Historical task source. The Eve task board holds current task state.
Phase 178: Autonomous Research & Agentic Scientist — Karpathy-Style Tweak Loops, Sakana-Class Tree Search, AlphaEvolve Evolutionary Discovery, Deep Research, MCTS Reasoning Substrate, Self-Play Learning, Physical-Lab Autonomy, Autoresearch Benchmarks & Governance#
Sovereign replication and extension of the April-2026 frontier
autonomous-research stack. Covers Karpathy's autoresearch (Mar 2026, ~630 LOC)
single-GPU overnight tweak loop and its async researcher-pool successor; Sakana
AI Scientist v2 agentic tree search with VLM figure critic and peer-reviewer
ensemble; Google DeepMind AI Co-Scientist (generate / debate / evolve) and
AlphaEvolve evolutionary program search (Gemini-powered, 4×4 matmul in 48
multiplications, 23% speed-up on Gemini training kernel); OpenAI Deep Research
(o3-deep-research / o4-mini-deep-research), Perplexity / xAI / Google deep
research siblings; AIDE + MLE-Bench, SWE-agent / SWE-ReX / mini-swe-agent, Neo,
Devin, Codegen; LATS, rStar-Math, Absolute Zero Reasoner, Voyager, EvoAgentX;
PaperBench, ScienceAgentBench, DiscoveryBench, RE-Bench, HLE; Coscientist (CMU,
Nature 2023), A-Lab (LBNL), Emerald Cloud Lab, Opentrons, Strateos.
This phase is the keystone that turns Oshun from a sovereign-AI platform into a sovereign-AI-research platform. It extends Phase 104 (Kalika theorem proving / research agents), Phase 128 (Kalika materials discovery workbench), Phase 130 (Kalika autonomous experimentation), Phase 47 (Nous agents), Phase 98 (test-time compute / A2A), and Phases 85–96 (training / eval / flywheel) with concrete autoresearch scaffolding that those phases do not implement.
Tech split: Python (~55%) for agent scaffolding, sandboxed code execution, and paper-generation tooling; TypeScript (~30%) for orchestration, dashboards, and workbench UI; Rust (~15%) for sandbox supervisor, parallel-execution runtime, and deterministic experiment-ledger serialisation.
Library layout: libs/nous/autoresearch-core/, libs/nous/lit-graph/,
libs/nous/tree-search-reasoning/, libs/nous/self-play-loop/,
libs/nous/ml-engineer-agent/, libs/nous/tweak-loop/,
libs/nous/evo-search/, libs/nous/paper-studio/, libs/nous/lab-driver/,
libs/nous/autoresearch-evals/, libs/nous/autoresearch-governance/.
178.1 Literature & Novelty Graph — @nous/lit-graph#
178.1.1 Scholarly Corpus Adapters#
- 178.1.1.1 Create
libs/nous/lit-graph/with project scaffolding. Build Semantic Scholar Graph API adapter exposingsearch(query, filters),paper(id),citations(id),references(id), and theembeddingendpoint for paper-vector nearest-neighbour novelty queries; exponential-backoff rate-limiting; batched bulk-fetch; local SQLite + MinIO cache - 178.1.1.2 Build OpenAlex adapter (300M+ records) with work / author / venue / concept endpoints and TSV-snapshot ingestion for offline full-corpus operations
- 178.1.1.3 Build arXiv OAI-PMH + API adapter with daily incremental
pull into the corpus; tie to existing
@sophia/crawlingfor dedup against the wider web corpus - 178.1.1.4 Build Crossref, PubMed E-Utilities, CORE, and bioRxiv / medRxiv adapters for biomedical and multi-publisher coverage
- 178.1.1.5 Build Connected Papers / Inciteful-style citation-graph traversal with bounded expansion budgets, dead-end detection, and relevance attenuation by hop distance and year
178.1.2 Novelty & Prior-Art Scoring#
- 178.1.2.1 Build
NoveltyScorer— given a proposed research idea (natural-language + keyword set), embed via Gemini-Embeddings-2 (Phase 47.1.4) and retrieve k-nearest prior papers from@nous/lit-graph; compute a multi-component novelty score (max cosine similarity, second-neighbour gap, citation-graph distance from proposal's seed paper, publication-year spread) - 178.1.2.2 Build prior-art collision detector — flag ideas whose top-3 neighbours exceed a similarity threshold and emit a diff explaining what is novel vs overlapping
- 178.1.2.3 Build method / dataset / claim structured extractor (LLM-driven, NeurIPS-guideline-shaped) over cached papers so novelty queries can match at the method-granularity, not just abstract-embedding
- 178.1.2.4 Build citation-quality ranking (venue tier + year + citation velocity + reviewer-agreement signal where available) so survey agents preferentially ground claims in high-quality sources
178.1.3 Survey & Deep-Research Agents#
- 178.1.3.1 Build
ScholarQA-class structured-survey agent producing citation-grounded multi-section reports with inline numbered references, per-section reference-quality score, and a hallucination-gate (every non-obvious claim must attach a paper id resolved through 178.1.1/178.1.2) - 178.1.3.2 Build STORM-class outline-then-expand survey agent with multi-perspective question generation and cross-referenced expansion, tuned for internal Oshun knowledge-bases (Sophia, Metis, Mnemosyne, Nisaba) in addition to public corpora
- 178.1.3.3 Build deep-research agent orchestrator wrapping
o3-deep-research/o4-mini-deep-research/ Gemini Deep Research / Claude Research / Perplexity Deep Research as interchangeable backends behind a single/research/deepAPI, with per-tenant budget and citation-quality enforcement - 178.1.3.4 Build self-hosted deep-research baseline using
@nous/llm+@nous/lit-graph+ browse-tool + citation-critic, benchmarked against the external deep-research APIs on a fixed Oshun-domain query set
178.2 Reasoning & Search Substrate — @nous/tree-search-reasoning#
178.2.1 Tree / Graph Search Kernels#
- 178.2.1.1 Implement Language Agent Tree Search (LATS) — MCTS with PUCT exploration, LM value + reflection nodes, reasoning-as-planning action space; reproduce published 94.4% HumanEval with Oshun self-hosted model backbone
- 178.2.1.2 Implement Tree-of-Thoughts (ToT) and Graph-of-Thoughts (GoT) variants as lighter-weight search patterns for cheap queries
- 178.2.1.3 Implement rStar-Math — policy SLM + code-augmented MCTS + Process Preference Model (PPM) trained via 4 rounds of self-evolution; target published 90.0% MATH with Qwen2.5-Math-7B-class backbone
- 178.2.1.4 Implement Reasoning-via-Planning (RAP) with world-model-based lookahead tying in Phase 176 RSSM / MuZero kernels for simulator-backed reasoning
- 178.2.1.5 Build unified
SearchKernel<State, Action>abstraction over LATS / MCTS / rStar / RAP, with pluggable value function, process-reward model, and expansion policy, reusable by every Phase 178 agent
178.2.2 Process & Outcome Reward Models#
- 178.2.2.1 Train process-reward model via MCTS-rollout labelling on Oshun internal math, code, and scientific-claim datasets (Phase 89.3.1.5 integration); use as step-level reward for search kernels
- 178.2.2.2 Train outcome-reward model for code (tests-pass) and math (answer-verified) via executor-grounded supervision, served as shared reward API across Phase 178 agents
- 178.2.2.3 Build reward-model drift monitor — periodically re-score historical high-reward trajectories, alert on silent miscalibration
178.3 Self-Improvement & Self-Play Reasoning — @nous/self-play-loop#
178.3.1 Self-Generated Curricula#
- 178.3.1.1 Implement Absolute Zero Reasoner (AZR, arXiv 2505.03335) — single model alternates proposer / solver roles generating abduction, deduction, induction tasks verified by a Python executor with learnability reward; TRR++ update; validate that training on zero external data improves baseline math/code eval
- 178.3.1.2 Implement STaR / V-STaR self-taught reasoner pipelines — bootstrap rationales, filter by final-answer correctness, fine-tune; V-STaR adds verifier-filtered chain-of-thought
- 178.3.1.3 Implement Self-Rewarding and Meta-Rewarding language models — LLM-as-judge loop producing its own DPO preference pairs, with meta-judge stabilising reward drift
- 178.3.1.4 Implement SPIRAL-style zero-sum multi-agent self-play for
reasoning transfer (per Phase 85.19.1.4 referenced baseline), now with
@nous/tree-search-reasoningsubstrate underneath
178.3.2 Skill Libraries & Voyager-Class Agents#
- 178.3.2.1 Build Voyager-class skill library — code skills indexed by embeddings with retrieval, composition, and iterative prompting; integrated with Phase 47 agent memory
- 178.3.2.2 Build automatic-curriculum generator proposing progressively harder tasks based on current-skill coverage gap analysis (Phase 85.17.1.6 quality-gated retraining hook)
- 178.3.2.3 Build EvoAgentX-style agent-workflow evolver — mutate agent graphs (roles, tools, prompts, search kernels), evaluate on held-out autoresearch tasks, keep-the-fittest with Pareto front on quality × cost × latency
178.4 Agentic ML Engineer — @nous/ml-engineer-agent#
178.4.1 AIDE-Class Tree-Search ML Engineer#
- 178.4.1.1 Implement AIDE (Weco, arXiv 2502.13138) — tree-search over Python ML code solutions, node = (code, score, diagnostics), expansion via LLM edits (improve / debug / diversify), beam + best-child selection; reproduce published MLE-Bench medal rate with Oshun self-hosted backbone
- 178.4.1.2 Build sandboxed execution backend on SWE-ReX-style parallel shell-exec runtime (one Docker sandbox per node expansion, seccomp + user namespaces + per-run filesystem snapshot + wall-clock + GPU-hour budget); deterministic resume and replay
- 178.4.1.3 Build MLE-Bench internal runner (75 Kaggle competitions, 24h/A10 budget) wired to the agent; publish continuous leaderboard vs Neo (34.2%), AIDE+o1-preview (16.9%), GPT-5.3 baselines
- 178.4.1.4 Build RE-Bench (METR, arXiv 2411.15114) internal runner — 7 ML-R&D tasks with human-expert baselines — publish agent vs human frontier plot per task per time budget (2h / 8h / 32h)
- 178.4.1.5 Build GPU-kernel autotuner agent — edits Triton / CUDA / PyTorch code, profiles wall-clock, verifies numerical equivalence vs reference; score on internal kernel library (attention, RMSNorm, rotary, RWKV-scan, state-space-scan)
178.4.2 SWE-Bench-Class Software Engineering Agent#
- 178.4.2.1 Fork and extend
mini-swe-agent/SWE-agent+SWE-ReXruntime; back with Claude Mythos / GPT-5.3 Codex / self-hosted@iris/codemodels; target ≥ 80% on SWE-Bench Verified self-test - 178.4.2.2 Build SWE-Bench Verified + SWE-Bench Pro + SWE-rebench (contamination-resistant) + SWE-Compass internal runners with leaderboard publication
- 178.4.2.3 Build SWE-Lancer freelance-job harness — score agents on economic-value-completed-per-dollar, not just pass-rate
- 178.4.2.4 Build SWE-EVO long-horizon repo-evolution harness — track agents over weeks-long feature-development windows with multi-PR state, CI signal, and review feedback integrated
178.5 Karpathy-Style Overnight Autoresearch Loop — @nous/tweak-loop#
178.5.1 Single-GPU Overnight Loop (Karpathy Pattern)#
- 178.5.1.1 Fork
karpathy/autoresearch(Mar 2026, ~630 LOC) as reference and port tolibs/nous/tweak-loop/adapting to Oshun conventions; preserve the core loop shape — agent edits a singletrain.py, runs a 5-minute training, readsval_bpb(or configured metric), keeps or discards based on delta, logs the decision with diff - 178.5.1.2 Build the Oshun target-training scripts — mirror
karpathy/nanochatat three scales (tweak-nano-125M,tweak-small-350M,tweak-mid-1B) each with a human-authoredprogram.mdprompt, frozenprepare.py, and one editabletrain.py; validation on held-out tokens, reproducible seed - 178.5.1.3 Build multi-backend agent driver — Claude Mythos / GPT-5.3
Codex / Gemini 3 Deep Think / self-hosted
@iris/codeas interchangeable code-editors with identical tooling contract; per-trial cost + latency + acceptance-rate reporting - 178.5.1.4 Build experiment-ledger (
@nous/tweak-loop/ledger) — immutable append-only log of{trial_id, parent, diff, command, metric_before, metric_after, delta, accepted, wall_clock, gpu_hours, model_rev}; visualised as a tree in the autoresearch workbench - 178.5.1.5 Build
/autoresearch/overnightone-command launcher — acceptstarget,model_backend,budget_gpu_hours,max_trials; emits morning summary (best delta, top-5 changes, failed-hypotheses histogram) to Slack / Oshun shell
178.5.2 Async Researcher-Pool Orchestration#
- 178.5.2.1 Extend the single-agent loop to an async researcher-pool — N agent-workers share the experiment-ledger, branch from any ancestor, avoid collision via advisory locks on subtrees, and gossip accepted improvements forward; the design goal is Karpathy's stated "community of PhD students, asynchronously massively collaborative"
- 178.5.2.2 Build subtree-assignment policy — UCB / Thompson sampling over active subtrees weighted by recent-delta and idle-time, to prevent N-1 agents from dog-piling the best branch
- 178.5.2.3 Build merge-conflict resolver — when two branches succeed on orthogonal changes, automatically attempt a combined trial and accept only if the composition improves
- 178.5.2.4 Build researcher-pool dashboard (
@nous/interp-dashboardintegration) — live tree view, per-agent throughput, cost-per-accepted-delta, hypothesis histogram, divergence metric across seeds
178.6 Evolutionary & Program-Synthesis Discovery — @nous/evo-search#
178.6.1 AlphaEvolve-Pattern Evolutionary Search#
- 178.6.1.1 Implement AlphaEvolve-pattern island model — N islands × M candidate programs per island, migration every k generations, LLM-driven mutation and crossover via Gemini 3 / Claude Mythos ensemble, evaluator pool running sandboxed scoring; elitism preserving per-island best
- 178.6.1.2 Build canonical test bed — 4×4 matrix-multiplication multiplication-count (target published 48), low-rank tensor decompositions, sorting-network depth, bin-packing, cap-set, kissing-number, and GPU-kernel autotune (Phase 178.4.1.5 interop) — fitness deterministically machine-checkable
- 178.6.1.3 Build Gemini-training-kernel optimisation harness — give the evolver the attention kernel source + the measured Gemini-training runtime benchmark; target the published ≥ 20% wall-clock improvement on equivalent Oshun kernels
- 178.6.1.4 Build math-problem sweep — run against ≥ 50 open problems from the Erdős / combinatorics / analysis canon, reproducing AlphaEvolve's reported SOTA-match-or-beat on ~95% of the set, and log any new SOTA discoveries with peer-review-ready write-ups
178.6.2 FunSearch-Pattern Program-of-Thought Search#
- 178.6.2.1 Implement FunSearch (DeepMind, Nature 2024) — evolutionary program search on problems with closed-form scorers; skeleton-preserving edits; Pareto elitism over (score, simplicity)
- 178.6.2.2 Build symbolic-regression integration — route Kalika Phase 104 symbolic-regression workloads through FunSearch for complex physical-law discovery benchmarks (Feynman-AI-Feynman, LLM-SR-Bench)
- 178.6.2.3 Build algorithm-discovery benchmark harness with known-optimal landmarks, to prevent silent regression as we scale the evolutionary backend
178.7 Paper Drafting, Review & Replication — @nous/paper-studio#
178.7.1 Authoring Pipeline#
- 178.7.1.1 Build Sakana-v2-class agentic tree search paper authoring — stages (ideation → hypothesis → experiment design → hparam tune → ablation → writeup), Experiment-Progress-Manager gating promotion, buggy-node classifier on failed trials
- 178.7.1.2 Build VLM figure critic — every generated figure passes a Gemini-Vision / Qwen3-VL check on caption clarity, axis labelling, unit hygiene, colour-blind-safe palette, and misleading-scale detection (Sakana v2 pattern); figures failing the critic trigger regeneration
- 178.7.1.3 Build LaTeX + Typst paper-build pipeline tied to Phase 139 scholarly-authoring stack — reproducible builds from a single YAML spec, deterministic bibliography, paper-embedding novelty check before emission
- 178.7.1.4 Build code-artifact release pipeline — every paper emits a companion reproducibility bundle (Docker image, lock-file, seed, env-capture, wall-clock log, cost ledger), bundled and hashed into the Phase 111 reproducibility store
- 178.7.1.5 Build plagiarism and LLM-regurgitation check — compare
against
@nous/lit-graphcorpus and internal training-data lineage (Phase 85.17.1.3) before publication
178.7.2 Peer-Review Ensemble & Replication#
- 178.7.2.1 Build Sakana-v2 peer-review ensemble — 5 LLM reviewers prompted on NeurIPS reviewer guidelines, producing soundness / contribution / clarity scores and narrative comments; area-chair LLM aggregates into an accept / reject recommendation; target the published 69% balanced accuracy against held-out OpenReview ground truth
- 178.7.2.2 Gate every Phase 178 autoresearch-generated artefact behind the peer-review ensemble; artefacts below the accept threshold are flagged for human review or auto-revision
- 178.7.2.3 Build PaperBench (OpenAI Preparedness) internal runner — replicate 20 ICML Spotlight papers from PDF + rubric (8,316 items), measure step-wise replication pass rate; publish trajectory over time (published human-baseline 41.4%, best current-agent 21.0% — track the closing gap)
- 178.7.2.4 Build ScienceAgentBench (ICLR 2025, 102 tasks, 44 papers) internal runner — agent scientist performs data-driven discovery tasks end-to-end with executable deliverables
- 178.7.2.5 Build DiscoveryBench (Ai2, 264 real + 903 synthetic tasks) internal runner — track agent discovery-rate across six disciplines over releases
178.8 Physical-Lab Autonomy — @nous/lab-driver#
This extends Phase 130 (Kalika autonomous experimentation) with the specific cloud-lab / robotic-lab driver abstractions that the bleeding edge now relies on, and with Coscientist / A-Lab / ChemCrow integration patterns.
178.8.1 Cloud-Lab & Robotic-Lab Drivers#
- 178.8.1.1 Build Emerald Cloud Lab driver — emit Symbolic Lab Language (SLL) programs, submit via ECL Command Center API, poll for completion, retrieve artefacts; catalogue 200+ instrument models with capability metadata
- 178.8.1.2 Build Strateos driver — Autoprotocol emission, instrument-reservation API, sample-tracking; same abstraction surface as ECL
- 178.8.1.3 Build Opentrons OT-2 / Flex driver via the Opentrons Python API for on-premise wet-lab automation (biology, chemistry pilot studies)
- 178.8.1.4 Build unified
LabDrivertrait —reserve(instrument),run(protocol),observe(artefactSpec),log,abort; Emerald / Strateos / Opentrons / self-hosted A-Lab drivers all conform
178.8.2 Closed-Loop Scientific Campaigns#
- 178.8.2.1 Build Coscientist-pattern agent (CMU Gomes/Boiko, Nature 2023) — hardware-manual retrieval tool, protocol code-gen tool, reaction-planner tool, ECL / Opentrons execution; canonical integration test is Pd-catalysed Suzuki–Miyaura + Sonogashira cross-coupling optimisation
- 178.8.2.2 Build A-Lab pattern (LBNL Ceder + DeepMind, Nature 2023) — Materials-Project candidate selector (Kalika Phase 127 hookup) + ML recipe generator + robotic-synthesis driver + XRD phase-identification + active-learning selection loop; target the published 41/58 targets in 17 days as an internal capability benchmark
- 178.8.2.3 Build ChemCrow-pattern agent (Nat Mach Intel 2024) — 18 expert chemistry tools (molecular editors, synthesis planners, safety checkers) orchestrated by an LLM executive for end-to-end synthesis planning
- 178.8.2.4 Build FAIR-Chem / UMA foundation-model integration for chemistry property prediction at scale inside the lab-driver planning loop
- 178.8.2.5 Build Boltz-2 / AlphaFold3 / ESM3 integration for
protein-design campaigns within
@sophia/research-enginecollaborative pipelines
178.8.3 Safety & Dual-Use Controls#
- 178.8.3.1 Build wet-lab hazard classifier — every proposed protocol passes a controlled-substance, precursor-export, biosafety-level, and CWC-Schedule check before dispatch
- 178.8.3.2 Build mandatory human-in-the-loop on: BSL-3+ organisms, Schedule 1 chemicals, CRISPR-in-germline, any high-energy reaction, or any protocol tagged dual-use by the classifier
- 178.8.3.3 Build audit log retention and tamper-evident chain-of-custody for every wet-lab artefact produced by autonomous agents (regulatory / insurance / IP requirement)
178.9 Autoresearch Benchmarks, Observability & Experiment Ledger — @nous/autoresearch-evals#
178.9.1 Unified Benchmark Harness#
- 178.9.1.1 Adopt UK AISI
inspect_ai+inspect_evalsas the top-level eval harness; wire in the suites already listed (MLE-Bench, SWE-Bench Verified, RE-Bench, PaperBench, ScienceAgentBench, DiscoveryBench) plus GAIA, Cybench, CTF, and Humanity's Last Exam subsets - 178.9.1.2 Build autoresearch-specific suites: KarpathyLoopBench
(overnight-tweak delta reproduction on a held-out
program.mdset), AlphaEvolveMathBench (the 50-problem open-math sweep with deterministic scoring), and OshunDomainBench (internal multi-domain scientific-question suite grounded in Veritas / Metis / Sophia / Kalika) - 178.9.1.3 Build continuous-leaderboard publisher — per-checkpoint, per-agent-config scores auto-posted to the Phase 91 evaluation hub with regression alarms and promotion gating
178.9.2 Observability & Ledger#
- 178.9.2.1 Integrate Langfuse (OTel-native) + W&B Weave as dual
agent-observability backends; every Phase 178 agent emits structured
traces (parent/child span, inputs, outputs, tokens, cost, latency, tool
calls) via a single
@nous/autoresearch-coremiddleware - 178.9.2.2 Build experiments-as-first-class diffing — pick two Langfuse experiments, auto-compute per-prompt deltas (output, score, cost, latency, token usage) and render as a side-by-side report (Langfuse 2026 feature; own the API shape ourselves so we are not vendor-locked)
- 178.9.2.3 Build the canonical experiment ledger — append-only,
hash-chained, signed entries per trial with
{agent_rev, model_rev, code_diff_hash, seed, env_hash, compute_cost, wall_clock, metric_delta}; ledger is the source of truth for reproducibility and audit - 178.9.2.4 Build SWE-ReX-class parallel sandbox pool as the shared execution substrate for every Phase 178 agent (Karpathy loop, AIDE, AlphaEvolve, Coscientist); sandbox isolates filesystem, network, CPU, GPU, and wall-clock per trial
- 178.9.2.5 Extend Phase 85 data flywheel with an
autoresearch-outcomescollector — ever trial's (prompt, proposed edit, measured delta, accepted) tuple is a high-value training example for improving the next generation of autoresearch agents
178.10 Multi-Agent Orchestration & Debate — @nous/autoresearch-core#
178.10.1 Generate / Debate / Evolve Multi-Agent#
- 178.10.1.1 Implement the Google AI Co-Scientist multi-agent pattern (arXiv 2502.18864) — specialised agents: Generation, Reflection, Ranking (Elo tournament over candidate hypotheses), Evolution, Meta-Review, and Proximity; scaled test-time compute controllable per query
- 178.10.1.2 Build Elo-tournament ranker with LLM-judge duels, diversity-bonus, and anti-collusion noise injection; calibration against human-expert rankings on held-out hypotheses
- 178.10.1.3 Validate on the published Co-Scientist case studies — drug-repurposing candidate proposal (liver fibrosis reference), AMR mechanism prediction — end-to-end within the Oshun stack
178.10.2 Backbone & SDK Choices#
- 178.10.2.1 Build interchangeable orchestration backends (LangGraph,
OpenAI Agents SDK, Anthropic Claude Agents, CrewAI, AutoGen, Google ADK)
behind a single
@nous/autoresearch-corefaçade; per-run selection based on feature coverage and cost - 178.10.2.2 Build A2A-interoperability adapter (Phase 98 integration) so Oshun autoresearch agents can collaborate with external third-party research agents under a signed capability-card exchange
- 178.10.2.3 Build per-run hermetic environment — every multi-agent run gets a deterministic model-revision pin, tool-version pin, and corpus-snapshot pin; records ledger entries sufficient to replay the exact run months later
178.11 Governance, Safety & Compute-Budget Controls — @nous/autoresearch-governance#
178.11.1 Budget & Scope Controls#
- 178.11.1.1 Build per-research-run compute budget governor — global GPU-hours, dollar-cost, and wall-clock caps per run, per researcher, and per org; hard stop on breach with auto-resume-on-approval
- 178.11.1.2 Build self-amplification guard — flagged any trial where the agent proposes self-modification of its own scaffold, budget, or governance code; route to human approval (Absolute Zero Reasoner explicitly called out this risk)
- 178.11.1.3 Build distribution-shift detector on autoresearch-generated artefacts — alert on rapid divergence from recent validated work, prompting a human-in-the-loop review before publication or deployment
178.11.2 Integrity & Oversight#
- 178.11.2.1 Build mandatory adversarial self-review pass before any trial is marked accepted — the agent must search its own diff for stub indicators (per CLAUDE.md stub list), LLM-regurgitated prior work, eval-contamination, and hardcoded shortcuts; only stub-free trials accumulate delta credit
- 178.11.2.2 Build dual-use risk classifier for proposed research directions (bioweapon, cyber-offense, CSAM, election-interference, CBRN); high-risk proposals halt and escalate
- 178.11.2.3 Build Kuan Yin integration (Phase 85.5) so autoresearch safety events feed the same moderation telemetry as user-product safety events, giving a single pane of glass
- 178.11.2.4 Publish "Autoresearch Accountability Report" per quarter — trials run, accepted deltas, novel results produced, cost, safety events, reviewer outcomes — sovereign transparency deliverable alongside the Phase 177 interpretability safety report