SLAF Framework

A single large generation step fails in ways that are hard to detect: part of the output is correct, part is invented, and nothing distinguishes them. Smaller units with narrow responsibilities make each result cheap to validate, and localise a failure to one block instead of the whole document.

A structured block is the atomic unit of generation. Each block produces one specific section of output and declares:

• Explicit data dependencies — what the block needs, so missing inputs are caught before generation rather than papered over during it.
• An output schema — the shape the result must conform to, so a malformed result is a validation error instead of a downstream surprise.
• A veracity check — how the result is verified against the source material it was drawn from.

Blocks do not require you to hand-assemble their inputs. Required values are extracted from the material already attached to a project — uploaded documents and chat history — and only what genuinely cannot be found is escalated.

An SOP states its completeness criteria: the inputs that must be present before it can produce trustworthy output. When something required is missing, the engine asks for it rather than guessing, and the answer feeds back into extraction. This is what keeps a partially-specified request from silently becoming a confidently-wrong result.

Schema conformance proves the output has the right shape, not that it is true. Veracity checks close that gap by validating generated content against the documents and data it claims to derive from, so unsupported assertions are surfaced instead of shipped.

SLAF is what SOP definitions express. See SOP Engine for the concrete format, and Getting Started to run one.