Rules that say when they apply

C:\> type briefing\2026-10-09.md

---
title: Rules that say when they apply
date: 2026-10-09
summary: Two preprints on rules that carry their own conditions and expiry, OpenAI's
  contracting evaluation with Ironclad, and a practitioner case for agreed business
  definitions.
pick: 'Items 1 and 2 together, as one small demonstration: ten exception cards plus
  a 12-case evaluation. That could produce evidence of whether structured context
  actually improves decisions.'
sources:
- title: 'SAGA: agent evolution via knowledge abstraction'
  url: https://arxiv.org/abs/2610.06964
  source: arXiv
  kind: Research preprint
  published: 2026-10-03
  what_changed: SAGA turns an agent's experience into a hierarchy of examples, procedures
    and reusable principles. Each principle says when it applies and keeps a link
    to the evidence behind it. The authors report better performance in ScienceWorld
    and ALFWorld.
  limits: Both are simulated environments, not customer operations.
  why_it_matters: It's a concrete research model for turning staff experience into
    reusable context without losing the circumstances it came from.
  idea_to_test: 'Take ten customer-service exceptions and write a card for each: the
    example, the rule, when it applies, its exceptions and the evidence. Compare AI
    answers that use the cards with answers that use the original notes.'
- title: OpenAI's contracting research with Ironclad
  url: https://openai.com/index/advancing-computer-use-with-ironclad/
  source: OpenAI
  kind: First-party research report
  published: 2026-10-06
  what_changed: Domain experts from Ironclad helped define 11 contracting tasks, covering
    approval routing, document configuration and reusable clauses, each graded against
    8 to 50 criteria. GPT-6 Astra averaged 55.0% against 41.6% for GPT-5.6 Sol.
  limits: A small internal evaluation. The two models ran at different reasoning settings,
    and the reported time savings are simulated estimates, not measured customer savings.
  why_it_matters: The evaluation checks whether business requirements survive a whole
    workflow, exceptions included. For context design, that suggests a written context
    brief is worth more when it comes with explicit acceptance criteria.
  idea_to_test: 'Build a 12-case evaluation for one business process: normal cases,
    cases right at a threshold, and exceptions. Score whether each required rule was
    followed.'
- title: 'From memory to guide: procedural coding memory'
  url: https://arxiv.org/abs/2610.04868
  source: arXiv
  kind: Research preprint
  published: 2026-10-04
  what_changed: The proposed Composer adapts stored procedures to the current workspace
    and the current stage of the work, including when a piece of guidance should expire.
    It targets the risk of applying a useful procedure where local conditions don't
    fit. On 13 long-horizon coding tasks, the authors report a pass rate 7.2 points
    higher on the hardest tasks, with the main agent using 32% fewer tokens.
  limits: The work is about coding agents. Whether it transfers to other business
    workflows is unproven.
  why_it_matters: Reusable knowledge needs a defined scope. A campaign rule may depend
    on stock, channel, audience or season, and retrieving it doesn't establish that
    it still applies.
  idea_to_test: 'Add three fields to one campaign''s context brief: applies when,
    overridden by, and expires when. Test it against a change in stock and a different
    channel.'
- title: The next enterprise AI advantage is better business context
  source: Express Computer
  kind: Practitioner analysis
  published: 2026-10-08
  what_changed: A practical account of governed definitions, machine-readable rules
    and keeping context maintained. It describes how answers go wrong when what a
    term means in the business changes, even though the data pipelines still work.
  limits: The examples are the author's own, with no independent evaluation.
  why_it_matters: 'It gives a specific place to start a diagnosis: the questions whose
    answers depend on disputed definitions or unwritten conventions.'
  idea_to_test: Ask three people on the team to define five terms, such as "active
    customer", "urgent order" and "available stock". Turn the disagreements into written
    conditions, each with a named owner.
---

C:\>