Skip to main content
When a research run has a paper-positive row and evidence context, the Enoch control plane can write and package generated paper artifacts. You control which provider performs that write through the paper_writer_provider field. The deterministic mode produces a fixed bounded template with no external dependencies, while synthetic.new connects to a hosted OpenAI-compatible API using the configured model. Treat provider-backed writing as an automated packaging/finalization path, not a publication guarantee or a substitute for independent validation. If claim extraction is unavailable, the writer must preserve existing ledgers or write a blocked empty ledger that cannot pass strict claim/evidence audit.

​
Providers

​
deterministic

The default provider. It produces template-based output without calling an external LLM. Use it for local testing and predictable behavior.

​
synthetic.new

Optional provider-backed writing path. Enoch sends paper context to the Synthetic.new API endpoint using the OpenAI-compatible chat-completion interface, using GLM-5.1 (hf:zai-org/GLM-5.1) as the default model in config.example.json. The response becomes the paper body. If the provider request fails and paper_writer_fallback_enabled is true, Enoch falls back to the deterministic template rather than failing the paper-write step.

​
Fields

  • paper_writer_provider
  • paper_writer_base_url
  • paper_writer_model
  • paper_writer_api_key
  • paper_writer_timeout_sec
  • paper_writer_temperature
  • paper_writer_max_tokens
  • paper_writer_fallback_enabled
If paper_writer_api_key is empty, provider code can look for SYNTHETIC_API_KEY. Do not commit API keys.

​
Fallback behavior

When fallback is enabled, provider failure can fall back to deterministic output. That is operationally useful, but the resulting artifact still needs explicit packaging/provenance and strict claim/evidence audit status before it is release-ready.

​
Publication rule

Paper-positive artifacts are intended to be auto-pushed as written after deterministic gates pass: evidence bundle, claim ledger, manifest, packaging/provenance lint, strict claim/evidence audit, corpus import ledger, and generated public indexes/manifests. Do not add a normal human editorial approval or taste-review stage between writer output and corpus push. If output is blocked, block on the concrete failed gate.
​
paper_writer_timeout_sec
number
default:"180"
Maximum seconds to wait for a response from the provider before treating the request as failed. Minimum value is 10. Increase this for very long papers or slow network links.
​
paper_writer_temperature
number
default:"0.2"
Sampling temperature passed to the model. Range 0.0–2.0. Lower values produce more deterministic output; the default 0.2 is appropriate for structured research writing.
​
paper_writer_max_tokens
number
default:"12000"
Maximum number of output tokens the model may generate per paper-write request. Minimum value is 512. Increase this if your papers are truncated.
​
paper_writer_fallback_enabled
boolean
default:"true"
When true and paper_writer_provider is synthetic.new, a failed provider request causes Enoch to fall back to the deterministic template instead of raising an error. Set to false if you want provider failures to surface explicitly.

​
Synthetic.new config

Add the following fields to your control VM’s config file to test provider-backed paper writing with Synthetic.new and GLM-5.1. Confirm corpus packaging/provenance scans and strict claim/evidence audit status before treating any output as release-ready: 
{
  "paper_writer_provider": "synthetic.new",
  "paper_writer_base_url": "https://api.synthetic.new/openai/v1",
  "paper_writer_model": "hf:zai-org/GLM-5.1",
  "paper_writer_api_key": "your-provider-key",
  "paper_writer_fallback_enabled": true,
  "paper_evidence_sync_enabled": true
}
Do not publish generated paper artifacts until corpus packaging/provenance lint and strict claim/evidence audit status are explicit. Once those deterministic gates pass, the artifact is pushed as written. A packaging/provenance pass is not a deep claim audit.