AI lab · The Wanderer’s Notes
Here I show how I work with a team of AI agents in valuation practice: who does what, where the work stops and waits for my decision, and how each step can be checked.
It is a working method, and it changes with the results of experiments; the log shows what changed and why. Valuer profile: nagul.com.ua.
Pipeline
01 · Brief
The task is set out in a file: sources, verification class, what counts as done. Before the brief goes to the executor, a separate agent looks for gaps in it.
02 · Execute
The executing agent works with files, not in a chat, and hands over a package: the result, a self-check and a list of checksums for every file.
03 · Verify
First scripts with no model, then agents that try to refute each claim; anything important is also checked by a model from another vendor. A separate judge session issues the verdict and edits nothing.
04 · Merge
Only after the verdict and my ‘ok’ does the result go into the library or to the addressee.
step · verification · waits for a person’s ‘ok’ · method as at 10 October 2026
Three ‘ok’s
Changes to the agents’ rules, roles and canonical files. Without an ‘ok’, a proposal stays a proposal.
The result after the judge’s verdict. Without an ‘ok’, it stays verified but goes nowhere.
A document to a client, an email to a person, any deletion, purchases and runs above the limit. This is enforced in the settings, not by agreement: the agent writes the draft, and I press send.
Everything before these points runs without me. If the pipeline stops at one of them, it loses nothing; it simply waits.
Depth of checking
| Class | What it is | Checks |
|---|---|---|
| A | a document for a client, a figure in it, a change to the rules, a merge into the library | scripts, several opposing agents, independent recalculation of spreadsheets from the inputs, a model from another vendor |
| B | internal material: consistency and completeness | scripts and opposing agents from the same vendor |
| C | a technical step with no judgement | scripts only |
The author of the brief sets the class, not the executor. When the executor rated its own work, it set the class too high and ran costly checks where lighter ones would do. Any check that was skipped is always named in the material.
Rules
Only one session edits a document at any one time.
In ‘checked X of Y’, Y is counted by a script from a list, not estimated by an agent.
A reference to a standard is confirmed only by a verbatim quotation from the primary source with its exact location: paragraph, page, edition.
The session that did the work does not declare it verified.
Anything important is checked by a model from another vendor, and an agent runs it, not a person carrying files between chats.
A new session restores the state of work from files: the decision log, the work register, the rules, the latest handover, the verdict. The service’s own memory is only a pointer.
Every agent run is recorded, every decision goes into the log, every package is frozen with checksums. A step without a trail did not happen.
A fact that changes over time, from a registered address to the effective date of a standard, is given with a date and a source. It is checked again before anything goes out.
Experiment log
Each entry here is a contract: which creature got in the way of working with agents, how I tracked it down and what trophy it left in the method. The creatures come from folklore; the cases are real.
Contracts
No. 2 · The Double
A large cost with zero tokens, a borrowed key in the run log and a key of its own for every event.
read the contractNo. 1 · The Shapeshifter
Four references to standards, zero matches with the primary source and a new column in the handover template.
read the contractBestiary
A paraphrase of a standard passing itself off as a quotation.
hunted down · contract No. 1
An expensive model quietly spreads to helper agents.
tracks spotted
One session takes on several tasks and loses its way.
tracks spotted
A system line swallows the real entry in the run log.
hunted down · contract No. 2
A brief seen only by its author.
tracks spotted
A plausible figure with no source.
tracks spotted
Limits
Agents read everything in full, calculate, look for inconsistencies and prepare drafts. The conclusion, its reasoning and responsibility for it stay with the valuer. I set out what IVS and EVS say about this in the note ‘Can a valuer use AI under IVS and EVS?’.