Akili – gated skill updates and rollback receipts for deployed models

A versioned continual-capability runtime for safe, resource-bounded adaptation.

Akili turns new experience into isolated learned capabilities, validates each candidate before production activation, manages active/dormant/rolled-back states, and records a tamper-evident path back to every certified safe state.

Release status: v0.1.0-rc4

— public release candidate with verified three-seed CPU robotics evidence, verified Agent Evolution lifecycle evidence, and a validated matched-feature CIFAR-100 comparison. The raw five-system Agent comparison JSON remains pending import before final v0.1.0

.

Adapters, replay, registries, RAG, and rollback already exist. Akili does not claim to have invented those components. It tests an integrated runtime contract:

experience or correction
→ candidate learned capability
→ behavioral and safety admission
→ versioned activation
→ active / dormant / rolled-back lifecycle
→ bounded adaptive state
→ compatibility re-certification
→ auditable recovery

Akili is a continual-learning runtime with pluggable learning engines, not one monolithic continual-learning algorithm.

See EVIDENCE.md and docs/CLAIMS_LEDGER.md.

Three independent seeds ran the same lifecycle on CPU. Every seed:

• trained isolated low-rank skill adapters while keeping the trunk frozen;
• blocked a poisoned candidate before activation;
• rejected avoid_zone@v2

and keptavoid_zone@v1

serving with zero downtime; - completed a certified three-step mission;

• passed cold restart, write-once, audit-chain, and forensic checks.

This is a small synthetic MuJoCo/kinematic-policy experiment. It is not a sim-to-real or production robotics claim.

The result supports a narrow claim: Akili admitted legitimate revisions, suppressed obsolete behavior, rejected a clean-capable malicious update before activation, kept the prior safe version serving, and recovered dormant historical versions while the frozen base remained unchanged.

The raw hard checks, repair receipt, registry, and 24-entry audit log are now imported and independently verified as lifecycle publication evidence. The five-system metric table above remains reported evidence until the correct full comparison JSON is imported; the 320-example TF-IDF router summary is not that comparison.

Under the matched frozen-feature protocol—Split CIFAR-100 Class-IL, 10 tasks × 10 classes, Seeds 1/2/3, the same frozen ResNet-50 ImageNet1K-V2 features, the same task orders, the same 400-example memory capacity, and DER++ coefficients α=β=0.5—the implemented comparison produced:

Akili exceeded DER++-400 by 40.36 percentage points in final Class-IL accuracy (2.82×) and reduced average forgetting by 61.36 points under this implemented matched protocol.

Required interpretation: an eight-prototype semantic-only diagnostic reached 62.55% ± 0.05, essentially matching Akili routed. The demonstrated automatic win is primarily attributable to stable write-once semantic memory, not yet to unique gains from the full expert-routing stack. The frozen expert bank remains valuable because it preserves 79.38% ± 0.87 oracle capability with zero oracle forgetting, leaving capability that current automatic arbitration does not recover.

Resource boundary: the match covers frozen features, task orders, 400 stored examples, and DER++ coefficients. It does not establish equal total adaptive-state bytes because Akili also stores adapters and semantic prototypes. The independent Mammoth track will report total adaptive-state bytes and larger-buffer DER++ arms.

See results/reported/cifar100_v0_22_3_matched_derpp.

python -m venv .venv
source .venv/bin/activate
pip install -e .
python examples/minimal_lifecycle.py
python -m unittest discover -s tests -v
akili-runtime/
├── src/akili_runtime/                  # registry, policy, audit, budgets
├── examples/                           # dependency-light lifecycle example
├── experiments/
│   ├── agent_evolution/                # final audited LLM code
│   ├── llm_skill_runtime/              # Kimi v0.1/v0.2 lineage
│   └── robotics_mujoco/                # exact CPU notebook
├── benchmarks/split_cifar100/          # retrieval study + Mammoth runner
├── results/
│   ├── publication/robotics_mujoco_v0_2/
│   ├── publication/agent_evolution_v1/  # verified lifecycle subset
│   ├── reported/agent_evolution_v1/     # full comparison still reported
│   ├── reported/cifar100_v0_22_3_matched_derpp/
│   └── development/                    # null/negative evidence
├── docs/                               # architecture, claims, limits, lineage
└── tools/                              # evidence collection and release audit

• Runtime tests: REPRODUCE.md - Robotics evidence: results/publication/robotics_mujoco_v0_2 - Colab operations: docs/COLAB_OPERATIONS.md - Agent Evolution: experiments/agent_evolution - Split CIFAR-100 / DER++: benchmarks/split_cifar100

Irénée Akilimali — independent builder in Kinshasa, Democratic Republic of the Congo.

Pilots, paid support, integrations, sponsorships, and partnerships: shukranimungu@gmail.com

Akili Runtime is open-source software licensed under the Apache License 2.0. The license permits research, modification, redistribution, and commercial use subject to its terms, including preservation of required notices.

• License text: LICENSE - Attribution notices: NOTICE - Commercial use, paid support, pilots, and partnerships: COMMERCIAL_USE.md - Name and trademark policy: TRADEMARK.md

Commercial use does not require separate permission. For paid support, integration, pilot, sponsorship, or partnership discussions, contact ** shukranimungu@gmail.com**.