MMLC Runtime 1.0.0

A ledger that can be re-run, and says what it did not conclude.

MMLC executes typed matrix-ledger documents. A document descends through nine stages and comes out as canonical JSON, hashes, an event log and a report. Each stage establishes something specific — and the specification records, stage by stage, what it refuses to establish.

in — MMLF

out — canonical JSON, hashes, events, manifest, report

stage 01 / 09

MMLF YAML / JSON

establishes

Schema and structural validation.

JSON Schema validates document shape. Runtime validation additionally checks IDs, layouts, references, operator domains, dependency cycles, intervention conflicts, and fixed-point contracts.

The nine stages of docs/ARCHITECTURE.md, with the separation-of-concerns statements attached to the stages they constrain. Select any stage.

Status

Runtime: 1.0.0
Stable document profile: MMLF 1.0
Backward-compatible loaders: MMLF 0.1 through 0.9
Python: 3.10+
Licence: Apache-2.0
Test suite: 71 tests before release validation
Historical experiments: E0–E9

established

Verified on a clean runner across Python 3.10, 3.11, 3.12 and 3.13: 71 passed, then validate, migrate and run --fail-on-audit on the stable example.

What 1.0 froze

Version 1.0 added no new mathematical capability. It is a stabilization and publication boundary — the point at which a run of prototypes became something installable, migratable and maintainable in public.

The Runtime 1.x public Python API.
The MMLF 1.0 document profile.
Loaders for MMLF 0.1 through 0.9.
Deterministic legacy-to-1.0 migration, with execution-equivalence verification.
Historical feature semantics, preserved through metadata.migrated_from.
All schemas, packaged inside the installed Python distribution.
Machine-readable CLI errors and documented exit codes.
Release benchmarks, compatibility documentation, and a release-verification experiment.

It also corrected a real defect: unknown future MMLF versions were being silently treated as v0.1 rather than rejected.

The capability stack

Ten years of separate tools would each own one of these. MMLC's claim is narrower and stranger: that they can share one execution model, and therefore one audit.

Typed transactions and operator invariants
Local, regional and global audit
Dependency taint and root-cause paths
Matrix layout and multidirectional traversal
Row, column and block constraints
Minimum-support repair proposals
Temporal references and fixed-point groups
Append-only corrections
Hard and soft counterfactual interventions
Branch differential ledgers
Discrete and deterministic continuous uncertainty
Policy scoring, information value and finite sequential decisions

Run it

bash

python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
python -m pip install -e ".[dev]"

bash

mmlc validate examples/mmlf_v1_stable.yaml
mmlc run examples/mmlf_v1_stable.yaml \
  --output outputs/quickstart \
  --deterministic \
  --fail-on-audit

Or from Python, where the audit status is a value you can assert on:

python

from mmlc import execute_file, validate_file

summary = validate_file("examples/mmlf_v1_stable.yaml")
result = execute_file("examples/mmlf_v1_stable.yaml", deterministic=True)

assert summary.version == "1.0"
assert result.global_audit["status"] == "PASS"

Scope

MMLCa replacement for spreadsheets, DAG engines, databases, symbolic systems, SMT solvers, causal libraries, or probabilistic programming frameworks
A declared causal graphthe true graph of the external world
Finite verificationproof

The full list of non-claims is on the limits page, and it is part of the specification rather than a disclaimer appended to it.