R* = min(R_ord, R_bin) · 408 atoms · 28 areas
496/748 atoms · gate:PASS
AMPHARE
GOVERNED_AUTONOMY
amphare.com / instruments

One kernel.
Four instruments.

Amphare ships as a family. atomikd decomposes intent into governed atoms. AmphSE builds them under the kernel. CodeSentinel measures what shipped. RepoGraph maps what depends on what. One substrate underneath; a different door for every way in.

atomikd decomposes AmphSE builds CodeSentinel validates RepoGraph maps
01 / THE METHOD
atomikdUNDER DEVELOPMENT

The front door from intent to atoms. atomikd applies atomic decomposition, the methodology behind everything Amphare does: work is broken down until each piece is the smallest reviewable, traceable, estimable thing. Grounded in the published literature: The Physics of Delivery and the Work / Effort / Value atom model from The Fractal Org.

"Add MFA to the admin login"
7 atoms · 12.0h · every one reviewable, every one traceable

Estimation stops being a negotiation and becomes arithmetic. Requirements go in; typed, sized, dependency-ordered atoms come out, ready for a sprint plan or a kernel dispatch.

read the methodology briefing on request
atoms, not quarks: we stop splitting exactly where work stops being reviewable.
02 / THE INSTRUMENT
CodeSentinelUNDER DEVELOPMENT

The AI code audit instrument. 408 quality atoms across 28 areas, five rubric archetypes, eight compliance overlays. Evidence-tagged findings with a full reliability bundle. The instrument that doesn't average away your catastrophes.

408
quality atoms
28
audit areas
8
compliance overlays
R*
reliability bundle
evidence OBSERVED / INFERRED / ABSENT  ·  governance WARNROUTEBLOCK
R* = min(R_ord, R_bin) · stochastic in, deterministic out

Findings recompute live with zero new model calls: reviewers change a parameter, the math layer replays. Quality becomes a defensible number instead of a vibe.

visit codesentinel →
03 / THE KERNEL
AmphSEUNDER DEVELOPMENT

The autonomous software development governance kernel. Backlog to verified, deployable code through governed multi-agent orchestration: append-only ledger, tenant isolation, hash-chained audit trail. Machines do the work. The system stays accountable.

amph_se · terminal
Sprint Engine Kernel v0.4.7 | Wave 7 - MOA Consumers (Clarification / AWTI)
Autonomous software development governance kernel
 
Stack Python 3.12 | PostgreSQL 16 + RLS | asyncpg 0.30 | temporalio
Arch Append-only ledger | Tenant-isolated | Hash-chained audit trail
Phase 496/748 atoms complete | 45,000+ automated tests | mypy --strict · ruff · bandit
 
Commands /dispatch /review /verify /replay /ingest /events /qhistory /show /help
Type / for command menu | Tab to autocomplete | Ctrl+C twice to exit
 
project: cli-served-model-attribution-b28fdd
amph_se >
04 / THE MAP
RepoGraphUNDER DEVELOPMENT

The code intelligence layer, generated as a byproduct of the same single ingestion pass that runs the audit. It owns the two jobs agents cannot grep for: deterministic blast-radius, what breaks if this changes, as an auditable transitive query rather than a probabilistic guess, and the stable semantic map, domain boundaries, ownership, coupling, the skeleton that changes monthly while the code changes hourly. Every edge is epistemically banded.

click a module · reverse-CALLS blast radius · OBSERVED solid · INFERRED dashed · selectedimpacted

Change the database module and the graph answers, deterministically, who transitively depends on it. Solid edges are compiler-grade OBSERVED facts; dashed edges are INFERRED dynamic links, labeled as such because the map never claims more certainty than it has.

explore RepoGraph → graph briefing on request