pea·aiTry itRegistry
Open PEA

docs / 03-one-analysis.md

One analysis

One analysis, from the question to a draft a board can read.

The workspace

The workspace CHAT Analyst: the union matters more than the commission. Work that. ▸ searched: union leadership ▸ read 4 passages, audit report ▸ minted 2 claims ▸ nothing found: the 2016 act Agent: 2 claims added to section II. One gap in section III. DESIGN EVIDENCE ACTORS TIMELINE GAPS ··· DOCUMENT I. The problem 3 claims II. The players 8 actors · 14 claims III. The rules 1 gap ⚠ Why the 2016 act was never brought into force would need: a senior civil servant · interview IV. The state of play empty V. Where reform can happen empty
Chat is for steering. The work lives on the right, and empty sections say so.

Chat is on the left. The analyst steers there, and the agents report what they did. Every search and every read shows as one line, including the ones that found nothing.

The analysis is on the right: the document, the claims behind it, the sources used, notes, and a count of what is covered. Closing the chat loses nothing, because the analysis never lived in it.

Every section the method asks for is always listed. An empty one says it is empty.

The steps

From question to reviewed draft Question country, problem, audience Skeleton 2 pages, bullets, every section Stop / go the board reads every line Full draft about 8,000 words, grown from it Review editors, experts; sign-off by people AGENTS AND ANALYST WRITE TOGETHER
The skeleton is the first target. Review stays with people.

The first target is the skeleton: two pages of bullets covering every section. A board already uses it to decide whether a full analysis is worth writing. It takes an analyst about two hours by hand, which gives a time to compare against.

The full draft grows from the skeleton. The approved skeleton is kept as it was when the board read it.

The sections come from a method document the agents read, not from code. The published method has five: the problem, the players, the rules of the game, the state of play, and where reform can happen. A different client can use different sections without changing any code.

The agents

RoleReadsWrites
PlannerThe plan and the claims so farWhat to work on next
ResearchThe registry, search, and the pages a reading model quotesSources it found, notes on them, claims from checked quotes, or a gap when it finds nothing
AnalysisClaims onlyActors, rules and disagreements, and judgements that rest on claims
WritingClaims and analysisThe skeleton and the draft

The analysis role reads claims, never documents. So it cannot form a view from a source that no claim records. This holds because it has no tool that reads a document, not because it is told not to.

Research runs as several agents in parallel, each on one part of the question. They all make claims through one step, which checks the quotes and sets the labels the same way every time.

In the first version one agent does all four jobs with every tool, so the analysis role can still read documents.

The check

The check reads the draft back against its claims Capital spending has sat near 60% for a decade. [12] ✓ cites claim 12 This matters for any reform that needs new money. — asserts nothing new The finance minister blocked the 2019 audit. ⚠ a fact, with no claim Source it find it, make it a claim Mark as judgement the author’s own view Cut it remove the sentence
An unsupported fact is flagged where it sits. Nothing is blocked.

In a published analysis, about one sentence in four carries an endnote. The rest connect, restate or argue. So the check does not ask for a citation on every sentence. It looks for sentences that state a fact and have no claim behind them.

Nothing is blocked. The author picks: find a source and make it a claim, mark it as their own judgement, or cut it.

The check runs on the analyst's writing too.

Where the analyst's knowledge goes

An expert knows things no open source says. When the check flags one of their sentences, they can record it as a claim from their own knowledge. It then sits in the evidence with everything else, marked as theirs.

Interview notes work the same way. A transcript is a source, and claims from it quote it. Running the same question with and without the interviews shows what the interviews added.