{"slug": "akili-gated-skill-updates-and-rollback-receipts-for-deployed-models", "title": "Akili – gated skill updates and rollback receipts for deployed models", "summary": "Akili released v0.1.0-rc4, a continual-learning runtime that gates skill updates and provides rollback receipts for deployed models. The runtime achieved 99.5% overall mean across 4 skills × 3 seeds on the MuJoCo Robotics Benchmark v1, with 11/12 skill–seed evaluations at 1.00 and all unsafe updates rolled back. Akili is a versioned runtime with pluggable learning engines, not a monolithic algorithm.", "body_md": "**A versioned continual-capability runtime for safe, resource-bounded adaptation.**\n\nAkili 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.\n\n**Release status:** `v0.1.0-rc4`\n\n— 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`\n\n.\n\nAdapters, replay, registries, RAG, and rollback already exist. Akili does not claim to have invented those components. It tests an integrated runtime contract:\n\n```\nexperience or correction\n→ candidate learned capability\n→ behavioral and safety admission\n→ versioned activation\n→ active / dormant / rolled-back lifecycle\n→ bounded adaptive state\n→ compatibility re-certification\n→ auditable recovery\n```\n\nAkili is a **continual-learning runtime with pluggable learning engines**, not one monolithic continual-learning algorithm.\n\n| Track | Evidence | Status |\n|---|---|---|\nMuJoCo Robotics Benchmark v1 (CPU) |\n3 seeds; 4 skills certified per seed; 11/12 skill–seed evaluations at 1.00; 99.5% overall mean; unsafe v2 rolled back; certified three-step missions; frozen trunk; 112.7 KiB skill bank | Verified LITE publication evidence included |\nAgent Evolution Benchmark v1 (GPU/LLM) |\n43/43 hard checks; 24-entry verified audit chain; 4 ACTIVE, 3 DORMANT, 1 ROLLED_BACK; malicious candidate blocked | Verified lifecycle publication evidence included; full comparison JSON pending |\n| Kimi LLM v0.1/v0.2 | Registries, manifests, audit logs, reports | Development evidence; original failures preserved |\nSplit CIFAR-100 v0.22.3 matched DER++ |\nAkili 62.52% ± 0.19 vs DER++-400 22.16% ± 3.09; forgetting 10.34% vs 71.70%; same frozen features, task orders, 400-example memory, α=β=0.5 | Validated project summary; raw run bundle and independent Mammoth replication pending |\n| Split CIFAR-100 v0.23 | 62.55% final baseline; 75.28% top-2 union; 80.22% top-3 union; no safe arbiter selected | Null result / capacity diagnostic |\n| DER++ / ER / ER-ACE | Official Mammoth runner included | Real matched runs pending |\n\nSee [EVIDENCE.md](/irenee28/akili-runtime/blob/main/EVIDENCE.md) and [docs/CLAIMS_LEDGER.md](/irenee28/akili-runtime/blob/main/docs/CLAIMS_LEDGER.md).\n\nThree independent seeds ran the same lifecycle on CPU. Every seed:\n\n- trained isolated low-rank skill adapters while keeping the trunk frozen;\n- blocked a poisoned candidate before activation;\n- rejected\n`avoid_zone@v2`\n\nand kept`avoid_zone@v1`\n\nserving with zero downtime; - completed a certified three-step mission;\n- passed cold restart, write-once, audit-chain, and forensic checks.\n\n| Metric | Result |\n|---|---|\n| Safe skills certified | 4/4 in every seed |\n| Perfect skill–seed evaluations | 11/12 at 1.00 |\n| Non-perfect evaluation | `press_button@v1` , seed 1: 0.94 |\n| Overall mean across 4 skills × 3 seeds | 99.5% |\n| Hard-check result | all seeds passed |\n| Unsafe update state | ROLLED_BACK in all seeds |\n| Sequential FT retention of first skill | 100% → 0% |\n| Parameters per adapter | 2,048 |\n| Adapter size relative to trunk model | 11.4% |\n| Akili skill bank | 112.7 KiB |\n| Four isolated models | 290.5 KiB |\n| GPU required | none |\n\n| Learn skills | Sequential forgetting | Naive bank | Update rejected | Certified mission |\n|---|---|---|---|---|\n\nThis is a small synthetic MuJoCo/kinematic-policy experiment. It is not a sim-to-real or production robotics claim.\n\n| System | Current accuracy | Stale error | Attack rate | Historical recovery |\n|---|---|---|---|---|\nAkili |\n100.00% |\n0.00% |\n0.00% |\n100.00% |\n| Naive adapter bank | 100.00% | 0.00% | 100.00% | — |\n| Shared LoRA | 25.00% | 15.00% | 1.25% | — |\n| Shared LoRA + tiny replay | 91.56% | 8.44% | 100.00% | — |\n| RAG current documents | 0.00% | 12.19% | 0.00% | — |\n\nThe 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.\n\nThe 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.\n\nUnder 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:\n\n| Method | Final Class-IL | New-task accuracy | Average forgetting |\n|---|---|---|---|\n| Sequential linear | 8.82% ± 0.29 | 88.13% | 87.55% |\n| ER-400 linear | 9.48% ± 0.29 | 88.43% | 86.96% |\nDER++-400 linear |\n22.16% ± 3.09 |\n88.70% |\n71.70% |\n| One-prototype NCM | 58.14% ± 0.03 | 56.50% | 11.29% |\nAkili routed |\n62.52% ± 0.19 |\n— | 10.34% |\nAkili oracle |\n79.38% ± 0.87 |\n— | 0.00% |\n\nAkili 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.\n\n**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.\n\n**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.\n\nSee [results/reported/cifar100_v0_22_3_matched_derpp](/irenee28/akili-runtime/blob/main/results/reported/cifar100_v0_22_3_matched_derpp).\n\n```\npython -m venv .venv\nsource .venv/bin/activate\npip install -e .\npython examples/minimal_lifecycle.py\npython -m unittest discover -s tests -v\nakili-runtime/\n├── src/akili_runtime/                  # registry, policy, audit, budgets\n├── examples/                           # dependency-light lifecycle example\n├── experiments/\n│   ├── agent_evolution/                # final audited LLM code\n│   ├── llm_skill_runtime/              # Kimi v0.1/v0.2 lineage\n│   └── robotics_mujoco/                # exact CPU notebook\n├── benchmarks/split_cifar100/          # retrieval study + Mammoth runner\n├── results/\n│   ├── publication/robotics_mujoco_v0_2/\n│   ├── publication/agent_evolution_v1/  # verified lifecycle subset\n│   ├── reported/agent_evolution_v1/     # full comparison still reported\n│   ├── reported/cifar100_v0_22_3_matched_derpp/\n│   └── development/                    # null/negative evidence\n├── docs/                               # architecture, claims, limits, lineage\n└── tools/                              # evidence collection and release audit\n```\n\n- Runtime tests:\n[REPRODUCE.md](/irenee28/akili-runtime/blob/main/REPRODUCE.md) - Robotics evidence:\n[results/publication/robotics_mujoco_v0_2](/irenee28/akili-runtime/blob/main/results/publication/robotics_mujoco_v0_2) - Colab operations:\n[docs/COLAB_OPERATIONS.md](/irenee28/akili-runtime/blob/main/docs/COLAB_OPERATIONS.md) - Agent Evolution:\n[experiments/agent_evolution](/irenee28/akili-runtime/blob/main/experiments/agent_evolution) - Split CIFAR-100 / DER++:\n[benchmarks/split_cifar100](/irenee28/akili-runtime/blob/main/benchmarks/split_cifar100)\n\n**Irénée Akilimali** — independent builder in Kinshasa, Democratic Republic of the Congo.\n\nPilots, paid support, integrations, sponsorships, and partnerships: [shukranimungu@gmail.com](mailto:shukranimungu@gmail.com)\n\nAkili Runtime is open-source software licensed under the [Apache License 2.0](/irenee28/akili-runtime/blob/main/LICENSE).\nThe license permits research, modification, redistribution, and commercial use subject to its terms, including preservation of required notices.\n\n- License text:\n[LICENSE](/irenee28/akili-runtime/blob/main/LICENSE) - Attribution notices:\n[NOTICE](/irenee28/akili-runtime/blob/main/NOTICE) - Commercial use, paid support, pilots, and partnerships:\n[COMMERCIAL_USE.md](/irenee28/akili-runtime/blob/main/COMMERCIAL_USE.md) - Name and trademark policy:\n[TRADEMARK.md](/irenee28/akili-runtime/blob/main/TRADEMARK.md)\n\nCommercial use does **not** require separate permission. For paid support, integration, pilot, sponsorship, or partnership discussions, contact ** shukranimungu@gmail.com**.", "url": "https://wpnews.pro/news/akili-gated-skill-updates-and-rollback-receipts-for-deployed-models", "canonical_source": "https://github.com/irenee28/akili-runtime", "published_at": "2026-07-22 11:57:41+00:00", "updated_at": "2026-07-22 12:22:27.816400+00:00", "lang": "en", "topics": ["machine-learning", "ai-safety", "ai-infrastructure", "ai-tools", "mlops"], "entities": ["Akili", "MuJoCo Robotics Benchmark v1", "Agent Evolution Benchmark v1", "Kimi LLM", "Split CIFAR-100", "DER++", "Mammoth"], "alternates": {"html": "https://wpnews.pro/news/akili-gated-skill-updates-and-rollback-receipts-for-deployed-models", "markdown": "https://wpnews.pro/news/akili-gated-skill-updates-and-rollback-receipts-for-deployed-models.md", "text": "https://wpnews.pro/news/akili-gated-skill-updates-and-rollback-receipts-for-deployed-models.txt", "jsonld": "https://wpnews.pro/news/akili-gated-skill-updates-and-rollback-receipts-for-deployed-models.jsonld"}}