How we audit: the RXed AI Periodic Table v1.0
Every AI tool we audit is scored against the same 20 elements, organized like chemistry organizes the elements: 4 rows (from primitives to emerging capabilities) × 5 families. The concept builds on the AI Periodic Table by IBM Technology (Martin Keen) — credit where due. RXed extended it: we filled the three cells IBM left open and made one swap, so the table is complete and audit-ready.
The rows
- R1 — Primitives
- R2 — Compositions
- R3 — Deployment
- R4 — Emerging
What RXed changed vs IBM's original — and why
+ Cx Context — context engineering became the core discipline
of applied AI in 2025–26; the context window is the atomic unit of orchestration.
+ Tr Tracing — LLM observability is now a production category of its own (OpenTelemetry GenAI conventions,
trace-based debugging); without traces there is no higher-level validation.
+ Pc Protocols — fills the emerging-orchestration cell IBM explicitly left open: MCP (agent-to-tool) and
A2A (agent-to-agent) became the standard stack under Linux Foundation governance.
Rt → Ev Evaluations — red-teaming is one facet of the broader evaluations discipline, so the element was
generalized; every serious 2026 AI stack ships evals.
How scoring works — deterministic, always
Each element gets 0–10 (steps of 0.5) with a written reason, or N/A with a written reason when the capability is outside the tool's stated scope. "Missing but expected" is a low score, not an N/A. The overall score is the plain mean of applicable elements, rounded to 0.1 — computed by the site generator, never set by hand, identical for every tool. The coverage badge (e.g. 17/20) shows stack breadth separately, so a brilliant narrow tool isn't punished and a shallow platform can't hide. Bands: ≥8.5 benchmark · 7.0–8.4 strong · 5.0–6.9 capable · <5.0 weak.
Tiers
- Frontier lab
- Enterprise platform
- SMB tool
- Specialist
- Open source
All 20 elements
Instruction design surface: system prompts, prompt tooling, templates, caching.
Scored when: Does the tool let users or developers shape instructions to the model?
Numerical representations of meaning powering semantic search and memory.
Scored when: Does the vendor ship or expose an embedding capability (API, semantic search feature)?
The context window and everything engineered into it: window size, file/knowledge attachment, context caching, context management. RXed addition v1.0.
Scored when: Does the tool manage what the model sees (files, history, knowledge, window)?
Logging of runs, decision paths and data lineage: run logs, trace trees, audit trails, cost attribution. RXed addition v1.0.
Scored when: Can an operator inspect what the AI did and why (logs, traces, history)?
The base language model(s): quality, choice, freshness. Third-party model access scores on quality-of-access, not ownership.
Scored when: Is an LLM at the core of the product?
The model invoking tools and APIs: breadth, reliability, schema strictness.
Scored when: Can the AI call external tools, actions or APIs?
Storage optimized for semantic search over embeddings.
Scored when: Does the vendor ship or host a vector/semantic store users can rely on?
Retrieval-augmented generation: grounding answers in user or org documents.
Scored when: Can the tool ground outputs in the user's own data?
Runtime safety: content filters, PII detection, policy enforcement, human-in-the-loop gates.
Scored when: Does the tool ship runtime safety or approval controls?
Native handling of images, audio, video alongside text.
Scored when: Does the product process or produce non-text modalities natively?
Think-act-observe loops working toward goals: autonomy, reliability, integration depth.
Scored when: Does the product ship agentic behavior (multi-step autonomous work)?
Adapting models on user/domain data: SFT, preference tuning, distillation.
Scored when: Does the vendor offer model adaptation to customers?
The scaffolding for building and running AI systems: SDKs, builders, orchestration logic, hooks. Renamed from IBM's Framework in v1.0 — Agent = Model + Harness.
Scored when: Does the vendor ship a builder, SDK or harness for composing AI systems?
Measuring quality: eval suites, benchmarks, regression testing, red-teaming (folded in from IBM's Rt in v1.0).
Scored when: Can users measure or test output quality systematically?
Efficient distilled/small model options: cost tiers, edge deployment.
Scored when: Does the vendor offer smaller/cheaper model tiers as a product choice?
Multiple specialized agents coordinating: handoffs, delegation, debate.
Scored when: Does the product support multiple agents working together?
AI-generated training data as a product capability.
Scored when: Does the vendor ship synthetic data generation for customers?
Open interoperability standards: MCP (agent-to-tool), A2A (agent-to-agent). RXed addition v1.0 — fills IBM's declared empty cell.
Scored when: Does the tool speak open agent protocols (MCP server/client, A2A)?
Understanding why the model does what it does: feature attribution, transparency reports exposed to users.
Scored when: Does the vendor expose interpretability surfaces to customers (not just research papers)?
Reasoning models with built-in chain-of-thought / test-time compute.
Scored when: Does the product offer reasoning-class models?
Table changelog
- v1.0 (2026-07-05): Initial frozen version. IBM's 17 elements + Cx Context (R1/G3), Tr Tracing (R1/G4), Pc Protocols (R4/G3). Swap: Rt Red-teaming folded into Ev Evaluations (R3/G4). Fw renamed Frameworks & Harnesses.
Audits are independent tool reviews, fully sourced on each passport, never legal or compliance advice. No affiliate links, ever.