Citadel

Citadel applies a YAML de-identification policy to JSON or JSONL records and writes the transformed data to an explicit output path.

uvx --from tacit-citadel citadel policy.yaml input.json output.json

For local development from this checkout:

uv run citadel policy.yaml sample.jsonl sample.citadel.jsonl

The PyPI package is tacit-citadel; the installed console command is citadel. A bare uvx citadel resolves the unrelated citadel package name on PyPI.

Project-specific policies can live with the dataset they are used for. The checked-in policy.yaml and sample.jsonl are generic support-ticket examples for local smoke tests.

Policy Shape

Policies have deterministic path-based rules and an optional whole-record rewrite step.

version: 1
name: support-ticket-sanitizer
description: De-identification policy for customer support tickets.

rewrite:
  backend: codex exec
  sandbox: read-only
  system_prompt: You are a conservative de-identification rewriter.
  user_prompt: |
    Rewrite the INPUT JSON object.

    INPUT JSON
    {{content}}

rules:
  - path: .account.id
    action: drop

rewrite.backend must be codex exec or claude -p. Rewrites always receive the whole record after deterministic rules have run and must return a complete JSON object. If a rewrite fails or returns invalid JSON after three attempts, Citadel logs an error, aborts the run, and does not write the output file.

Rules

path is a small jq-like selector. It supports dotted object fields, list wildcards, numeric list indexes, quoted bracket fields, and simple | select(.field == "value") / | select(.field != "value") filters.

required defaults to true. Use required: false for sparse paths.

drop

Deletes matched object fields.

- path: .account.id
  action: drop

fuzz_number

Perturbs numeric values with either percentage or range mode.

- path: .customer.contract_value_usd
  action: fuzz_number
  params:
    mode: percent
    max_percent: 3
    precision: 0

- path: .risk_score
  action: fuzz_number
  params:
    mode: range
    min_delta: -1
    max_delta: 1
    step: 1

date_offset

Replaces date or datetime strings with a day offset from an anchor date.

- path: .events[].timestamp
  action: date_offset
  required: false
  params:
    anchor_path: .reported_at
    output: human_relative

Outputs are same day, N day later, N days later, N day ago, or N days ago.

Inputs

Citadel accepts:

• a JSON object
• a JSON array of objects
• JSONL with one object per non-empty line

The output format matches the input shape.

Development

The implementation is intentionally contained in run.py.

uv run pytest
uv run ruff check .
uv run ty check