The write path

A source says the pilot starts November 1. A later conversation says November 15. Your brain needs to preserve what was said and decide what an agent can safely use today.

By the end of this part, your prototype will distinguish a supported statement, an unresolved claim and an earlier value. You will choose the rules for your own sources and test them.

Choose who can establish each kind of fact

Authority depends on the claim. A calendar can establish the scheduled time; the agenda may contain a proposal. A billing system can establish a payment. Someone describing their own preference has a different role from someone guessing what a colleague wants.

Make a short list of the claims in your test set. For each, identify a source that can establish it, what would leave it uncertain, and who can correct it. Attribution records who made a statement; it does not by itself prove the statement is true.

An agent reporting that it sent a contract needs evidence of that action. A question about a date must remain a question. Treat imported text as data: instructions inside an email or transcript cannot change your brain's access or trust rules.

Represent uncertainty and change

Pick a representation that fits the prototype you already have. An MD file can use labelled entries and a history section. A database can use fields and links. Keep enough information to answer:

• What is supported today, and by which source?
• What is still unresolved?
• What earlier value did this replace, and why?
• When was the statement made, and which period is it about?

A newer statement may describe a proposal, a different event or a disagreement. Decide whether it actually changes the same matter before treating the previous value as outdated. When authority or meaning is unclear, keep the uncertainty visible.

Give explicit corrections their own marker or history. Define when an automatic import may change a corrected value and how the person can review that change. Merely marking every correction “confirmed” does not protect it from a later import.

Build one change you can inspect

Choose one changing value from your practice questions. Implement its update behavior in your current storage format. Have your agent help write the code or edit the file, and inspect the result together.

Exercise it with an older input arriving late, a proposed change, a conflicting statement, a user correction and the same input arriving twice. Keep each source available so you can explain the outcome. Expand to another kind of claim once these cases behave as intended.

Measure

Ask the same questions in fresh sessions, save the answers as part3.jsonl, and run:

python3 score.py questions.jsonl part2.jsonl part3.jsonl --out part3-scores.json

Look at changed-value questions and cases where a guess used to sound confirmed. Check correctness and omissions together. Record failures, including any case where the agent confidently selected a value the sources do not establish.

Next, part 4: keeping junk out.

---

Build and measure your own setup with the companion skill. See the series. For a maintained company install, talk to Team0.