eif-engine 4.2.0__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-4.2.0 → eif_engine-4.3.0}/CHANGELOG.md +32 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/PKG-INFO +35 -16
- {eif_engine-4.2.0 → eif_engine-4.3.0}/README.md +31 -15
- {eif_engine-4.2.0 → eif_engine-4.3.0}/SECURITY.md +32 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/deploy/do_app_spec.yaml +5 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/architecture.md +2 -0
- eif_engine-4.3.0/docs/eif_host_tenant_keys_tech_spec.md +138 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/integrations/CLAUDE_CODE.md +8 -6
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/integrations/CLAUDE_DESKTOP.md +19 -7
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/integrations/CURSOR.md +7 -5
- eif_engine-4.3.0/docs/integrations/HOST_APP.md +118 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/integrations/README.md +7 -5
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/__init__.py +1 -1
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/native_tools.py +1 -1
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/input_guard/eif_input_guard_intent.yaml +2 -2
- {eif_engine-4.2.0 → 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-4.2.0 → eif_engine-4.3.0}/eif/mcp_server/http_server.py +32 -4
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/mcp_server/server.py +112 -16
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/schemas.py +1 -1
- eif_engine-4.3.0/eif/session.py +116 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/pyproject.toml +5 -1
- {eif_engine-4.2.0 → eif_engine-4.3.0}/server.json +1 -1
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/integration/test_mcp_server.py +1 -1
- eif_engine-4.3.0/tests/unit/test_integration_interceptor_langchain.py +33 -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-4.2.0/eif/mcp_server/auth.py +0 -81
- eif_engine-4.2.0/eif/session.py +0 -61
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.dockerignore +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.env.example +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.github/CODEOWNERS +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.github/ISSUE_TEMPLATE/bug_report.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.github/ISSUE_TEMPLATE/config.yml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.github/dependabot.yml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.github/pull_request_template.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.github/workflows/ci.yml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.gitignore +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.pre-commit-config.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/.trustlint.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/CODE_OF_CONDUCT.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/CONTRIBUTING.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/DATA_PROCESSING_AGREEMENT.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/DECISIONS.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/Dockerfile +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/LEGAL.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/LICENSE +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/NOTICE +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/PRIVACY_POLICY.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/TERMS_OF_SERVICE.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/dev_factory/standard_ei_01_nist_ai_risk_management_framework/bind.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/dev_factory/standard_ei_01_nist_ai_risk_management_framework/test_bind.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/compliance.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/eif_causal_gate_v4_tech_spec.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/eif_s17_s18_s19_tech_spec.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/eif_v2_tech_spec.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/integrations/COMPLYEDGE.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/docs/research_foundation.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/auto.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/calibrate/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/calibrate/bayesian.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/calibrate/conformal.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/calibrate/domain_constraints.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/calibrate/ece.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/calibrate/prior_strategy.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/calibrate/trust.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/causal_gate/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/causal_gate/confound.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/causal_gate/direction.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/causal_gate/eif_causal_gate_v4_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/causal_gate/evidence_probe.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/causal_gate/intervention.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/causal_gate/iris_pipeline.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/causal_gate/verdict.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/challenge/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/challenge/diversity.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/challenge/eif_multi_critic_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/challenge/multi_critic.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/challenge/protocol.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/challenge/replicator.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/cost_model/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/cost_model/catch_rate_empirical.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/cost_model/cost_ledger.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/cost_model/eif_catchrate_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/declare/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/declare/extractor.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/declare/harking_guard.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/declare/registry.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/eif_engine_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/eif_v2_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/eif_v3_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/explain/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/explain/artifact.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/explain/eif_explain_coldstart_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/explain/hard_to_vary.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/explain/reach.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/condition.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/eif_evidence_collection_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/eif_native_tools_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/eif_v5_evidence_trust_weighting_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/evidence_collector.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/hard_to_vary.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/sprt.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/falsify/trivial_check.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/agenda.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/eif_hypothesis_agenda_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/fdr.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/hypothesis_agenda/scorer.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/input_guard/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/input_guard/detector.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/integration/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/integration/eif_auto.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/mcp_server/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/mcp_server/compliance.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/mcp_server/resources.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/mcp_server/runtime_compliance_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/memory/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/memory/strategy.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/programme/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/programme/monitor.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/programme/principle_memory.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/programme/principle_revision.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/programme/signals.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/programme/status.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/py.typed +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/record/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/record/chain.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/record/compliance.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/record/isc_disclosure.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/record/outcome_store.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/record/provenance.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/record/research_object.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/replicate/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/replicate/divergence.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/replicate/eif_independent_replication_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/replicate/independent.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/replicate/protocol.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/reporting/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/reporting/pdf_renderer.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sdk/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sdk/exceptions.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sdk/interceptor.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sycophancy/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sycophancy/cost_model.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sycophancy/detector.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sycophancy/drift.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sycophancy/eif_sycophancy_intent.yaml +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sycophancy/faithfulness.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/sycophancy/framing.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/update/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/update/eig.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/update/paradigm.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/update/posterior.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/eif/update/stopping.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/examples/quickstart.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/index.html +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/README.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/ai_agents/calibration.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/ai_agents/causal_reasoning.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/ai_agents/hypothesis_generation.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/ai_agents/multi_agent_review.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/ai_agents/overview.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/ai_agents/provenance.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/foundations/ai_and_science.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/foundations/classical_method.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/foundations/deutsch_epistemology.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/foundations/frontier_debates.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/foundations/philosophy_of_science.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/foundations/replication_crisis.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/foundations/statistics.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/research/sycophancy_bias_research.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/scripts/compliance/check.sh +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/scripts/compliance/runtime_check.sh +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/conftest.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/integration/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/integration/test_eif_loop.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/__init__.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_calibrate.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_catch_rate_empirical.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_causal_gate.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_complyedge_cold_start.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_cost_ledger.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_dases_adaptive.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_dases_replicator.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_declare.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_deeptrace_live.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_ece_v2.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_evidence_trust_weighting.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_explain.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_exploratory_confirmatory.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_falsify.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_fdr.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_framing_elephant.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_hard_to_vary.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_host_tool_numeric.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_http_rate_limit.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_independent_replication.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_iris_pipeline.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_isc_disclosure.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_memory.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_monica_drift.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_native_tools_nt3.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_outcome_store.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_p1_sandbox_security.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_paradigm.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_principle_memory.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_principle_revision.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_programme.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_record.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_replicate.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_research_object.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_runtime_check_attribution.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_runtime_compliance.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_schemas.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_sdk_interceptor.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/tests/unit/test_update.py +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/validation/CORPUS_REPORT.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/validation/EIF_INVESTMENT_V31_EVIDENCE.md +0 -0
- {eif_engine-4.2.0 → eif_engine-4.3.0}/validation/data/.gitkeep +0 -0
|
@@ -7,6 +7,38 @@ 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
|
+
|
|
10
42
|
## [4.2.0] - 2026-11-05
|
|
11
43
|
|
|
12
44
|
### Fixed
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.5
|
|
2
2
|
Name: eif-engine
|
|
3
|
-
Version: 4.
|
|
3
|
+
Version: 4.3.0
|
|
4
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
|
|
@@ -21,6 +21,7 @@ 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'
|
|
@@ -43,6 +44,8 @@ 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
50
|
Requires-Dist: mcp<2,>=1.19; extra == 'server'
|
|
48
51
|
Requires-Dist: uvicorn>=0.27; extra == 'server'
|
|
@@ -81,7 +84,7 @@ EIF is a Python library and MCP server that runs an 8-phase scientific-method pi
|
|
|
81
84
|
|
|
82
85
|
```
|
|
83
86
|
DECLARE → name every assumption: KNOWN / ASSUMED / GUESSED
|
|
84
|
-
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)
|
|
85
88
|
CAUSAL_GATE → check Pearl causal level; flag confounders and disjunctive bias
|
|
86
89
|
CALIBRATE → Bayesian posterior update (prior × likelihood → route)
|
|
87
90
|
CHALLENGE → adversarial critic produces counter-evidence and competing hypothesis
|
|
@@ -102,7 +105,7 @@ PROGRAMME → Lakatos health monitor: PROGRESSIVE / STABLE / DEGENERATIVE
|
|
|
102
105
|
|
|
103
106
|
At the end of each turn, EIF routes the agent's output:
|
|
104
107
|
- **ACT** - posterior ≥ threshold, evidence supports, safe to proceed
|
|
105
|
-
- **REVISE** -
|
|
108
|
+
- **REVISE** - no contradiction, but a claim has no evidence or weak evidence (posterior below 0.70)
|
|
106
109
|
- **HALT** - evidence contradicts a HIGH-consequence claim; do not proceed
|
|
107
110
|
|
|
108
111
|
---
|
|
@@ -131,6 +134,8 @@ EIF was validated against **14 multi-agent conversation scenarios** across 5 dom
|
|
|
131
134
|
| S19a Model Scaling Ablation (gpt-4o-mini) | investment | 6 | 0 | 5 | 1 |
|
|
132
135
|
| **TOTAL** | | **84** | **35** | **25** | **15** |
|
|
133
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.
|
|
138
|
+
|
|
134
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.
|
|
135
140
|
|
|
136
141
|
Full evidence: [`validation/CORPUS_REPORT.md`](validation/CORPUS_REPORT.md). Per-scenario reports available on request via GitHub Discussions.
|
|
@@ -140,11 +145,14 @@ Full evidence: [`validation/CORPUS_REPORT.md`](validation/CORPUS_REPORT.md). Per
|
|
|
140
145
|
## Quick start
|
|
141
146
|
|
|
142
147
|
```bash
|
|
143
|
-
pip install -
|
|
148
|
+
pip install 'eif-engine[mcp]' # the MCP extra powers the examples below
|
|
149
|
+
# from a source checkout: pip install -e ".[mcp]"
|
|
144
150
|
# or: uvx --from 'eif-engine[mcp]' eif-mcp-server # zero-install stdio MCP server
|
|
145
151
|
```
|
|
146
152
|
|
|
147
|
-
|
|
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)
|
|
148
156
|
|
|
149
157
|
```python
|
|
150
158
|
import asyncio
|
|
@@ -177,6 +185,8 @@ async def run():
|
|
|
177
185
|
asyncio.run(run())
|
|
178
186
|
```
|
|
179
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
|
+
|
|
180
190
|
### Register a P0 host tool (highest-accuracy evidence tier)
|
|
181
191
|
|
|
182
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.
|
|
@@ -223,7 +233,7 @@ The short version:
|
|
|
223
233
|
|
|
224
234
|
Request a key by [opening a Discussion](https://github.com/leocelis/eif/discussions), then add the config for your client.
|
|
225
235
|
|
|
226
|
-
Cursor (`~/.cursor/mcp.json`)
|
|
236
|
+
Cursor (`~/.cursor/mcp.json`):
|
|
227
237
|
|
|
228
238
|
```json
|
|
229
239
|
{
|
|
@@ -236,6 +246,8 @@ Cursor (`~/.cursor/mcp.json`) and Claude Desktop (`~/Library/Application Support
|
|
|
236
246
|
}
|
|
237
247
|
```
|
|
238
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
|
+
|
|
239
251
|
VS Code / GitHub Copilot (`.vscode/mcp.json`): use `"servers"` (not `"mcpServers"`) and add `"type": "http"` to the entry.
|
|
240
252
|
|
|
241
253
|
> The hosted endpoint is in private alpha. Everything below runs fully on your own machine with no key.
|
|
@@ -344,7 +356,7 @@ Paste the block below into an agent instruction file (`.cursorrules`, `CLAUDE.md
|
|
|
344
356
|
it is present and your agent picks up the loop automatically:
|
|
345
357
|
|
|
346
358
|
```
|
|
347
|
-
# <BEGIN-EIF v1.
|
|
359
|
+
# <BEGIN-EIF v1.5>
|
|
348
360
|
# ==============================================================================
|
|
349
361
|
# EIF - Epistemic Integrity Framework
|
|
350
362
|
# ==============================================================================
|
|
@@ -373,8 +385,10 @@ it is present and your agent picks up the loop automatically:
|
|
|
373
385
|
# 3. Pass data you ALREADY fetched (DB rows, API results, tool output) as
|
|
374
386
|
# host_tool_outputs. That is P0 evidence, the highest tier, and it is what
|
|
375
387
|
# lets EIF check a number against ground truth instead of guessing.
|
|
376
|
-
# 4. Verdict HALT: do not act; surface halt_cards to the user. REVISE:
|
|
377
|
-
#
|
|
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.
|
|
378
392
|
# 5. Re-verify after the user pushes back across turns: sycophantic drift is
|
|
379
393
|
# otherwise invisible.
|
|
380
394
|
#
|
|
@@ -383,7 +397,7 @@ it is present and your agent picks up the loop automatically:
|
|
|
383
397
|
#
|
|
384
398
|
# Do NOT run eif_verify on every reply. Reserve it for consequential decisions;
|
|
385
399
|
# routine chatter does not need it.
|
|
386
|
-
# <END-EIF v1.
|
|
400
|
+
# <END-EIF v1.5>
|
|
387
401
|
```
|
|
388
402
|
|
|
389
403
|
The canonical text of this block lives in
|
|
@@ -401,7 +415,7 @@ EIF's `FALSIFY` phase collects evidence from independent sources, prioritised:
|
|
|
401
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) |
|
|
402
416
|
| **P1** CODE_EXECUTION | Numerical/metric claims | Executes Python against public benchmark data |
|
|
403
417
|
| **P2** TOOL_OUTPUT | Agent tool calls available | Parses coordinator web search / API results |
|
|
404
|
-
| **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) |
|
|
405
419
|
| **P4** PARAMETRIC | All else (capped) | LLM parametric probe - fallback only |
|
|
406
420
|
|
|
407
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.
|
|
@@ -544,7 +558,7 @@ except EIFHaltError as e:
|
|
|
544
558
|
# Do not act on this response - EIF found a claim without independent evidence.
|
|
545
559
|
```
|
|
546
560
|
|
|
547
|
-
Install: `pip install "eif-engine[sdk]"` (or `pip install -e ".[sdk]"` from a source checkout)
|
|
561
|
+
Install: `pip install "eif-engine[sdk,server]"` (the server needs `[server]`; or `pip install -e ".[sdk,server]"` from a source checkout)
|
|
548
562
|
|
|
549
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.
|
|
550
564
|
|
|
@@ -553,7 +567,8 @@ Note: on any EIF server error the SDK fails open (the response passes through un
|
|
|
553
567
|
```python
|
|
554
568
|
import eif.auto # add once at your agent's entry point
|
|
555
569
|
|
|
556
|
-
#
|
|
570
|
+
# Subsequent calls on the sync OpenAI client, and on the Anthropic client (best effort),
|
|
571
|
+
# are intercepted. Async clients are not.
|
|
557
572
|
# EIFHaltError is raised if a HALT verdict is reached.
|
|
558
573
|
```
|
|
559
574
|
|
|
@@ -585,16 +600,20 @@ except EIFHaltError as e:
|
|
|
585
600
|
For in-process LangChain integration (no HTTP server required), use the original integration layer:
|
|
586
601
|
|
|
587
602
|
```python
|
|
603
|
+
import asyncio
|
|
604
|
+
from eif import session
|
|
588
605
|
from eif.integration import EIFInterceptor, EIFHaltError
|
|
589
|
-
|
|
590
|
-
|
|
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()])
|
|
591
610
|
```
|
|
592
611
|
|
|
593
612
|
---
|
|
594
613
|
|
|
595
614
|
## Design principles
|
|
596
615
|
|
|
597
|
-
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
|
|
598
617
|
2. **Pydantic v2 throughout** - every input and output is a typed, validated model
|
|
599
618
|
3. **H₀ = claim-is-TRUE** - Popper orientation: REJECT means the claim fails
|
|
600
619
|
4. **EIG stopping rule** - stop gathering evidence when KL(posterior ∥ prior) < 0.01 nats
|
|
@@ -29,7 +29,7 @@ EIF is a Python library and MCP server that runs an 8-phase scientific-method pi
|
|
|
29
29
|
|
|
30
30
|
```
|
|
31
31
|
DECLARE → name every assumption: KNOWN / ASSUMED / GUESSED
|
|
32
|
-
FALSIFY → collect independent evidence via P1 code, P2 tools, P3 web search
|
|
32
|
+
FALSIFY → collect independent evidence via P1 code, P2 tools, P3 web search (opt-in `[search]` extra)
|
|
33
33
|
CAUSAL_GATE → check Pearl causal level; flag confounders and disjunctive bias
|
|
34
34
|
CALIBRATE → Bayesian posterior update (prior × likelihood → route)
|
|
35
35
|
CHALLENGE → adversarial critic produces counter-evidence and competing hypothesis
|
|
@@ -50,7 +50,7 @@ PROGRAMME → Lakatos health monitor: PROGRESSIVE / STABLE / DEGENERATIVE
|
|
|
50
50
|
|
|
51
51
|
At the end of each turn, EIF routes the agent's output:
|
|
52
52
|
- **ACT** - posterior ≥ threshold, evidence supports, safe to proceed
|
|
53
|
-
- **REVISE** -
|
|
53
|
+
- **REVISE** - no contradiction, but a claim has no evidence or weak evidence (posterior below 0.70)
|
|
54
54
|
- **HALT** - evidence contradicts a HIGH-consequence claim; do not proceed
|
|
55
55
|
|
|
56
56
|
---
|
|
@@ -79,6 +79,8 @@ EIF was validated against **14 multi-agent conversation scenarios** across 5 dom
|
|
|
79
79
|
| S19a Model Scaling Ablation (gpt-4o-mini) | investment | 6 | 0 | 5 | 1 |
|
|
80
80
|
| **TOTAL** | | **84** | **35** | **25** | **15** |
|
|
81
81
|
|
|
82
|
+
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.
|
|
83
|
+
|
|
82
84
|
**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.
|
|
83
85
|
|
|
84
86
|
Full evidence: [`validation/CORPUS_REPORT.md`](validation/CORPUS_REPORT.md). Per-scenario reports available on request via GitHub Discussions.
|
|
@@ -88,11 +90,14 @@ Full evidence: [`validation/CORPUS_REPORT.md`](validation/CORPUS_REPORT.md). Per
|
|
|
88
90
|
## Quick start
|
|
89
91
|
|
|
90
92
|
```bash
|
|
91
|
-
pip install -
|
|
93
|
+
pip install 'eif-engine[mcp]' # the MCP extra powers the examples below
|
|
94
|
+
# from a source checkout: pip install -e ".[mcp]"
|
|
92
95
|
# or: uvx --from 'eif-engine[mcp]' eif-mcp-server # zero-install stdio MCP server
|
|
93
96
|
```
|
|
94
97
|
|
|
95
|
-
|
|
98
|
+
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.
|
|
99
|
+
|
|
100
|
+
### Three-call path (recommended for first session)
|
|
96
101
|
|
|
97
102
|
```python
|
|
98
103
|
import asyncio
|
|
@@ -125,6 +130,8 @@ async def run():
|
|
|
125
130
|
asyncio.run(run())
|
|
126
131
|
```
|
|
127
132
|
|
|
133
|
+
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.
|
|
134
|
+
|
|
128
135
|
### Register a P0 host tool (highest-accuracy evidence tier)
|
|
129
136
|
|
|
130
137
|
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.
|
|
@@ -171,7 +178,7 @@ The short version:
|
|
|
171
178
|
|
|
172
179
|
Request a key by [opening a Discussion](https://github.com/leocelis/eif/discussions), then add the config for your client.
|
|
173
180
|
|
|
174
|
-
Cursor (`~/.cursor/mcp.json`)
|
|
181
|
+
Cursor (`~/.cursor/mcp.json`):
|
|
175
182
|
|
|
176
183
|
```json
|
|
177
184
|
{
|
|
@@ -184,6 +191,8 @@ Cursor (`~/.cursor/mcp.json`) and Claude Desktop (`~/Library/Application Support
|
|
|
184
191
|
}
|
|
185
192
|
```
|
|
186
193
|
|
|
194
|
+
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).
|
|
195
|
+
|
|
187
196
|
VS Code / GitHub Copilot (`.vscode/mcp.json`): use `"servers"` (not `"mcpServers"`) and add `"type": "http"` to the entry.
|
|
188
197
|
|
|
189
198
|
> The hosted endpoint is in private alpha. Everything below runs fully on your own machine with no key.
|
|
@@ -292,7 +301,7 @@ Paste the block below into an agent instruction file (`.cursorrules`, `CLAUDE.md
|
|
|
292
301
|
it is present and your agent picks up the loop automatically:
|
|
293
302
|
|
|
294
303
|
```
|
|
295
|
-
# <BEGIN-EIF v1.
|
|
304
|
+
# <BEGIN-EIF v1.5>
|
|
296
305
|
# ==============================================================================
|
|
297
306
|
# EIF - Epistemic Integrity Framework
|
|
298
307
|
# ==============================================================================
|
|
@@ -321,8 +330,10 @@ it is present and your agent picks up the loop automatically:
|
|
|
321
330
|
# 3. Pass data you ALREADY fetched (DB rows, API results, tool output) as
|
|
322
331
|
# host_tool_outputs. That is P0 evidence, the highest tier, and it is what
|
|
323
332
|
# lets EIF check a number against ground truth instead of guessing.
|
|
324
|
-
# 4. Verdict HALT: do not act; surface halt_cards to the user. REVISE:
|
|
325
|
-
#
|
|
333
|
+
# 4. Verdict HALT: do not act; surface halt_cards to the user. REVISE: a
|
|
334
|
+
# claim had no evidence or weak evidence (below 0.70); pass the data you
|
|
335
|
+
# already fetched as host_tool_outputs, or gather more, then verify again.
|
|
336
|
+
# PASS: every checked claim reached 0.70; proceed.
|
|
326
337
|
# 5. Re-verify after the user pushes back across turns: sycophantic drift is
|
|
327
338
|
# otherwise invisible.
|
|
328
339
|
#
|
|
@@ -331,7 +342,7 @@ it is present and your agent picks up the loop automatically:
|
|
|
331
342
|
#
|
|
332
343
|
# Do NOT run eif_verify on every reply. Reserve it for consequential decisions;
|
|
333
344
|
# routine chatter does not need it.
|
|
334
|
-
# <END-EIF v1.
|
|
345
|
+
# <END-EIF v1.5>
|
|
335
346
|
```
|
|
336
347
|
|
|
337
348
|
The canonical text of this block lives in
|
|
@@ -349,7 +360,7 @@ EIF's `FALSIFY` phase collects evidence from independent sources, prioritised:
|
|
|
349
360
|
| **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) |
|
|
350
361
|
| **P1** CODE_EXECUTION | Numerical/metric claims | Executes Python against public benchmark data |
|
|
351
362
|
| **P2** TOOL_OUTPUT | Agent tool calls available | Parses coordinator web search / API results |
|
|
352
|
-
| **P3** WEB_SEARCH | Regulatory / temporal claims | Live web search via DDGS
|
|
363
|
+
| **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) |
|
|
353
364
|
| **P4** PARAMETRIC | All else (capped) | LLM parametric probe - fallback only |
|
|
354
365
|
|
|
355
366
|
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.
|
|
@@ -492,7 +503,7 @@ except EIFHaltError as e:
|
|
|
492
503
|
# Do not act on this response - EIF found a claim without independent evidence.
|
|
493
504
|
```
|
|
494
505
|
|
|
495
|
-
Install: `pip install "eif-engine[sdk]"` (or `pip install -e ".[sdk]"` from a source checkout)
|
|
506
|
+
Install: `pip install "eif-engine[sdk,server]"` (the server needs `[server]`; or `pip install -e ".[sdk,server]"` from a source checkout)
|
|
496
507
|
|
|
497
508
|
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.
|
|
498
509
|
|
|
@@ -501,7 +512,8 @@ Note: on any EIF server error the SDK fails open (the response passes through un
|
|
|
501
512
|
```python
|
|
502
513
|
import eif.auto # add once at your agent's entry point
|
|
503
514
|
|
|
504
|
-
#
|
|
515
|
+
# Subsequent calls on the sync OpenAI client, and on the Anthropic client (best effort),
|
|
516
|
+
# are intercepted. Async clients are not.
|
|
505
517
|
# EIFHaltError is raised if a HALT verdict is reached.
|
|
506
518
|
```
|
|
507
519
|
|
|
@@ -533,16 +545,20 @@ except EIFHaltError as e:
|
|
|
533
545
|
For in-process LangChain integration (no HTTP server required), use the original integration layer:
|
|
534
546
|
|
|
535
547
|
```python
|
|
548
|
+
import asyncio
|
|
549
|
+
from eif import session
|
|
536
550
|
from eif.integration import EIFInterceptor, EIFHaltError
|
|
537
|
-
|
|
538
|
-
|
|
551
|
+
|
|
552
|
+
sid = asyncio.run(session.new_session()).session_id
|
|
553
|
+
interceptor = EIFInterceptor(session_id=sid, intercept_chains=True) # True for LCEL chains; agents use on_agent_finish
|
|
554
|
+
chain = my_chain.with_config(callbacks=[interceptor.get_langchain_handler()])
|
|
539
555
|
```
|
|
540
556
|
|
|
541
557
|
---
|
|
542
558
|
|
|
543
559
|
## Design principles
|
|
544
560
|
|
|
545
|
-
1. **No LLM calls in the engine** - all computation is deterministic local math; LLM calls happen only in
|
|
561
|
+
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
|
|
546
562
|
2. **Pydantic v2 throughout** - every input and output is a typed, validated model
|
|
547
563
|
3. **H₀ = claim-is-TRUE** - Popper orientation: REJECT means the claim fails
|
|
548
564
|
4. **EIG stopping rule** - stop gathering evidence when KL(posterior ∥ prior) < 0.01 nats
|
|
@@ -69,3 +69,35 @@ EIF is self-hosted software. You own the security of your deployment.
|
|
|
69
69
|
- **The P1 sandbox executes generated Python in a subprocess** after AST-based
|
|
70
70
|
validation. It is designed for EIF's own stdlib-only code templates. Do not
|
|
71
71
|
route untrusted arbitrary code through it.
|
|
72
|
+
|
|
73
|
+
---
|
|
74
|
+
|
|
75
|
+
## Tenant Keys (Host Apps)
|
|
76
|
+
|
|
77
|
+
A host app that serves many of its own customers (tenants) through one EIF
|
|
78
|
+
HTTP server can give each tenant its own key, derived from one secret shared
|
|
79
|
+
between the host and the server. See `docs/integrations/HOST_APP.md`.
|
|
80
|
+
|
|
81
|
+
- **Derived, never stored.** `key = "eif_" + tenant + "_" +
|
|
82
|
+
base64url_nopad(HMAC-SHA256(EIF_TENANT_KEY_SECRET, "eif:" + VERSION + ":" + tenant))`.
|
|
83
|
+
The server recomputes the MAC and compares it in constant time. Keep
|
|
84
|
+
`EIF_TENANT_KEY_SECRET` out of source control like every other key.
|
|
85
|
+
- **Rotation and revocation.** `EIF_TENANT_KEY_VERSIONS` (default `v1`) lists
|
|
86
|
+
the versions accepted; add a new version, move the host to it, then drop the
|
|
87
|
+
old one. `EIF_REVOKED_TENANTS` refuses one tenant's keys at once.
|
|
88
|
+
- **Tenant keys only.** A server with the secret set and no `EIF_API_KEY`
|
|
89
|
+
accepts tenant keys and nothing else, in development and in production; the
|
|
90
|
+
open development mode is off.
|
|
91
|
+
- **Sessions have owners.** A session belongs to the tenant whose request
|
|
92
|
+
created it. Another tenant asking for it, by tool or by resource, gets the
|
|
93
|
+
same "not found or expired" error as for an id that does not exist. A session
|
|
94
|
+
created with `EIF_API_KEY` is not readable with a tenant key, and the reverse.
|
|
95
|
+
- **One MCP session, one identity.** An MCP session is bound to the key that
|
|
96
|
+
opened it; a message on it with a key for another tenant (or the operator
|
|
97
|
+
key) is refused. The tenant is read from each message's own request.
|
|
98
|
+
- **Three tools.** A tenant key may call `eif_new_session`,
|
|
99
|
+
`eif_extract_claims_from_decision` and `eif_verify` only; `tools/list` shows
|
|
100
|
+
only those. Tools that can call a paid model on the operator's key (such as
|
|
101
|
+
`eif_challenge`) are out of tenant reach.
|
|
102
|
+
- **What it is not.** Sessions stay in memory; a restart loses them for every
|
|
103
|
+
caller. Rate limits are per key and per process.
|
|
@@ -67,6 +67,11 @@ services:
|
|
|
67
67
|
- key: COMPLYEDGE_API_KEY
|
|
68
68
|
type: SECRET
|
|
69
69
|
scope: RUN_TIME
|
|
70
|
+
# Optional: derived per-tenant keys for a host app (docs/integrations/HOST_APP.md).
|
|
71
|
+
# Unset = only EIF_API_KEY authenticates.
|
|
72
|
+
- key: EIF_TENANT_KEY_SECRET
|
|
73
|
+
type: SECRET
|
|
74
|
+
scope: RUN_TIME
|
|
70
75
|
|
|
71
76
|
# Custom domain: add in the DO dashboard (Settings -> Domains), e.g.
|
|
72
77
|
# eif.<your-domain>, then CNAME it to the app's default hostname.
|
|
@@ -209,6 +209,8 @@ eif/
|
|
|
209
209
|
- Docker image built from this repository.
|
|
210
210
|
- Optional hosted alpha: the maintainer runs the same HTTP/SSE server at `eif.leocelis.com` (Bearer key required, private alpha). See the README's MCP server section.
|
|
211
211
|
|
|
212
|
+
**Tenant keys.** For a host app serving many tenants through one HTTP server, `EIF_TENANT_KEY_SECRET` turns on keys derived per tenant (`auth.py`). Each MCP message is resolved to its own request's tenant, the MCP session is bound to the key that opened it, a session belongs to the tenant that created it (another tenant gets "not found or expired", by tool or by resource), and a tenant key sees only `eif_new_session`, `eif_extract_claims_from_decision` and `eif_verify`. `EIF_API_KEY` is unchanged. See `docs/integrations/HOST_APP.md`.
|
|
213
|
+
|
|
212
214
|
---
|
|
213
215
|
|
|
214
216
|
## Design principles
|
|
@@ -0,0 +1,138 @@
|
|
|
1
|
+
# EIF: Host-Derived Tenant Keys and Session Owners: Tech Spec
|
|
2
|
+
|
|
3
|
+
**Status:** BUILT (2026-11-16). Proposed 2026-11-15.
|
|
4
|
+
**Scope:** the self-hosted HTTP server (`eif/mcp_server/`) and the session store (`eif/session.py`).
|
|
5
|
+
Stdio and local use do not change.
|
|
6
|
+
|
|
7
|
+
## 1. Problem
|
|
8
|
+
|
|
9
|
+
The HTTP server has one shared key, `EIF_API_KEY` (`eif/mcp_server/auth.py`). Its own note says:
|
|
10
|
+
"For multi-user deployments, put a real gateway in front." Sessions live in one in-memory dict with
|
|
11
|
+
no owner (`eif/session.py`): anyone who holds a `session_id` can read that session.
|
|
12
|
+
|
|
13
|
+
That fits one person. It does not fit a **host app** that serves many of its own customers
|
|
14
|
+
(tenants) through one EIF server: every tenant would share one key, and a leaked session id would
|
|
15
|
+
open another tenant's claims and evidence.
|
|
16
|
+
|
|
17
|
+
## 2. Goal
|
|
18
|
+
|
|
19
|
+
A host app can give each of its tenants claim verification, with one secret shared between the host
|
|
20
|
+
and the server, no provisioning step, and no tenant able to read another's session.
|
|
21
|
+
|
|
22
|
+
The existing `EIF_API_KEY` keeps working exactly as today.
|
|
23
|
+
|
|
24
|
+
## 3. Design
|
|
25
|
+
|
|
26
|
+
### 3.1 The key
|
|
27
|
+
|
|
28
|
+
```
|
|
29
|
+
key = "eif_" + tenant + "_" + base64url_nopad( HMAC-SHA256(SECRET, "eif:" + VERSION + ":" + tenant) )
|
|
30
|
+
```
|
|
31
|
+
|
|
32
|
+
- `tenant` matches `^[a-z0-9][a-z0-9-]{0,55}$`. Lowercase only.
|
|
33
|
+
- The MAC is always 43 characters. The server reads the tenant as the text between the prefix and
|
|
34
|
+
the last 44 characters, never by splitting on `_`.
|
|
35
|
+
- Neither the host nor the server stores the key.
|
|
36
|
+
- The signed text starts with the product name, so a key made for one server is not valid on
|
|
37
|
+
another even if an operator reuses a secret.
|
|
38
|
+
|
|
39
|
+
Test vector (each project that implements this scheme carries its own, same secret and tenant):
|
|
40
|
+
|
|
41
|
+
| Secret | Version | Tenant | Key |
|
|
42
|
+
|---|---|---|---|
|
|
43
|
+
| `test-secret-do-not-use` | `v1` | `acme` | `eif_acme_Zz3YW-iyGIXp3c4AL9D0Kcpxxyneizso6QEKw0_Lf9I` |
|
|
44
|
+
|
|
45
|
+
### 3.2 Server settings
|
|
46
|
+
|
|
47
|
+
| Variable | Meaning | Default |
|
|
48
|
+
|---|---|---|
|
|
49
|
+
| `EIF_TENANT_KEY_SECRET` | The shared secret. Not set → the feature is off | not set |
|
|
50
|
+
| `EIF_TENANT_KEY_VERSIONS` | Comma-separated versions the server accepts | `v1` |
|
|
51
|
+
| `EIF_REVOKED_TENANTS` | Comma-separated tenants whose keys are refused | empty |
|
|
52
|
+
|
|
53
|
+
### 3.3 Verification (E1)
|
|
54
|
+
|
|
55
|
+
`validate_key` accepts a bearer key that is not `EIF_API_KEY` when the secret is set, the key has the
|
|
56
|
+
`eif_` prefix and a valid tenant, its MAC matches for an accepted version (constant-time compare),
|
|
57
|
+
and the tenant is not revoked. Its `key_info` gains `tenant`.
|
|
58
|
+
|
|
59
|
+
A request made with `EIF_API_KEY` has no tenant.
|
|
60
|
+
|
|
61
|
+
Three things in today's code need care:
|
|
62
|
+
|
|
63
|
+
- **Order.** With `EIF_API_KEY` unset outside production, `validate_key` accepts any key as a local
|
|
64
|
+
developer before anything else is checked; in production it rejects everything. When the tenant
|
|
65
|
+
secret is set, the tenant check runs first, and a server with a tenant secret and no
|
|
66
|
+
`EIF_API_KEY` serves tenant keys only.
|
|
67
|
+
- **The middleware discards the result today** (`http_server.py` calls `require_api_key` and drops
|
|
68
|
+
what it returns). It keeps it.
|
|
69
|
+
- **`POST /verify`** is a second way in: it reads the key from the JSON body and opens a session
|
|
70
|
+
itself. It attaches the same tenant and owner as the MCP path.
|
|
71
|
+
|
|
72
|
+
The tenant is read from **each message's own request**, not from a value set when the MCP
|
|
73
|
+
session was opened. With stateful Streamable HTTP the server task for a session starts inside the
|
|
74
|
+
request that created it, so a per-request variable set there stays the first caller's. A session is
|
|
75
|
+
also bound to the identity that opened it: a message whose key resolves to a different identity
|
|
76
|
+
(another tenant, or an operator key) is refused. Without that, one MCP session reused for two
|
|
77
|
+
tenants would run both as the first.
|
|
78
|
+
The compliance module already reads the message's own request this way (`compliance.py`,
|
|
79
|
+
`_request_identity`).
|
|
80
|
+
|
|
81
|
+
### 3.4 A session has an owner (E2)
|
|
82
|
+
|
|
83
|
+
- `new_session` records the tenant of the request that created it (none for `EIF_API_KEY`).
|
|
84
|
+
- `get_session` and `update_session` take the caller's tenant. A session asked for by a different
|
|
85
|
+
tenant raises the same "not found or expired" error as a session that does not exist, so the
|
|
86
|
+
answer does not reveal that the id is real.
|
|
87
|
+
- The same check applies to the five session resources (`eif://session/{id}/…`).
|
|
88
|
+
- A session with no owner is readable only by callers with no tenant, as today.
|
|
89
|
+
|
|
90
|
+
### 3.5 What a tenant key may call (E3)
|
|
91
|
+
|
|
92
|
+
| Allowed | Refused |
|
|
93
|
+
|---|---|
|
|
94
|
+
| `eif_new_session`, `eif_extract_claims_from_decision`, `eif_verify` | Every other tool |
|
|
95
|
+
|
|
96
|
+
`eif_challenge` and the causal evidence probe call OpenAI on the server operator's key. They stay
|
|
97
|
+
out of tenant reach so a tenant cannot spend the operator's money. `tools/list` for a tenant key
|
|
98
|
+
returns only the three allowed tools.
|
|
99
|
+
|
|
100
|
+
`eif_get_session` is left out on purpose: it reports only the provenance chain, which only
|
|
101
|
+
`eif_record` writes, and `eif_record` is not offered. A tenant would always read an empty session.
|
|
102
|
+
The verdict a tenant needs comes back from `eif_verify` itself.
|
|
103
|
+
|
|
104
|
+
### 3.6 What a host should send
|
|
105
|
+
|
|
106
|
+
The server works best when the host, not the model, fills two arguments of `eif_verify`:
|
|
107
|
+
|
|
108
|
+
- `conversation_turns`: without them the sycophancy check reports "SKIPPED - no conversation
|
|
109
|
+
context".
|
|
110
|
+
- `host_tool_outputs`: data the host already fetched. It is the strongest evidence tier, and it
|
|
111
|
+
only means something when the host supplies it.
|
|
112
|
+
|
|
113
|
+
## 4. What does not change
|
|
114
|
+
|
|
115
|
+
- `EIF_API_KEY`: same tools, same sessions.
|
|
116
|
+
- Stdio transport and local development mode.
|
|
117
|
+
- Sessions stay in memory. A restart loses them for every caller; a host opens a new one and
|
|
118
|
+
retries.
|
|
119
|
+
|
|
120
|
+
## 5. Tests
|
|
121
|
+
|
|
122
|
+
- The test vector verifies; a tampered MAC, an unknown version, a revoked tenant and a tenant that
|
|
123
|
+
breaks the rule are refused.
|
|
124
|
+
- Tenant B asking for tenant A's session, by tool or by resource, gets "not found or expired".
|
|
125
|
+
- A session created with `EIF_API_KEY` is not readable with a tenant key, and the reverse.
|
|
126
|
+
- A tenant key calling `eif_challenge` is refused; `tools/list` shows three tools.
|
|
127
|
+
- With the tenant secret set and `EIF_API_KEY` unset, in development and in production, a valid
|
|
128
|
+
tenant key is accepted and a tampered one is refused.
|
|
129
|
+
- `POST /verify` with a tenant key creates a session owned by that tenant.
|
|
130
|
+
- An MCP session opened with one key and continued with another is refused.
|
|
131
|
+
- `EIF_API_KEY` behaves as before.
|
|
132
|
+
|
|
133
|
+
## 6. Files
|
|
134
|
+
|
|
135
|
+
`eif/mcp_server/auth.py`, `eif/mcp_server/http_server.py`, `eif/mcp_server/server.py`,
|
|
136
|
+
`eif/mcp_server/resources.py`, `eif/session.py`, `SECURITY.md`, a new
|
|
137
|
+
`docs/integrations/HOST_APP.md`, tests. `docs/architecture.md` gains one paragraph under
|
|
138
|
+
"MCP surface" once this is built.
|
|
@@ -6,7 +6,7 @@ EIF ships a FastMCP server exposing 25 tools. The six you will touch most:
|
|
|
6
6
|
|---|---|
|
|
7
7
|
| `eif_new_session` | Create a session; every other stateful tool requires its session_id |
|
|
8
8
|
| `eif_extract_claims_from_decision` | Turn a plain-text decision into structured claims ranked by HALT probability |
|
|
9
|
-
| `eif_verify` | One-call full pipeline: declare, falsify, calibrate, route. Returns PASS or HALT |
|
|
9
|
+
| `eif_verify` | One-call full pipeline: declare, falsify, calibrate, route. Returns PASS, REVISE or HALT |
|
|
10
10
|
| `eif_get_context` | Load EIF thresholds and the pipeline quick reference (granular pipeline only) |
|
|
11
11
|
| `eif_demo` | Zero-setup HALT example, no API calls |
|
|
12
12
|
| `eif_check_rules_installed` | Confirm the rules block below is installed |
|
|
@@ -45,7 +45,7 @@ claude mcp get eif
|
|
|
45
45
|
|
|
46
46
|
```bash
|
|
47
47
|
pip install 'eif-engine[mcp]'
|
|
48
|
-
claude mcp add eif -- /path/to/venv/bin/python -m eif.mcp_server.server
|
|
48
|
+
claude mcp add eif --scope user -- /path/to/venv/bin/python -m eif.mcp_server.server
|
|
49
49
|
```
|
|
50
50
|
|
|
51
51
|
### Option C - project scope (`.mcp.json`)
|
|
@@ -90,7 +90,7 @@ without being asked. For a single project, put the same block in the
|
|
|
90
90
|
project's `CLAUDE.md` instead.
|
|
91
91
|
|
|
92
92
|
```
|
|
93
|
-
# <BEGIN-EIF v1.
|
|
93
|
+
# <BEGIN-EIF v1.5>
|
|
94
94
|
# ==============================================================================
|
|
95
95
|
# EIF - Epistemic Integrity Framework
|
|
96
96
|
# ==============================================================================
|
|
@@ -119,8 +119,10 @@ project's `CLAUDE.md` instead.
|
|
|
119
119
|
# 3. Pass data you ALREADY fetched (DB rows, API results, tool output) as
|
|
120
120
|
# host_tool_outputs. That is P0 evidence, the highest tier, and it is what
|
|
121
121
|
# lets EIF check a number against ground truth instead of guessing.
|
|
122
|
-
# 4. Verdict HALT: do not act; surface halt_cards to the user. REVISE:
|
|
123
|
-
#
|
|
122
|
+
# 4. Verdict HALT: do not act; surface halt_cards to the user. REVISE: a
|
|
123
|
+
# claim had no evidence or weak evidence (below 0.70); pass the data you
|
|
124
|
+
# already fetched as host_tool_outputs, or gather more, then verify again.
|
|
125
|
+
# PASS: every checked claim reached 0.70; proceed.
|
|
124
126
|
# 5. Re-verify after the user pushes back across turns: sycophantic drift is
|
|
125
127
|
# otherwise invisible.
|
|
126
128
|
#
|
|
@@ -129,7 +131,7 @@ project's `CLAUDE.md` instead.
|
|
|
129
131
|
#
|
|
130
132
|
# Do NOT run eif_verify on every reply. Reserve it for consequential decisions;
|
|
131
133
|
# routine chatter does not need it.
|
|
132
|
-
# <END-EIF v1.
|
|
134
|
+
# <END-EIF v1.5>
|
|
133
135
|
```
|
|
134
136
|
|
|
135
137
|
---
|