Approach

Data first, then agents.

How we build, in three parts: traceable data before any agent works with it, many agents with one defined job each, and a person who stays in charge. Each part moves forward only on evidence.

1 · Data first

Every figure has a path back to its source.

Fragmented markets report the same fact in different shapes. Before any agent works with it, we reconcile it into one record and keep the path back to every source, with the time it was fetched. Where sources disagree, the value is held, not guessed.

Four sources, four shapes
REGISTRY A
ID A-1234 · issued 142,500 tCO₂e · retired 38,200 t · vintage 2021
fetched 06:00 UTC
REGISTRY B
“Proj. 1234” · 142.5 kt issued · 36.9 kt retired · 2021
fetched 06:00 UTC
EXCHANGE LISTING
“REDD+ Project 1234 V21” · listed · active
fetched 06:02 UTC
RATING REPORT
Project 1234 · flag: reversal risk
fetched 06:05 UTC
RECONCILEOne project, one recordMatch the identifiers across sourcesConvert units and formatsCompare values, flag conflictsKeep the source and fetch time of every field
Click a value to see which sources it came from.
RECORD · PROJECT 1234REDD+ · vintage 2021
Registry A says 38,200 t, Registry B says 36,900 t. The values differ, so the field is held and an alert goes to the person on the loop. No agent uses it until it is resolved.
FETCH LOG · 2026-10-19
06:00:12 UTCregistry A1 record fetched · ok
06:00:14 UTCregistry B1 record fetched · ok
06:02:40 UTCexchange listing1 listing fetched · ok
06:05:09 UTCreconcileretired: A and B differ · held for review · alert sent

Illustrative record with invented values. The method is the one behind QAVCM, our carbon market prototype, where data lineage is documented.

2 · Many agents, two loops

Not one agent, but many, each with a defined job.

Each agent does one thing, such as collecting data, reconciling it, running analytics or drafting a report. It is small enough to check, replace and move up on its own, and a failure stays in one part. A person stays in charge in two ways.

On the loop

A person watches the agents, not each result.

  • Sets the limits up front
  • Gets an alert when a limit is broken or an agent is unsure
  • Can pause or stop any agent at any time
In the loop

A fixed approval step for what matters.

  • Every result of a defined kind passes it, every time
  • At minimum, anything new that reaches a client: new figures, answers, advice
  • An alert from the loop can send a result here for review
ExampleWhat happens
Monthly controlling report
An agent compiles the figures from internal data. Routine, so it goes straight to internal processing.watched from the loop
A client asks for last month's report again
An agent finds the report that was already approved and sends it back. No new figures, no advice.low risk · no approval step
A client asks why a valuation changed
An agent drafts the answer from market and internal data. A person approves it before it goes out.in the loop · approved
Backtest of a new strategy variant
The result breaks a set limit. An alert reaches the person on the loop, and the result is held for review.alert · held for review

These examples show where every product is heading. Today our own operations run on the loop; our products are one step earlier, with a person approving every result.

3 · Moving up a step

Evidence decides when an agent takes over.

People move from in the loop to on the loop, never out of it. Each move is run like an experiment: a hypothesis, an agreed measurement period, and evidence that decides whether one part of a product moves up a step.

01
The data layer and its lineage are documented.
02
Agent results were accepted without change over an agreed period.
03
It is clear what a person approves, what they watch, and how problems are escalated.

Example conditions; the exact ones are agreed per product. See where each product stands today.