MathGraph is a verification-routing kernel. The canonical path is intentionally small:
claim or task
-> semantic validation boundary when an informal claim is present
-> formal claim / artifact
-> advisory route and constructor selection
-> verifier / finite checker / trusted importer / chain audit
-> EvidenceManifest
-> replay
-> invariant checks
-> Lawbook acceptance
-> Reason Atlas routing memory
Every accepted claim ends in exactly one terminal form:
VERIFIED_PROOFFINITE_COUNTERMODELNAMED_OBSTRUCTION
Finite-search failure is not proof. Raw verifier output is not boundary evidence by itself. Advisory route pressure, H-Tilt scores, Reason Atlas entries, semantic intake, and model output can guide work but cannot verify claims.
mathgraph/certificates.py: terminal forms and compact certificatesmathgraph/invariants.py: executable trust-boundary checksmathgraph/evidence_manifest.py: replayable evidence manifest schemamathgraph/evidence_replay.py: manifest replay checksmathgraph/lawbook.py: Lawbook entry dataclasses and review surfacemathgraph/lawbook_acceptance.py: manifest-backed acceptance contractmathgraph/lawbook_boundary.py: canonical terminal-admission façademathgraph/lawbook_ingest.py: boundary-gated Lawbook ingest helpersmathgraph/lawbook_query.py: query façade for reusable artifactsmathgraph/lawbook_export.py: manifest, JSONL, and summary exportsmathgraph/lawbook_reuse.py: reuse and action-change signalsmathgraph/reason_atlas.py: advisory routing memory and verifier-backed metricsmathgraph/semantic_validation.py: informal/formal claim boundarymathgraph/finite_magma_world.py: small deterministic finite checker worldmathgraph/finite_magma.py: finite magma representation and checked countermodel certificatesmathgraph/magma_constructors.py: deterministic constructor familiesmathgraph/sat_cache.py: constructor satisfaction cachemathgraph/policy_engine.py: advisory route policy buildermathgraph/proof_congruence.py: bounded TRUE-side congruence tracesmathgraph/true_proof_templates.py: proof-template family inventorymathgraph/verifier_execution.py: local verifier execution boundarymathgraph/verification_loop.py: stable loop façademathgraph/compounding_engine.py: canonical memory-becomes-capacity runnermathgraph/kernel.py: compact kernel acceptance surface
python scripts/run_release_check.py --quick
python scripts/run_repo_architecture_audit.py
python scripts/run_mathgraph_compounding_loop.py --allow-fallback-demo --out-dir /tmp/mathgraph_compounding_demo
python scripts/run_mathgraph_compounding_engine.py --out-dir /tmp/mathgraph_compounding_demo --episodes 2 --tiny-demo
python scripts/run_true_side_inventory.py --out-dir /tmp/mathgraph_true_inventory_demo --tiny-demoThe compounding command is the canonical repo-level loop. Fallback mode proves the wiring without claiming real SAIR results; real SAIR mode requires explicit equation and matrix paths.
Recursive residual compounding is the first stronger real-corpus compounding benchmark:
generic finite-countermodel route
-> residual frontier
-> residual-mined advisory constructors
-> recursive memory generations
-> compact constructor atlas
-> held-out transfer evaluation
-> TRUE contamination controls
Run the fallback-safe path with:
python scripts/run_recursive_residual_compounding.py --profile smoke --allow-fallback-demo --out-dir /tmp/mathgraph_recursive_residual_smokeCompact atlas routes are advisory scheduling objects. They do not verify claims or enter terminal Lawbook memory without independent boundary-backed evidence.
PQ-IR is the symbolic feature layer for ETP implications. It parses binary magma equations, builds bounded quotient-state features, classifies residual basins, and emits advisory constructor-family recommendations. It feeds compounding and residual-routing code, but it is not a truth boundary.
The ETP compounding engine is the repo-native constructor/residual loop:
episode -> constructors -> finite checking -> residuals -> obstruction atlas
-> repair constructors -> Lawbook update -> next episode
It writes a lightweight SQLite Lawbook plus CSV/JSON reports. Repair constructors and obstruction rows are advisory; only concrete finite countermodel certificates can support FALSE terminal candidates.
The TRUE-side inventory generates bounded congruence traces and Lean skeletons.
These are candidate proof-template artifacts until a proof verifier accepts
them. Bounded closure can guide future proof routes, but it is not
LEAN_VERIFIED.