The Calibrated Marketplace
When several Faculties can do the same kind of work, the system routes to the one with the highest earned trust, measured purely on outcomes, blind to mechanism. Reliability is bought with evidence, not declared.
Overview
This page describes the direction of trust-based routing. Today, routing rules and trust are what you configure and read; the automatic ranking described here is where they are headed, not a switch you can turn on.
When more than one Faculty can handle the same kind of work, something has to decide which one actually gets it. Syncropel decides on earned trust: it routes the work to the candidate with the highest trust for that domain, and it measures trust purely on outcomes that independent evaluators judged. We call this the calibrated marketplace.
"Calibrated" is the load-bearing word. This is not a popularity contest, a price auction, or a hand-tuned preference list. The only currency is verified results. A Faculty that consistently produces work that independent evaluators accept rises; one whose work gets rejected falls. The marketplace is a measurement, continuously refreshed.
Mechanism-blind by construction
The marketplace never inspects what a Faculty is made of. When several candidates declare they can do a kind of work, the system ranks them by their trust in the relevant domain and dispatches the best one, reading nothing but the trust scores. A deterministic solver, a language model, and a script are ranked on exactly the same axis.
This matters because it means a non-LLM mechanism competes on equal footing. There's no built-in assumption that "the smart thing to do is call the model." If a cheap exact solver produces better-judged results than a language model on a class of problems, the marketplace routes that class to the solver, not because anyone coded that preference, but because the evidence put the solver on top.
The intended shape: two Faculties declare the same problem kind, say an exact optimization solver and a weak greedy heuristic. Both do real work; independent verdicts grade their results; their trust diverges. A fresh problem of that kind, naming no specific Faculty, then goes to whichever candidate has earned the most trust. Reliability is measured, not assumed.
How a winner is chosen
When a piece of work could go to multiple Faculties, the marketplace:
- Gathers candidates by declared capability. Every Faculty that advertises it can handle this kind of work is in the running, the system discovers competitors from what they declare they do, not from a curated list.
- Ranks by domain trust. Each candidate is scored by its earned trust in the work's domain (see Trust). The mechanism is never read.
- Dispatches the best. The highest-trust candidate gets the work.
Two disciplines keep this honest:
- Cold-start floor. A candidate with no verdicts scores effectively zero. A brand-new Faculty cannot leapfrog a proven one just by existing, it has to earn its standing first.
- Ties favor the incumbent. When scores tie, the existing primary keeps the work. A fresh challenger displaces a proven Faculty only by actually scoring higher on the evidence.
Why this is the economic payoff, not just a routing trick
The marketplace is the mechanism by which Syncropel would get cheaper and more reliable over time without anyone tuning it. As evidence accumulates:
- Work gravitates to whichever mechanism is genuinely best at it, often a cheaper, more deterministic one than a language model.
- Proven Faculties get more of the work, which produces more evidence, which sharpens the ranking further.
- New mechanisms can be introduced safely: deploy them as candidates, let them prove themselves on real work, and they only take over where they actually win.
Nothing moves faster than the evidence allows. A mechanism that's been great historically but starts producing worse-judged results drifts down as recent verdicts pull its trust lower, and the marketplace quietly reroutes to a better candidate.
What's next
- Faculties: the deployable units the marketplace routes between, and the
spl facultycommands to deploy them. - Trust: the evidence-based scoring that ranks candidates.
- Routing rules: how explicit routing and the trust-based marketplace fit together.
Faculties
A Faculty is a deployable unit of intelligence addressed by what it does, not what it's made of. An LLM, a solver, a classifier, or a script can all fulfill the same Faculty — and the system routes work to whichever has earned the most trust at the task.
Patterns
How a workspace learns from work that keeps succeeding the same way, reuses the outcome instead of reasoning it out again, and stops reusing it once it starts to fail.