Precedent

Knowledge that cannot be derived

The case for a corpus rests on the exceptions no amount of reasoning can settle.

The first time it meets a counterparty

nothing in the corpus about them

A payment arrives short of the invoice by a proportion matching no statutory band, with no refund and no fee explaining the gap. The agent can measure the shortfall precisely and still cannot say why it exists.

Escalating is correct here. The evidence is genuinely insufficient, and a guess would close an invoice that was never settled.

After a reviewer has resolved one

one precedent about that counterparty, written from that decision

The same arithmetic now matches something the system has seen. The precedent carries the arrangement — a negotiated rebate, an advance already adjusted — which no tool could have computed from the case.

This is the whole mechanism. Not a better prompt: a fact the system did not previously have.

the corpus is the difference between those two columns

What that is worth, measured

Held-out cases replayed against the corpus at two sizes. The left-hand figure is every exception; the right-hand one is the subset that can only be answered from a precedent.

70.0% → 86.7% all exceptions, corpus 42 → 151
0.0% → 83.3% the only-a-precedent cases below
56.7% → 61.7% random control — flat, which is what rules out “more text in the prompt”

What it has actually remembered

No counterparty precedents in the corpus yet.