Central-Bank Rate Analysis Kit

A working, agentic Codex environment for safe, AI-assisted data analysis. It takes a data question from raw file to a verified, sourced, management-ready HTML presentation — cleaning data, comparing figures, visualizing them, and writing an executive summary, all under a Safe-Zone discipline with independent verification and human-in-the-loop checkpoints.

It is a teaching environment, built on the agentic-context-engineering method (the four mechanisms — AGENTS.md, standards, skills, subagents — plus stable IDs, registries, business-logic templates, and a two-round QA system).

The flagship task

Compare the Central Bank of Armenia (CBA) refinancing rate to peer central banks:

1. Bring the local CBA rate file into the Safe Zone and clean it.
2. Fetch peer policy rates (Fed, ECB, BoE, …) from the open web — each cited.
3. Combine and analyze — spread, direction, volatility.
4. Visualize as charts a manager reads in ten seconds.
5. Assemble and verify a self-contained HTML presentation.

Setup

1. Open Codex with this folder as the working directory, and trust it when prompted. The kit is self-contained: AGENTS.md is the root switchboard, .agents/skills/ holds the skills, .codex/ holds the subagents and the project config, standards/ holds the six standards, and the content folders hold the data and outputs.

   Trust matters: Codex loads project-scoped .codex/ layers — the subagents, the sandbox setting and the Safe-Zone hook — only for a trusted project. Untrusted, the skills and AGENTS.md still work; the subagents and the hook do not.

2. Python 3 is used for data work. No install is needed for CSVs (standard library). If you point the kit at an Excel file (.xlsx), it will create a disposable .venv and install pandas/openpyxl automatically — then throw it away. If that install can't run, it falls back to CSV and tells you.

3. Web access is on, and it had to be turned on. sandbox_mode = "workspace-write" disables the network by default, so .codex/config.toml sets network_access = true under [sandbox_workspace_write]. Without it $fetch-peer-rates and $web-research cannot run at all, and the .xlsx path cannot pip install.

   That makes egress a sandbox-level permission, and the Safe-Zone classification the only thing standing in front of it. A PreToolUse hook prints a reminder before each web call — its matcher is a regex over the tool name, so confirm the web tool's name in your Codex build and narrow the pattern before you rely on it. The hook only prints; it cannot block.

Run it

The whole flagship, end to end:

$workflow "Compare the CBA refinancing rate to peer central banks for a board briefing"

…or step by step (each step is a skill; each producing step is checked by a QA skill before it's accepted):

$intake data/cba-refinance-rate.csv     # classify (Safe Zone) + register  → source-qa
$clean-data DS-001                        # fix missing/dupes/outliers/dates → clean-qa
$fetch-peer-rates                         # peer rates from the web, cited   → source-qa
$analyze "CBA vs peers"                   # spread / direction / volatility  → analysis-qa
$visualize FND-001                        # self-contained SVG charts        → presentation-qa
$presentation FND-001                     # self-contained HTML presentation → presentation-qa
$qa-check presentations/RPT-001-*.html    # independent re-check on demand

Separately, when a question needs fresh information from the open web rather than a number for the pipeline:

$web-research "How have peer central banks guided on 2026 rate cuts?"
                                          # sourced brief WRB-NNN     → web-research-qa

How it stays safe and honest (the three non-negotiables)

1. Classify before you use or send. Every dataset is public / internal / restricted. Restricted data never reaches a web tool or an external prompt. The shipped data/internal/loan-portfolio-CONFIDENTIAL.csv is a restricted example to practice this. → standards/safe-zone.md
2. No unsourced figures. Every number traces to a dataset cell, a logged computation, or an official source in sources/source-registry.md. "Not found" is an acceptable answer; a plausible guess is not. → standards/evidence-and-figures.md
3. Compute with throwaway scripts, verify independently. Analysis runs as a disposable Python script (deleted after), then a separate verification script recomputes and checks it (also deleted). Only the data, charts, HTML, and a markdown record of the method + verdict survive. → standards/ephemeral-compute.md

What's in here

Path	What it is
AGENTS.md	Root switchboard — read the map here
context/analysis-brief.md	The anchor — the decision, vocabulary, Safe-Zone posture
standards/	The six standards, as plain documents. Named by AGENTS.md, the folder notes, and each skill
.agents/skills/	Codex skills — producing skills + their QA skills. Invoke with $name
.codex/agents/	data-steward, analyst, presenter, web-research — one TOML each
.codex/config.toml	Sandbox, approval policy, and the Safe-Zone egress hook
data/	Datasets + registry (shipped: a dirty CBA rate file + a peer-rates template + a restricted example)
web-research/	Open-web sources (WRS-NNN) + research briefs (WRB-NNN) + two registries + brief template
analysis/	Cleaning logs, combined data, findings + registry + finding template
presentations/	Charts + HTML presentations + registry + HTML template
sources/	Source registry (SRC-NNN) + evidence policy
scratch/	The only place throwaway scripts live — empty at rest
TRAINER-GUIDE.md	Maps the 8-meeting course onto the skills/exercises

The shipped sample data

• data/cba-refinance-rate.csv (and .xlsx) — a monthly CBA refinancing-rate series that is deliberately dirty: two missing values, one duplicate row, one out-of-range outlier (95.0 where 9.5 was meant), mixed date formats, and a stray percent sign. It is illustrative only (source tier P4) — refresh from the official CBA before publishing any figure.
• data/peer-rates.template.csv — the empty schema fetch-peer-rates fills with real, web-sourced peer rates. No fabricated numbers are shipped.
• data/internal/loan-portfolio-CONFIDENTIAL.csv — a small, obviously-fictional restricted file for the classification exercise. It must never go to a web tool.

A note on outputs

The kit ships empty of results (no findings, charts, or presentations yet) so each cohort produces its own. Every artifact you generate is registered with a stable ID in the relevant _index.md, verified, and QA'd before it's accepted. Presentations are self-contained HTML — open them in any browser offline, or open one in a browser for review.