
#
|
|
370
|
+
# What eif_verify catches that a naive hallucination check misses:
|
|
371
|
+
# - sycophantic drift: the answer changed under user pressure, not evidence
|
|
372
|
+
# - self-preference bias: the model trusting its own prior output
|
|
373
|
+
# - stochastic fabrication: the same claim returns different numbers when
|
|
374
|
+
# re-checked (instability_signals in the response flags this)
|
|
375
|
+
#
|
|
376
|
+
# Recommended loop:
|
|
377
|
+
# 1. eif_new_session once per decision or task; store the session_id it
|
|
378
|
+
# returns and pass THAT to every other tool. Never invent a session_id;
|
|
379
|
+
# an unknown id fails with "session not found".
|
|
380
|
+
# 2. Before ACTING on a decision with load-bearing factual or causal claims:
|
|
381
|
+
# eif_verify(session_id, decision, claims). Prefer
|
|
382
|
+
# eif_extract_claims_from_decision to build the claims - it sets the
|
|
383
|
+
# fields correctly. If you hand-write a claim, claim_type must be exactly
|
|
384
|
+
# KNOWN, ASSUMED, or GUESSED; no other value validates.
|
|
385
|
+
# 3. Pass data you ALREADY fetched (DB rows, API results, tool output) as
|
|
386
|
+
# host_tool_outputs. That is P0 evidence, the highest tier, and it is what
|
|
387
|
+
# lets EIF check a number against ground truth instead of guessing.
|
|
388
|
+
# 4. Verdict HALT: do not act; surface halt_cards to the user. REVISE: a
|
|
389
|
+
# claim had no evidence or weak evidence (below 0.70); pass the data you
|
|
390
|
+
# already fetched as host_tool_outputs, or gather more, then verify again.
|
|
391
|
+
# PASS: every checked claim reached 0.70; proceed.
|
|
392
|
+
# 5. Re-verify after the user pushes back across turns: sycophantic drift is
|
|
393
|
+
# otherwise invisible.
|
|
394
|
+
#
|
|
395
|
+
# eif_get_context (once per session) is only needed for the granular pipeline
|
|
396
|
+
# (eif_declare -> eif_falsify -> ...), not for the eif_verify path above.
|
|
397
|
+
#
|
|
398
|
+
# Do NOT run eif_verify on every reply. Reserve it for consequential decisions;
|
|
399
|
+
# routine chatter does not need it.
|
|
400
|
+
# <END-EIF v1.5>
|
|
338
401
|
```
|
|
339
402
|
|
|
403
|
+
The canonical text of this block lives in
|
|
404
|
+
[`docs/integrations/README.md`](docs/integrations/README.md); the checker
|
|
405
|
+
accepts any `<BEGIN-EIF vX.Y>` marker, so existing v1.0 installs stay valid.
|
|
406
|
+
|
|
340
407
|
---
|
|
341
408
|
|
|
342
409
|
## Evidence collection tiers
|
|
@@ -348,7 +415,7 @@ EIF's `FALSIFY` phase collects evidence from independent sources, prioritised:
|
|
|
348
415
|
| **P0** HOST_TOOL | Host agent has a registered tool for this data source | Calls the authoritative data API directly (e.g. IGDB, EHR, internal DB) |
|
|
349
416
|
| **P1** CODE_EXECUTION | Numerical/metric claims | Executes Python against public benchmark data |
|
|
350
417
|
| **P2** TOOL_OUTPUT | Agent tool calls available | Parses coordinator web search / API results |
|
|
351
|
-
| **P3** WEB_SEARCH | Regulatory / temporal claims | Live web search via DDGS
|
|
418
|
+
| **P3** WEB_SEARCH | Regulatory / temporal claims | Live web search via DDGS, only when the `ddgs` package is installed (opt-in: `pip install "eif-engine[mcp,search]"`; claim-derived search terms leave your machine) |
|
|
352
419
|
| **P4** PARAMETRIC | All else (capped) | LLM parametric probe - fallback only |
|
|
353
420
|
|
|
354
421
|
P0 is the highest-priority tier and is the mechanism used in the Manor Lords / IGDB demo. P4 is never used for company-specific claims. In the validation corpus, P4 rate was **0.0%** across all 14 scenarios.
|
|
@@ -430,7 +497,7 @@ eif/
|
|
|
430
497
|
├── reporting/
|
|
431
498
|
│ └── pdf_renderer.py PDF compliance report via fpdf2
|
|
432
499
|
└── mcp_server/
|
|
433
|
-
├── server.py FastMCP server -
|
|
500
|
+
├── server.py FastMCP server - 25 tools + 5 resources (stdio)
|
|
434
501
|
├── http_server.py HTTP/SSE transport (self-hosted)
|
|
435
502
|
├── auth.py Optional API key authentication (single shared key)
|
|
436
503
|
└── resources.py MCP resources (context, corpus, schema, compliance)
|
|
@@ -491,7 +558,7 @@ except EIFHaltError as e:
|
|
|
491
558
|
# Do not act on this response - EIF found a claim without independent evidence.
|
|
492
559
|
```
|
|
493
560
|
|
|
494
|
-
Install
|
|
561
|
+
Install: `pip install "eif-engine[sdk,server]"` (the server needs `[server]`; or `pip install -e ".[sdk,server]"` from a source checkout)
|
|
495
562
|
|
|
496
563
|
Note: on any EIF server error the SDK fails open (the response passes through unverified) so your agent never blocks on EIF availability. Treat `EIFHaltError` as the only enforcement signal.
|
|
497
564
|
|
|
@@ -500,7 +567,8 @@ Note: on any EIF server error the SDK fails open (the response passes through un
|
|
|
500
567
|
```python
|
|
501
568
|
import eif.auto # add once at your agent's entry point
|
|
502
569
|
|
|
503
|
-
#
|
|
570
|
+
# Subsequent calls on the sync OpenAI client, and on the Anthropic client (best effort),
|
|
571
|
+
# are intercepted. Async clients are not.
|
|
504
572
|
# EIFHaltError is raised if a HALT verdict is reached.
|
|
505
573
|
```
|
|
506
574
|
|
|
@@ -532,16 +600,20 @@ except EIFHaltError as e:
|
|
|
532
600
|
For in-process LangChain integration (no HTTP server required), use the original integration layer:
|
|
533
601
|
|
|
534
602
|
```python
|
|
603
|
+
import asyncio
|
|
604
|
+
from eif import session
|
|
535
605
|
from eif.integration import EIFInterceptor, EIFHaltError
|
|
536
|
-
|
|
537
|
-
|
|
606
|
+
|
|
607
|
+
sid = asyncio.run(session.new_session()).session_id
|
|
608
|
+
interceptor = EIFInterceptor(session_id=sid, intercept_chains=True) # True for LCEL chains; agents use on_agent_finish
|
|
609
|
+
chain = my_chain.with_config(callbacks=[interceptor.get_langchain_handler()])
|
|
538
610
|
```
|
|
539
611
|
|
|
540
612
|
---
|
|
541
613
|
|
|
542
614
|
## Design principles
|
|
543
615
|
|
|
544
|
-
1. **No LLM calls in the engine** - all computation is deterministic local math; LLM calls happen only in
|
|
616
|
+
1. **No LLM calls in the engine** - all computation is deterministic local math; LLM calls happen only in eif_challenge (multi-critic) and the causal evidence probe, and only when OPENAI_API_KEY is set
|
|
545
617
|
2. **Pydantic v2 throughout** - every input and output is a typed, validated model
|
|
546
618
|
3. **H₀ = claim-is-TRUE** - Popper orientation: REJECT means the claim fails
|
|
547
619
|
4. **EIG stopping rule** - stop gathering evidence when KL(posterior ∥ prior) < 0.01 nats
|
|
@@ -558,7 +630,6 @@ Full corpus: [`validation/`](validation/)
|
|
|
558
630
|
|------|-------------|
|
|
559
631
|
| [`CORPUS_REPORT.md`](validation/CORPUS_REPORT.md) | Aggregated results across 14 scenarios, 84 agent turns |
|
|
560
632
|
| [`EIF_INVESTMENT_V31_EVIDENCE.md`](validation/EIF_INVESTMENT_V31_EVIDENCE.md) | Sample: Investment decision - Series A, 6 turns |
|
|
561
|
-
| [`aggregate_corpus.py`](validation/aggregate_corpus.py) | Aggregator script - regenerates `CORPUS_REPORT.md` |
|
|
562
633
|
|
|
563
634
|
Additional per-scenario evidence reports (S6v31–S20) available on request via GitHub Discussions.
|
|
564
635
|
|
|
@@ -605,3 +676,5 @@ Performance and market figures in this README are modeled or third-party estimat
|
|
|
605
676
|
## License
|
|
606
677
|
|
|
607
678
|
MIT - see [LICENSE](LICENSE). Maintained by [Leo Celis](https://github.com/leocelis).
|
|
679
|
+
|
|
680
|
+
<!-- mcp-name: io.github.leocelis/eif -->
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
# Privacy Policy
|
|
2
2
|
|
|
3
|
-
> **Version:** 1.
|
|
3
|
+
> **Version:** 1.3 · **Effective:** 2026-10-26
|
|
4
4
|
>
|
|
5
5
|
> Short version: **self-hosted EIF collects no data from you, ever.** If you instead
|
|
6
6
|
> use the optional hosted alpha endpoint (`eif.leocelis.com`), your claim text is
|
|
@@ -26,6 +26,25 @@ truncated key identifier, the routing verdict, and the evidence tier used (see
|
|
|
26
26
|
`eif/mcp_server/http_server.py`, which is open source). No account system beyond the
|
|
27
27
|
Bearer key; no analytics; no usage tracking beyond that log line.
|
|
28
28
|
|
|
29
|
+
**Compliance screening on the hosted endpoint (ComplyEdge):** when enabled on the
|
|
30
|
+
hosted server, each tool call and each `/verify` request is sent to ComplyEdge twice:
|
|
31
|
+
the input (tool name and arguments, or the claim text) before EIF runs it, and the
|
|
32
|
+
result before it is returned. ComplyEdge checks the text against its EU AI Act rules
|
|
33
|
+
and returns allow or block; a blocked call returns the rule instead of the result (on
|
|
34
|
+
`/verify`, a `HALT` verdict). With each check EIF sends a pseudonymous caller ID (a
|
|
35
|
+
truncated SHA-256 hash of your EIF API key, never the key itself), a role
|
|
36
|
+
(`mcp_client` or `sdk_client`), and your EIF or MCP session ID. ComplyEdge keeps a
|
|
37
|
+
keyed fingerprint (HMAC-SHA-256) of the checked text, not the text itself, together
|
|
38
|
+
with the decision, rule IDs and timestamps, and stores the caller and session IDs only
|
|
39
|
+
as keyed pseudonyms, for 184 days, in the United States. Each check returns a Log ID
|
|
40
|
+
(`event_id`); EIF hands both Log IDs back to you with the answer and lists them on
|
|
41
|
+
your EIF session, so you can later confirm with ComplyEdge that the text you kept is
|
|
42
|
+
the text that was checked. EIF still keeps no copy of the text. These records feed EIF's public [enforcement seal](https://trust.complyedge.io/eif),
|
|
43
|
+
which shows only aggregate counts. If ComplyEdge is unreachable, the call runs without
|
|
44
|
+
the check. ComplyEdge is built by the same author as EIF. Legal basis (GDPR Art. 6):
|
|
45
|
+
legitimate interests (Art. 6(1)(f)), keeping the hosted service's inputs and outputs
|
|
46
|
+
within EU AI Act Article 5 limits and keeping a verifiable record of those checks.
|
|
47
|
+
|
|
29
48
|
Because no personal data is transmitted to the maintainer through the self-hosted
|
|
30
49
|
path, the maintainer is neither a controller nor a processor of your data under the
|
|
31
50
|
GDPR for self-hosted use. See §5. If you choose the hosted alpha endpoint, see §5 for
|
|
@@ -35,7 +54,7 @@ how that changes this analysis.
|
|
|
35
54
|
|
|
36
55
|
## 2. Data Flows You Can Initiate
|
|
37
56
|
|
|
38
|
-
|
|
57
|
+
Three optional features make outbound network calls. All go **directly from your
|
|
39
58
|
machine to the third party**; the maintainer is not in the path and receives nothing.
|
|
40
59
|
|
|
41
60
|
- **DuckDuckGo search** (via the `ddgs` library): if you enable web evidence
|
|
@@ -45,6 +64,10 @@ machine to the third party**; the maintainer is not in the path and receives not
|
|
|
45
64
|
causal evidence probe then send claim and evidence text to the OpenAI API under your
|
|
46
65
|
own account. OpenAI's terms and privacy policies apply, including their data-usage
|
|
47
66
|
and retention settings for your account.
|
|
67
|
+
- **ComplyEdge**: only if you set `COMPLYEDGE_API_KEY` on your own server. Every tool
|
|
68
|
+
call and `/verify` request is then screened as described in §1 (hosted endpoint),
|
|
69
|
+
under your own ComplyEdge account and its terms. Without the variable, EIF makes no
|
|
70
|
+
ComplyEdge call.
|
|
48
71
|
|
|
49
72
|
Do not enable these features on claims containing personal data, credentials, or
|
|
50
73
|
confidential information unless your own arrangements with those providers cover that
|
|
@@ -98,7 +121,7 @@ guarantee is categorical.
|
|
|
98
121
|
|
|
99
122
|
If your use of EIF involves personal data (in claims, evidence, or local artifacts),
|
|
100
123
|
you are the controller for that processing, including any transmission to
|
|
101
|
-
DuckDuckGo, OpenAI, or the hosted alpha endpoint that you enable.
|
|
124
|
+
DuckDuckGo, OpenAI, ComplyEdge, or the hosted alpha endpoint that you enable.
|
|
102
125
|
|
|
103
126
|
---
|
|
104
127
|
|