Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add AnthonyAlcaraz/agentic-graph-rag-skills --skill four-layer-eval-cascadegit clone --depth 1 https://github.com/AnthonyAlcaraz/agentic-graph-rag-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/anthonyalcaraz/agentic-graph-rag-skills/four-layer-eval-cascade)<a href="https://agentmods.dev/skills/anthonyalcaraz/agentic-graph-rag-skills/four-layer-eval-cascade"><img src="https://agentmods.dev/badge/skills/anthonyalcaraz/agentic-graph-rag-skills/four-layer-eval-cascade/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/anthonyalcaraz/agentic-graph-rag-skills/four-layer-eval-cascade"><img src="https://agentmods.dev/badge/skills/anthonyalcaraz/agentic-graph-rag-skills/four-layer-eval-cascade.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00211 | $0.03270 |
| Opus 5 | $0.00105 | $0.01635 |
| Sonnet 5 | $0.00042 | $0.00654 |
| Haiku 4.5 | $0.00021 | $0.00327 |
Grade A, and why
four-layer-eval-cascade scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 10d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 191 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Four-Layer Evaluation Cascade
Overview
The execution graph tells you what happened. This framework is the diagnostic engine that tells you why. It operates as a sequential filter, moving from the most general failure cause to the most specific, and it stops at the first layer that catches the failure. Each layer asks a progressively narrower question (Ch7 Figure 7-1):
- Layer 0 (hallucination gate) gates on grounding. A lightweight NLI-style classifier scores whether the answer is grounded in the retrieved premise. The threshold is 0.85. It catches 60-70% of hallucinations at less than 5% of the compute of a full LLM judge call. Scores between 0.5 and 0.85 escalate to the full pipeline (the SLM-LLM flywheel); scores below 0.5 hard-block.
- Layer 1 (context evaluator) asks: did the agent even possess the
information it needed? A failure here is a knowledge-representation or
retrieval failure, not a reasoning failure. The verdict is binary
(sufficient or not) with a
missing_informationlist (J1-style). - Layer 2 (cognitive fault isolator) splits a cognitive failure into two mutually exclusive categories. A KNOWLEDGE failure means coherent reasoning over wrong facts (low Knowledge Index). A REASONING failure means the right facts, badly connected (near-zero or negative InfoGain steps).
- Layer 3 (TIR-Judge) verifies quantitative claims by executing code
against the KG. Its reward is
correctness * format * tool, multiplicative so a confidently wrong but well-formatted answer scores zero.
The output is a structured diagnostic report (the chapter's diagnostic-report example): it names the failure mode, locates it by node ID, and prescribes an intervention. That report is what drives the self-improvement engine.
The worked anchor is the DevOps premature-closure autopsy (the chapter's InfoGain-trace and
diagnostic-report examples): sufficient context, Knowledge Index 0.91, InfoGain trace
[0.34, 0.29, 0.22, 0.03, -0.01, 0.19], low-InfoGain steps [4, 5], fault at
CausalAttributionNode, recommended intervention PROMPT_REFINEMENT.
What ships with it
2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 10d ago First seen · 191 lines · 211 tokens per session scan A d5decb58552d
four-layer-eval-cascade is a skill published in the GitHub repository AnthonyAlcaraz/agentic-graph-rag-skills (10 stars, last pushed 2mo ago), licensed MIT. It adds 211 tokens to every session and 3,270 once invoked, about $0.0011 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
potpie-cli
Use when the task is centered on running, explaining, configuring, or troubleshooting the potpie command: doctor, login, pot management, source registration, search, graph workbench reads/writes, and pot scope behavior.
ios-simulator
Verify and debug native, React Native, Expo, or Flutter apps on an iOS Simulator with agent-device. Use when an agent needs to launch an app, inspect its live UI, tap, type, scroll, validate a code change, collect failure evidence, or reproduce a workflow on an iPhone or iPad Simulator.
cvm-ai-doctor
A health-diagnosis workflow for servers, computers, virtual machines, and containers on Linux, macOS, or Windows.
ai-discover
Parallel discovery of performance hotspots (perf track) and failure surfaces (bug track) for the auto-improvement loop. Fans out one subagent per hot-path area or failure surface; each returns ONE concrete, behavior-preserving fix candidate (perf) or a reproducing test plus fix (bug). Discovery only — no code changes…
debug
Debug issues in the MCP Gateway Registry using first-principles thinking. Invoke when something is broken, timing out, returning errors, or behaving unexpectedly. Forces structured root-cause analysis before any code change is proposed.
nw-root-why
Root cause analysis and debugging.