Decision memory
The part nobody can reconstruct later.
Research platforms can find the document. Work-product tools can produce the
memo. Neither has the answer a PM gave on a Tuesday in March, or the evidence that
was on the screen when they gave it. That only exists if something wrote down the
pre-outcome state before the outcome was known.
What was believed
The assumption, its state, and the date it was adopted, as it stood before anyone knew how it would resolve.
Evidence at the time
Every document read, with its verified identity and located spans.
A decision cannot later be replayed against evidence that did not exist when it
was made.
What LyraMind proposed
The agent's recommendation, its confidence and how that confidence
was computed, kept as its own record.
What the human decided
Accepted, rejected, deferred or dismissed, with the person, the
reason and the session. Kept apart from the proposal, always.
What actually happened
The reported figure against the stated condition. Whether 76 is
less than 90 is not a judgement, so the comparison is done in code, and it
refuses rather than guessing when the inputs are unusable.
What was learned
Which decisions rested on that assumption, what each concluded,
and what the firm systematically got wrong.
Those six things are only separable because they were never merged.
A single field saying “assumption weakened” answers none of the questions an
investigator, an allocator or a post-mortem actually asks.