Architecture & formats
Separating canonical data, decidable policy, and rendered views.
Project Meno explores what happens when data files are never executable, policy gates are finite and exhaustively checked, and human-facing prose is projected deterministically from structured records.
1 · Structured formats
.mnd documents and .mnv vocabularies.
In conventional repositories, operational truth drifts across prose markdown files, shell scripts, and configuration formats that require ad hoc parsers or full-file reads. Meno separates canonical records from rendered views:
- Canonical data (
.mnd) and verified vocabularies (.mnv): Flat, deterministic directives and typed slots that hold zero executable code (enforcing write-xor-execute discipline across documentation and state). - Bounded node slicing: Automated tools read and update individual addressed nodes by identifier and content digest rather than loading entire files into memory or into a language model’s context window.
- Deterministic human views: People read clean rendered projections — tables, manuals, and plain-English summaries — generated from the underlying structured records without manual synchronization drift.
2 · Decidable gates
Finite-state policy tables (.mni) with 2^k verification.
Security and workflow policies written in general-purpose programming languages or complex rule engines quickly become impossible to verify across every combination of inputs. Meno constrains policy gates deliberately:
- Each
.mnigate accepts only finite boolean or enumerated inputs, capped atk <= 16inputs so the complete truth space (2^k <= 65,536states) can be evaluated exhaustively in milliseconds. - Compiler and linter checks reject non-total tables, vacuous branches, and surviving test mutants before a policy artifact is accepted.
- Ambient enforcement separates polite, deterministic command reroutes from hard fail-closed security denials on credentials, off-limits paths, and irreversible operations.
3 · Storage substrate
Content-addressed storage and content-blind indexing (MNFS).
MNFS investigates storage and indexing primitives tailored to verifiable workspaces and multi-process local state:
- Immutable content-addressed objects: Read-only SHA-256 content-addressed blobs paired with atomic swap updates, eliminating torn writes and silent in-place mutation.
- Deduplicated volume snapshots and loopback mounts: Block-aligned deduplication across iterative workspace generations, paired with unprivileged local virtual-filesystem projection rather than kernel-extension workarounds.
- Content-blind relational whiteboard: Signed group and isolation-wall edges that answer structural neighborhood queries without opening or indexing opaque, encrypted, or restricted files.
4 · Agent workflows
Context-metered execution and continual-learning guardrails.
Beyond static formats, Project Meno studies practical harnesses for software-engineering and research agents:
- Context virtual memory: Bounded slice reads, folded command outputs, and explicit session resume cards so long-running tasks preserve output quality without context-window thrashing.
- Two-beat step-back and failure circuit breakers: Requiring an explicit check for simpler paths or upstream environmental barriers before an agent writes retry loops or fragile workarounds.
- Tiered rule lifecycle: Candidate operational rules progress through observation and multi-session cooling-off evidence before human confirmation promotes them into standing policy.