nullorigin.yaml

# NullOrigin — Default Configuration
# Author: Muhammad Rakibul Islam <imrakibul@gmail.com>
# License: Apache-2.0
#
# Every value here can be overridden by a NULLORIGIN_* environment variable
# (see nullorigin/config.py) and then by an explicit CLI flag.

proxy:
  # Loopback by default: this proxy relays your upstream API keys. Binding a
  # public interface requires either a telemetry token or --allow-public-bind,
  # and should sit behind a TLS-terminating reverse proxy.
  host: "127.0.0.1"
  port: 8080

  # Guards /telemetry and /telemetry/reset. Generate with:
  #   openssl rand -hex 32
  # Leave empty only while bound to loopback.
  telemetry_token: ""

  # Reject request bodies larger than this many bytes (default 100 MiB).
  max_request_bytes: 104857600
  # Used when a request carries no x-nullorigin-provider header.
  default_provider: "openai"
  timeout_seconds: 120.0
  upstream_map:
    anthropic: "https://api.anthropic.com"
    openai: "https://api.openai.com"
    gemini: "https://generativelanguage.googleapis.com"

text:
  clean_unicode: true
  # Fold Cyrillic/Greek confusables to their Latin equivalents.
  fold_homoglyphs: true
  # Deltas buffered before a streaming span is handed to the rewriter.
  stream_window_tokens: 40

  # Keep fenced code blocks out of the rewriter. Paraphrasing code produces
  # code that does not run, and the z-score drops either way, so nothing
  # downstream would notice. Code inside a response still has invisible and
  # bidi characters stripped; it is never NFKC-normalized (that rewrites
  # string literal values) and never sent to the model.
  protect_code_blocks: true
  # Also shield inline `code` spans, which usually hold identifiers or commands.
  protect_inline_code: true

  paraphraser:
    # none | ollama | openai_compatible | llama_cpp | lexical
    #
    # Choosing "none" leaves the statistical watermark intact — only unicode
    # artifacts are removed. "lexical" needs no local model but is a much
    # weaker attack than a neural rewrite; `nullorigin benchmark` measures the
    # difference rather than assuming it.
    backend: "ollama"
    endpoint: "http://localhost:11434"
    model: "llama3.2:3b"
    # Used only by the llama_cpp backend.
    model_path: ""
    # Used only by the openai_compatible backend.
    api_key: ""
    temperature: 0.75
    top_p: 0.9
    timeout_seconds: 120.0

  # Gate on whether a rewrite preserved meaning. A z-score drop is not evidence
  # of preserved meaning: a fluent rewrite that changes a fact scores exactly
  # like a faithful one. Failed checks retry at lower temperature and finally
  # fall back to the original, because measuring drift after the fact does not
  # help if the damaged text has already been returned.
  fidelity:
    enabled: true
    # Retries at progressively lower temperature; drift is temperature-driven.
    max_retries: 2
    temperature_step: 0.25
    # Bidirectional entailment catches role swaps that invariant counting
    # cannot. Requires nullorigin[nli]; without it, invariants still gate and
    # the limitation is reported rather than hidden.
    use_nli: true
    nli_threshold: 0.5
    nli_model: "cross-encoder/nli-deberta-v3-small"

  kgw:
    gamma: 0.5
    hash_key: 15485863
    # z > 4.0 corresponds to p < 3e-5 (Kirchenbauer et al.).
    detection_threshold: 4.0

media:
  purge_metadata: true
  disrupt_stego: true
  crop_ratio: 0.005
  # trim     — shift the coordinate origin without resampling (default;
  #            retained pixels stay bit-exact, output is ~1% smaller)
  # resample — restore exact original dimensions, costing roughly
  #            SSIM 0.81 / PSNR 31 dB even at a 0.5% crop
  # none     — no geometric change
  crop_mode: "trim"
  dwt_threshold: 0.8
  fourier_strength: 0.05
  dither_sigma: 0.6

audio:
  disrupt_watermarks: true
  # Phase below this frequency is preserved to protect intelligibility.
  low_cut_hz: 800.0
  max_shift_rad: 0.25
  notch_bands: 3