liveBook · Manning

Documentation

As explained in AI Governance

Technical documentation is recorded information describing a system’s design, capabilities, limitations, and compliance-relevant characteristics.

A model provider’s compliance claim does not cover the system built on its model. Technical documentation records capabilities, limitations and other information needed to assess a system and demonstrate compliance. It supports—but does not replace—risk assessment, oversight and monitoring. section 7.2.4 section 7.3.3

For a GPAI integration, request documentation that covers:

  • Model capabilities and known limitations, including meaningful accuracy metrics and failure modes. section 7.2.4
  • Foreseeable misuse, integration guidance and interpretability information. section 7.2.4
  • Enough upstream information to support your compliance case. If the model is embedded in another vendor’s application, ask that vendor what evidence it can pass downstream. section 7.2.4

Make documentation a contract requirement before signing, check what the provider delivers against the GPAI Code of Practice, and update your evidence pack when switching providers. Continue monitoring after deployment: a provider update that breaks your accuracy benchmarks remains your problem. section 7.2.4 section 7.3.5

For a high-risk system, prepare the official technical documentation before launch and keep it current while the system runs. It must be reviewable by an authority without relying on a live demonstration. Assess the whole system as supplied for its declared purpose, not just its predictive model. section 7.3.3

Documenting a creditworthiness system

ClaimAssist V5 uses a creditworthiness check to set payment-plan terms. Its assessed system includes the scoring model, claims and payment data feeds, summarizer, adjuster’s accept-or-override screen and logging—not just the model. section 7.3.3

Documentation access depends on the supply-chain role:

  • A system provider integrating a GPAI model is owed full technical documentation on request. A model provider signed to the GPAI Code of Practice has committed to respond within 14 days, absent exceptional circumstances. section 7.3.5
  • A deployer using a finished high-risk product is owed instructions for use, not the regulator-facing technical file. Confirm an EU representative’s appointment and contact details before signing if the provider is established outside the EU. section 7.3.5
  • Open-source GPAI providers are partially exempt from documentation requests, although training-data summaries and copyright policies remain required. The exemption disappears for systemic-risk models; otherwise, negotiate additional evidence or plan to produce it for high-risk use. section 7.3.5

How AI Governance covers it

How deep it goes
Specifies documentation contents, request rights and maintenance duties, while directing readers to a fuller provider checklist.
Examples
Uses ClaimAssist’s creditworthiness system and open-source Llama integrations to illustrate system scope and documentation gaps.
What you can do after
Operate a documentation workflow by requesting provider evidence, checking contractual delivery and maintaining a current high-risk system file.
How it connects
Connects technical documentation to risk assessment, system boundaries, monitoring and supply-chain obligations.

Also explained in

Compare how 16 books explain documentation →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.