research-package RP-2026-07-23-002
AI Knowledge Gap Analysis
AI Knowledge Gap Analysis
Summary
The repository has unusually strong architectural hypotheses about externalized state, research relays, context, agent coordination, and knowledge management. Its dominant weakness is empirical: concepts are not connected to executable experiments, measured baselines, formal registries, or validated decision thresholds.
| Topic | Known in repository | Unknown / assumptions | Conflicting evidence | Research opportunity | Recommended experiment | Deliverable |
|---|---|---|---|---|---|---|
| Agent evaluation | Verification is required; REP defines quality metadata | Which measures predict useful completion; evaluator validity | Public benchmark tasks can be broken; final-state success misses process risk | Build repository-native eval suite | Run representative tasks across models/harnesses and repeat seeds | Eval specification, harness, baseline report |
| Long-horizon reliability | Durable artifacts and bounded cycles are preferred | Failure probability by horizon; recovery value | Capability horizons rise, but realistic tasks remain difficult | Model task decomposition and recovery | Vary task length, checkpoints, and intervention policy | Reliability curves and failure taxonomy |
| Context engineering | Context should be compiled, compressed, and budgeted | Optimal retention/summarization policy | Full context can underperform selective context in some workflows | Test policies on identical tasks | Full history vs recent window vs structured summary vs retrieval | Context policy decision framework |
| Memory | Working/project/retrieval/organizational layers are described | What deserves persistence; stale-memory detection; update policy | More memory may introduce obsolete state | Develop memory lifecycle tests | Inject stale/conflicting memories and measure recovery | Memory schema and freshness protocol |
| Agent orchestration | Coordinator, roles, contracts, and artifact communication are described | When extra agents add net value | Parallel agents add coverage but coordination can dominate | Quantify orchestration economics | Single vs parallel independent vs hierarchical execution | Orchestration threshold framework |
| Human-agent collaboration | Humans should retain consequential control | Best intervention timing and review granularity | Autonomy improves throughput; review capacity becomes bottleneck | Optimize escalation | Compare fixed gates, risk-triggered gates, and end review | Human effort/quality frontier |
| Permissions and security | Controlled delegation is a candidate principle | Usable action tiers; prompt-injection resilience | More controls may slow benign work | Risk-based capability design | Adversarial tool outputs under tiered permissions | Threat model and permission matrix |
| Observability | Metrics and durable artifacts are recommended | Minimum trace needed for diagnosis without excessive cost | Full traces aid diagnosis but inflate storage/context | Define diagnostic trace schema | Remove trace fields and measure diagnosis accuracy | Observability standard |
| Research evidence | Primary sources, contradictions, and provenance are required | Operational evidence quality scoring; citation drift | First-party data is current but biased | Calibrate source grading | Blind reviewers score claims with/without provenance | Evidence rubric and registry |
| Retrieval/RAG | Hybrid search and canonical metadata are proposed | Retrieval quality on actual repository questions | Large context and retrieval each fail differently | Create retrieval test collection | Compare lexical, vector, hybrid, and curated context | Retrieval benchmark |
| Knowledge graphs | Proposed for concept relationships | Whether graph maintenance produces enough value now | Graphs improve navigation but can create curation debt | Delay implementation pending task evidence | Compare graph-assisted vs hybrid-search answers | Go/no-go decision |
| MCP/tool protocols | Protocols reduce custom integration | Trust, lifecycle, schema drift, authorization ergonomics | Standard transport does not ensure safe or correct tools | Validate integration contracts | Mutate schemas, permissions, and tool outputs | MCP readiness checklist |
| Coding agents | Repository work is a primary target | Generalization from benchmarks to this repo | Benchmark quality and harness effects distort ranks | Use real issue history as tasks | Replay held-out repository changes | Coding-agent baseline |
| Browser/computer agents | Important future interface | Reliability under dynamic UI, auth, and injection | Strong demos coexist with fragile state handling | Build bounded UI tasks | Repeat tasks under layout/content perturbations | UI agent risk profile |
| Model routing | Cost-constrained computation is a candidate principle | Routing thresholds by task/risk | Frontier models cost more but may reduce retries | Optimize total verified cost | Route same suite across model tiers | Routing policy |
| Local vs cloud models | Listed as research area | Privacy/performance/cost crossover | Hardware and model releases change quickly | Monitor, do not architect early | Quarterly representative benchmark | Deployment scorecard |
| Multimodal reasoning | Listed but undeveloped | Value for repository/document workflows | Rapid improvement, weak internal use case definition | Develop use cases before broad study | Diagram/PDF/UI tasks with observable grading | Multimodal task set |
| Synthetic data/RL | Recognized as technology | Whether AI-ROS has enough task/eval data to benefit | Optimization can overfit flawed graders | Defer until eval foundation exists | Adversarial holdout after sufficient traces | Training-data readiness gate |
| Explainability | Constitution values traceability | Which explanations help debugging and trust | Verbal rationale may be unfaithful | Focus on observable execution first | Compare trace/state evidence with generated explanations | Explainability evidence note |
| Economics | Token economics is documented | Cost per verified outcome and human oversight cost | Token cost alone ignores retries and review | Define full cost accounting | Capture compute, latency, retries, and review minutes | Verified-outcome cost model |
Missing Infrastructure
The following canonical systems described by the research program do not yet exist:
- append-only scientific journal;
- evidence registry;
- hypothesis registry;
- theory registry and evolution log;
- experiment registry;
- open-question registry;
- metrics registry;
- automated reference and identifier validation.
Creating all registries now would be speculative. The next mission should instantiate only the minimum schemas needed for the evaluation experiment, then revise them from observed use.
Cross-Disciplinary Gaps
- Measurement science: construct validity, inter-rater agreement, measurement error
- Software testing: mutation testing, property-based testing, coverage adequacy
- Human factors: automation bias, review fatigue, escalation design
- Security engineering: capability control, identity, secret handling, prompt injection
- Economics: cost per verified outcome, option value of reversibility
- Operations research: scheduling, queues, parallelism, resource allocation
- Cognitive science: working-memory limits and external representations
- Knowledge management: provenance, freshness, canonicalization
- Safety engineering: hazard analysis and defense in depth
Largest Knowledge Gap
AI-ROS cannot yet distinguish a system that produces persuasive artifacts from one that reliably produces correct, traceable, safe, and economical outcomes. Until that distinction is operationalized, theory updates and automation decisions remain vulnerable to evaluator error.