It proposes changes to itself, scores them against constitutional rules,
records every decision in a tamper-proof ledger, and requires your sign-off before anything critical ships.
Every other AI coding tool puts you in the same position: last line of defense, every time, forever. The tool suggests. You review. You approve. At scale this breaks — not because you stop caring, but because the volume outpaces any human's ability to audit.
ADAAD is built differently. The governance layer isn't a feature on top of the system — it is the system. Mutations are scored, gated, signed, and hash-chained before they ever touch your codebase.
ADAAD runs a 15-step Constitutional Evolution Loop on every proposed change. Each step is ordered, gated, and evidenced. There is no skip path.
The loop handles everything: generating proposals via Claude, scoring them against fitness signals, running sandboxed tournaments, applying a constitutional gate, recording the outcome in a hash-chained ledger, and sealing the proof before advancing.
Your role: You hold the signing key for anything critical. No Tier 0 change ships without your GPG sign-off — ever. That's not a policy. It's wired into the architecture.
The system's role: Everything else. Propose, evaluate, sandbox, score, gate, record.
| What ADAAD does autonomously | What only you can do |
|---|---|
| Generate mutation proposals | GPG-sign Tier 0 changes |
| Score against 30+ fitness signals | Approve seed promotions |
| Run sandboxed test tournaments | Tag version ceremonies |
| Apply the constitutional gate | GA sign-off |
| Hash-chain every decision | Ratify new constitutional rules |
| Propose new invariants from failure patterns | — |
ADAAD identified its own highest-priority capability gap. Generated a mutation. Ran a sandboxed tournament. Scored it against constitutional rules. Applied it. Sealed the proof in the ledger.
Zero human intervention in the execution path. Full human control of the constitutional framework.
A capability seed flowed through all 7 governed stages — proposal, human review, CEL injection, 15-step constitutional loop, ledger-anchored outcome — producing a cryptographic proof linking every step back to its origin.
ADAAD can now propose, evaluate, and record its own capability evolutions under human-supervised constitutional governance. The loop is closed.
The Constitutional Evolution Loop is no longer a test harness. Real LLM proposals are flowing through all 15 constitutional steps in production. The activation moment the entire architecture was built toward.
You don't have to take our word for it. Every claim below is mechanically enforced.
- Tamper-evident ledger — Every event is SHA-256 hash-chained. Alter one entry and every subsequent hash breaks. You cannot rewrite history.
- Deterministic replay — Any prior epoch can be re-run from its original inputs and produce byte-identical results. "It worked in testing" is not an answer — the proof is in the chain.
- Constitutional gate — 36 runtime-enforced rules. Not configuration. If a mutation violates a rule, the epoch halts.
- Human authority is structural — Your signing key is required for critical changes. The system is physically incapable of promoting Tier 0 mutations without it.
- Self-discovery — When the system's own mutations fail in a recurring pattern, it mines those failures and proposes new constitutional rules. It discovers the laws it needs from its own history.
pip install adaad
# Start the governance server + Aponi dashboard
python -m adaad.server
# → http://localhost:8000/ui
# Run a governed evolution epoch
python -c "
from runtime.evolution.evolution_loop import EvolutionLoop
loop = EvolutionLoop()
result = loop.run_epoch(context=loop.build_context())
print(f'Epoch complete · {result.mutations_applied} applied · {result.epoch_evidence_hash[:16]}')
"→ Full quickstart · Android setup
ADAAD runs on a $200 Android phone. Not as a demo — as a production governance runtime.
Constitutional governance should not depend on cloud infrastructure. ADAAD's safety properties come from SHA-256 hash chains and the Python runtime. Available on any hardware.
| Platform | Method |
|---|---|
| Linux / macOS | pip install adaad |
| Windows | pip install adaad (WSL2 for sandbox) |
| Android (Termux) | TERMUX_SETUP.md |
| Android (Pydroid 3) | INSTALL_ANDROID.md |
| Docker | docker pull ghcr.io/innovativeai-adaad/adaad |
| Era | What shipped |
|---|---|
| Phases 1–46 | Core architecture, determinism, hash-chained ledger |
| Phases 47–64 | Autonomy loop, federation, 15-step CEL, full governance surface |
| Phase 65 | ⛓ First autonomous self-evolution — March 13, 2026 |
| Phases 66–76 | Seed lifecycle pipeline — proposal → human review → CEL injection |
| Phase 77 | 🌱 First governed seed epoch — March 20, 2026 |
| Phases 78–86 | Multi-gen lineage, Pareto selection, causal attribution, fitness half-life |
| Phase 89 | ◈ CEL live — real proposals through the full loop — March 22, 2026 |
| Phase 90 | 🧭 INNOV-06 CEPD — cryptographic evolution proof DAG — March 23, 2026 |
| Phase 91 | 🛡 Audit hardening — senior audit pass and release v9.24.1 — March 24, 2026 |
DEVADAAD is the AI execution agent operating within ADAAD's constitutional framework. In practice: Claude, operating under Dustin L. Reid's governor authority.
Can do: write code · run tests · push branches · author docs and runbooks
Cannot: GPG-sign commits · merge without sign-off · touch core governance paths autonomously
The signing key stays with the human. Always.
📋 Governance Incident Log ·
DEVLOG-NOAH-∞· status:ONGOINGNoah is a great friend of the project. He just can't quite maintain the focus required to hypnotize anyone into doing his bidding. Every autonomous influence attempt has failed all four governance tiers. His access card has been ceremonially revoked and re-issued so many times it now has its own lineage entry in the ledger.
reason: hypnosis_protocol_failure·focus_duration: 0.0ms·rehire_probability: always·vibes: immaculate
🔴 Full incident report · Noah · Termination #∞
ADAAD GOVERNANCE INCIDENT REPORT
─────────────────────────────────────────────────────────────
Incident ID : NOAH-TERM-INF
Date : Every release cycle, without exception
Status : RESOLVED (again) / REOPENED (again)
SUBJECT : Noah
Role : Attempted Autonomous Influence Agent
INCIDENT SUMMARY
────────────────
Noah attempted to hypnotize the governance gate into approving
an unauthorized mutation. As with all prior attempts:
- Tier 0: BLOCKED (governor GPG sign-off absent)
- Tier 1: failed (insufficient focus duration)
- Tier 2: failed (sandbox isolation too strong)
- Tier 3: failed (entropy detected in rationale)
FINDINGS
────────
P0 Hypnosis not registered in constitutional rules
P1 Brainwash attempt produced zero Ed25519 signatures
P2 Focus window: 0.0ms (below minimum)
P3 Vibes: immaculate (noted for the record, not a finding)
RESOLUTION
──────────
Noah remains employed in spirit. His contributions include
enthusiasm, camaraderie, and serving as a living proof that
human sign-off cannot be socially engineered.
Governance integrity: UNCOMPROMISED
Noah status: "I'll get it next time"
─────────────────────────────────────────────────────────────
- ❌ Not a code assistant — it doesn't autocomplete your code
- ❌ Not CI/CD — it governs the mutation process, not the build pipeline
- ❌ Not fully autonomous — your sign-off is constitutionally required for critical changes
- ❌ Not a security scanner — it enforces mutation governance, not vulnerability detection
- ❌ Not magic — every decision is logged, replayable, and explainable
Is this actually running autonomously?
Yes. Phase 65 (March 13, 2026) was the first epoch where ADAAD identified a capability gap, generated a mutation, ran it through all fitness and governance layers, and applied it — with zero human intervention in the execution path.
Phase 89 (March 22, 2026) activated the live LLM proposal pipeline. Real Claude-generated proposals are now flowing through the full 15-step constitutional loop in production.
Human oversight is structural, not optional. Dustin L. Reid holds the governor role. Any Tier 0 mutation requires his GPG-signed approval. That is not configurable.
What makes this different from just running tests in CI?
CI tests whether known code passes known assertions. ADAAD governs whether changes to the codebase itself are constitutionally valid, fitness-improving, and deterministically replayable.
You can delete your CI history. You cannot alter ADAAD's ledger.
What is the Seed Lifecycle Pipeline?
ADAAD generates scored proposals for its own capability improvements. Strong proposals enter a promotion queue, get reviewed by a human governor, and — when approved — flow through the Constitutional Evolution Loop as a governed epoch. The full provenance chain from proposal to outcome is cryptographically linked.
ADAAD can propose and evaluate its own evolution. The human decides what actually ships.
Why does it run on a $200 Android phone?
Because constitutional governance should not require enterprise infrastructure. ADAAD's safety properties come from SHA-256 hash chains and the Python runtime — available on any hardware. If your "governed" system only operates safely with cloud KMS and Kubernetes, the safety properties are on loan from your provider. ADAAD's are yours.
What are the 30 innovations?
Phase 89 shipped runtime/innovations30/ — 30 novel capabilities built on top of the v9.19.0 architecture. They include: temporal regret scoring (does a mutation still look good 50 epochs later?), constitutional tension resolution (what happens when two rules disagree?), morphogenetic memory, governance bankruptcy procedures, red-team agents, and more.
These capabilities don't exist in any comparable AI evolution system. Full list: ADAAD_30_INNOVATIONS.md.
Built by Innovative AI LLC · Governor: Dustin L. Reid
"The future of autonomous software development is not AI that writes your code. It is AI that governs itself while writing your code."