WAYS OF WORKING · APPLIED NINO CHAVEZ
An agent's research report states everything with the same confidence — the number it read at the source and the number it absorbed from a search snippet look identical on the page. The technique: make every load-bearing claim wear a label that says how it was obtained, and audit the labels adversarially before anything ships.
→ or scroll to advance · ← to go back
02 THE FAILURE THIS FIXES
Two things go wrong in agent research, and they compound. First: search summaries are hearsay — the tool's synthesis of a page is not the page, but the agent quotes it with page-level confidence. Second: the byline vanishes — a vendor ranking itself, a rival attacking its competitor, and a neutral review platform all arrive as identical-looking "results."
In the session behind demo 09, the first research pass compared e-commerce platforms using the recommended vendor's own comparison blog. When challenged, the "corrected" pass swapped in a damning statistic — sourced from a competing payment processor's marketing. Two rounds, two biases, one failure: nobody asked who wrote the page. the originating case — confident, tabulated, and marketing all the way down
The reader of the final deliverable — a client, an exec, your future self — can't re-run the research. If the confidence isn't labeled, they inherit yours. Labels move the trust decision to where the evidence actually is.
03 THE TECHNIQUE · PART 1
Every load-bearing number — price, rating, statistic, anything that gates the decision — carries one of three labels. The legend ships with the deliverable, so the reader knows what each color promises:
fetched at source
The agent read the actual page, ran the actual query, hit the actual API — in this session, with the fetch in the transcript. The only tier allowed to carry "verified" language.
The bar: a fetch you can point to, not a memory of one.
search-derived
The number arrived in a search summary or a secondary page. Might be right — often is — but nobody read the source. The label tells the reader: confirm this yourself before you commit money to it.
Not a failure. An honest amber beats a false green every time.
blocked / unverifiable
The source refuses direct reads — many review platforms 403 automated fetches. Say so, in the deliverable, next to the number. Then check whether the recommendation survives without it. If it doesn't, it needs different footing.
The test: would the advice still stand if this number vanished?
04 THE TECHNIQUE · PART 2
Labels the author assigns to their own draft inherit the author's blind spots — the drafting agent believed its green tags. The audit is a separate pass, fresh context, told to refute rather than confirm:
Pull every load-bearing claim: prices, ratings, statistics, and each instance of "verified / confirmed / checked directly." Mechanical work — a regex finds most of it.
For each claim, answer from the transcript, not from memory: was the source fetched, or did this arrive in a search result? Who owns the domain it came from?
Fetch what carries "verified" language. Confirms → keep. Contradicts → fix the claim. Blocked → downgrade the label and say why.
Turn the deliverable's own confidence legend on itself: every green tag must trace to a fetch. A label system that lies on one row is worse than none.
What does the draft imply is free, complete, or settled that was never priced or checked? The originating case shipped "costs $0 to start" while the plan's only real recurring cost went unmentioned.
Corrections first, with what changed and why. Keep the conclusions only if they survive on the downgraded evidence — "the labels were wrong but the advice holds" is a legitimate, common outcome.
A fresh context refuses the draft's framing in a way its author can't — a different model, a different session, or just different instructions all work. What matters is that the auditor treats every "verified" as a hypothesis.
05 KEEPING IT
Run manually, this technique lasts exactly as long as you remember to invoke it. The session behind demo 09 needed three operator-prompted audit rounds — so the discipline got promoted into standing equipment the same afternoon (demo 03's pattern):
A standing rule
One paragraph in the preferences document every agent session loads: snippets are hearsay, "verified" requires a fetch, vendor and competitor content are both marketing. Rules the agent reads every session outlive rules you say once.
A deterministic checker
A pipeline stage that extracts claims from a draft and validates each against a provenance ledger — any "verified" without a fetched-at-source entry fails the build. No AI in the loop: regex, a ledger, an exit code.
A named audit pass
The six-step loop as an invocable skill, run before anything client-facing ships. Naming it matters: "run the evidence audit" is a checklist item; "be more careful about sources" is a hope.
Honest limits: re-deriving everything would consume the session — scope the audit to claims that gate decisions, touch money, or ship to someone who can't check them. And some sources simply can't be read by an agent; the technique's answer isn't to work around the block, it's to say so on the page.
06 DO THIS TODAY
If you only prompt
Add one line to any research ask: "label every number: did you read the source page, or a search result about it? Note who owns each domain you cite." Then ask the follow-up on anything load-bearing: "re-derive that one at the source."
If you review AI output
Run the audit as a second session: paste the draft, instruct it to extract every claim and refute the verified ones. Expect the common outcome — conclusions hold, labels don't — and treat a draft that survives unchanged with suspicion.
If you build the deliverables
Make the label a first-class field in your template — green, amber, blocked — with the legend printed for the reader. Let "this couldn't be verified, and here's why the advice stands anyway" be a sentence your format allows. That sentence is where trust comes from.