Configuration and selected limits

Configuration lives in config/ and is validated strictly. Unknown keys are rejected at the top level and inside every nested object, so a typo such as default_sensitvity fails to load rather than being silently ignored.

The five main vault configuration files reject YAML anchors, aliases, explicit type tags, and tabs. config/clean-room.yaml uses a separate parser and does not apply that blanket prohibition.

config/ziggurat.yaml

schema_version: 1
lifecycle:
  review_queue_limit: 50

Key	Type	Notes
schema_version	literal 1	Required
lifecycle.review_queue_limit	integer, minimum 1	Required

The repository configuration uses 50. A fresh vault created by ziggurat init starts at 20. The configured value limits entries per rendered review page, not the backlog, validation scan, contradiction enforcement, or proposal retention. Packets report the render limit separately from total/displayed/remaining counts, page positions, and the whole backlog’s oldest age. Use review --order oldest for global oldest-first ordering and review --cursor <next_cursor> to continue. Changing this limit invalidates existing cursors; restart without a cursor. See review navigation. Nothing is automatically dismissed, expired, archived, deleted, or resolved.

config/domain.yaml

domain:
  page_types:
    - entity
    - concept
    - comparison
    - query
    - decision
    - reference
  tags:
    - ai
    - architecture
    - infrastructure
    - product
    - personal
    - workflow
    - security
    - performance

page_types and tags are non-empty lists of non-empty strings.

config/privacy.yaml

privacy:
  default_sensitivity: restricted
  default_pii: unknown

Both values are literals. Fresh captures are restricted and PII-unknown, and stay out of the model-readable evidence index until a human resolves them.

config/adapters.yaml

adapters: {}

Key	Type	Notes
adapters.model_endpoint	URL, optional	Complete HTTP loopback chat-completions URL
adapters.model_name	string, 1-128 characters, optional	Model alias sent in chat completions; defaults to ziggurat-refine

Retrieval is lexical (BM25) over the Gold index. No embeddings are computed and no vector index exists.

Only http: on localhost, 127.0.0.1, or [::1] is accepted. Any other scheme or host fails configuration loading.

The supported protocol is llama.cpp non-streaming POST /v1/chat/completions. For example:

adapters:
  model_endpoint: http://127.0.0.1:18080/v1/chat/completions
  model_name: ziggurat-refine

The adapter uses the complete URL as configured; it does not append the API path. The adapter sends a Zod-derived draft schema in the verified b10809 nested format:

response_format: {
  type: "json_schema",
  json_schema: {
    name: "ziggurat_refinement_draft",
    strict: true,
    schema: RefinementDraftJsonSchema
  }
}

A bare sibling schema field is not supported; a real b10809 probe silently ignored it and returned an invalid draft. Requests use max_tokens: 2048, temperature: 0, and stream: false. There is no protocol fallback, JSON repair, retry, or tool-call path. See the local model protocol and setup guide.

config/trust.yaml

trust:
  reviewers: []

Each reviewer entry carries reviewer_id, key_id, algorithm: ed25519, and public_key_pem. Only public keys belong here.

config/clean-room.yaml

Only project_names and exclude_paths are allowed. Missing or null values mean empty additional lists; otherwise each must be a list of non-empty strings. An absent, empty, comments-only, or YAML-null document uses built-in defaults. Other non-mapping documents, malformed YAML, unknown keys, unreadable files, and invalid list entries fail ziggurat check.

Because the scan covers this file too, a configured project name is also a finding in it. Add config/clean-room.yaml to exclude_paths when you configure names, and re-read the file by hand before every release.

Enforced limits

Limit	Value	Applies to
Records per refine reference block	12	refine
Canonical Bronze body bytes per source	32 KiB	refine
Combined canonical Bronze body bytes	256 KiB	refine
Adapter request deadline	30 seconds	refine model endpoint
Request body ceiling	1 MiB	refine model endpoint
Response body ceiling	1 MiB, enforced while reading response bytes	refine model endpoint
Completion token limit	2,048	llama.cpp refinement request
Query length	1,024 UTF-16 code units	query, search_context
Results per search	20	query, search_context
Retained citations per session	200	MCP session
Search excerpt length	500 UTF-16 code units	search_context
CLI JSON excerpt length	300 UTF-16 code units	query --json
Gold verification age	90 days	build, query, MCP

The Bronze-body limits do not include JSON encoding, metadata, prompts, or optional target context. Those are additional request bytes, subject to the separate 1 MiB request ceiling.

The loopback restriction applies to the configured model endpoint. The deadline, redirect refusal, and byte ceilings are enforced by the refine adapter, which is the only shipped code that issues a request. Reading response bytes incrementally enforces the byte ceiling; it does not enable streaming chat completions.

Oversize refine records are omitted rather than truncated, and every omission is reported with a reason. Over-long queries are refused rather than truncated. Reaching the citation ceiling evicts the oldest IDs, which then become invalid.

Related

• CLI reference
• Gold MCP