Arch-MIND · Architectural governance

Deterministic architectural governance.

Bit-identical scores, cryptographically-chained session deltas, rules as compile-time invariants (not CI warnings).

Request commercial license View showcase

Arch-MIND is STARGA Commercial-licensed. Source distribution and on-prem deployment by written agreement.

arch-mind

STARGA Commercial

Nine Q16.16 metric kernels
Four tree-sitter languages plus heuristic MIND scoring
HMAC-SHA256/Ed25519 evidence log

Two machines, one score

The same source state scores the same on x86_64 and aarch64

Nine Q16.16 metrics in integer arithmetic, so the published golden hash of the medium_realistic fixture reproduces byte for byte on either chip.

The two questions Arch-MIND answers

“Did the agents I just ran make my codebase structurally worse?”

arch-mind session-start snapshots architectural state before an agent works; session-end re-snapshots and emits an HMAC-chained (tamper-evident) delta to the previous event via HMAC-SHA256 (or Ed25519 when the key is initialised). The history is replayable, tamper-evident, and recallable across sessions.

answer: session deltas, chained, HMAC-authenticated (tamper-evident), replayable

“Is the code still architecturally consistent with my declared intent?”

.arch-mind/rules.mind declares compile-time invariants (acyclicity_q16 == 10000, q16_determinism_purity >= 9000, …). arch-mind check-rules --mode enforce fails closed when any rule breaks: not a warning, a hard stop.

answer: rules.mind invariants, compile-time, fail-closed

session-startsnapshot pre-agent state
agent works
session-endre-snapshot, emit delta
chained deltaHMAC-SHA256 or Ed25519
history --verifyreplay the chain

One session, drawn from the CLI below: each delta is chained to the previous event, and history --verify replays the whole chain.

Scoreboard: Mind-Mem (dated fixture)

Nine Q16.16 metric kernels over a static, dated fixture scan of six commits of star-ga/mind-mem, STARGA's flagship persistent-memory engine for AI coding agents (pip install mind-mem). This is a checked-in fixture, not a live feed; it is not refreshed on deploy.

fixture scan

Loading scan data…

Bit-identical reproducibility

The same source state produces the same Q16.16 score on any x86 or aarch64 machine running the same Arch-MIND version, byte-for-byte. Two machines, identical hash. No floating-point non-determinism, no machine-dependent rounding. Every audit, every delta, every session event is anchored to a verifiable integer that travels cleanly across systems.

medium_realistic fixture: golden hash

unsigned hex: 0xa9d26a15
signed i32: -1_445_828_075

This exact hash reproduces on any x86 / aarch64 machine running the same Arch-MIND version on the same source state. Run arch-mind scan --fixture medium_realistic.json and verify yourself after signing a commercial agreement.

The nine metrics

All Q16.16 fixed-point in [0, 10000] raw scale. The kernel returns the value << 16 packed as i32; the demo displays the raw integer.

modularity_q16Newman 2006
Fraction of dep-edges whose source and destination share a package
acyclicity_q16Tarjan 1972
Fraction of dep-edges NOT inside a non-trivial strongly-connected component
depth_q16Kahn 1962
Longest dep-chain depth, normalised to [0, 10000]
equality_q16Gini 1912
Inverted Gini coefficient of per-module symbol counts
redundancy_q16Baker 1995
Duplicate-pattern density (variance-of-variance proxy pre-AST)
q16_determinism_puritySTARGA-native
Fraction of modules with no f32/f64/nondet primitive in scope
evidence_chain_densitySTARGA-native
Σ evidence-chain calls / Σ decision points across all modules
mcp_tool_isolationSTARGA-native
Inverse of max transitive-dependency overlap between MCP tools
governance_kernel_coverageSTARGA-native
Fraction of declarations under [protection] or [invariant] annotation

How Arch-MIND compares

Factual claims drawn from each tool's public documentation. No screenshots, no quoted marketing text.

Criterion	Arch-MIND	CodeScene	Sigrid	SonarQube
Failure model	kernel-invariant (fail-closed compile-time)	CI warning	dashboard alert	CI warning
Score type	Q16.16 fixed-point, bit-identical	float, machine-dependent	float	float
Cross-machine reproducible	yes (verified)	no (claimed)	no	no
Evidence chain	HMAC/Ed25519 per session, replayable	none	none	none
Per-repo rule grammar	.arch-mind/rules.mind parseable by mindc	XML config	dashboard	XML profile

CLI surface

Ten subcommands. Every command is also available as a JSON-RPC MCP tool.

Arch-MIND CLI reference

arch-mind scan --fixture <path> [--out <path>]: Score a pre-built fixture JSON; emit nine Q16.16 metrics
arch-mind sidecar-scan --repo <path> [--lang python,typescript,go,rust,mind] [--out <path>]: Walk a live repo via tree-sitter sidecar; produce canonical-AST fixture
arch-mind rules --rules <path> --scan <path> [--mode enforce|report]: Evaluate .arch-mind/rules.mind invariants against a scan result
arch-mind explain --scan <path>: Print a human-readable breakdown of each metric for a scan
arch-mind baseline --fixture <path> --out <path>: Snapshot the current scores as an approved baseline for delta tracking
arch-mind delta --before <baseline> --after <baseline>: Compute signed score delta between two baseline snapshots
arch-mind session-start --repo <path> --fixture <path> --agent <name> --commit <sha>: Open a session; snapshot pre-agent architectural state, initialise HMAC chain
arch-mind session-end --repo <path> --fixture <path>: Close a session; emit HMAC-chained (tamper-evident) delta chained to the previous event
arch-mind history --repo <path> [--verify]: Replay and optionally verify the full HMAC/Ed25519 evidence chain
arch-mind check-rules --repo <path> --fixture <path> [--mode enforce|report]: Fail-closed invariant check; non-zero exit if any rule is violated

Arch-MIND is STARGA Commercial

Source distribution and on-prem deployment by written agreement. Want to run Arch-MIND on your repo? We'll set up a private trial with your code staying entirely on your infrastructure. The binary phones home only for license validation.

Request commercial license