Ambition with a leash.
That is Padua.
Padua is our proprietary orchestration and observer layer. It exists to turn search evidence into improvement that a serious organisation can actually authorise: bound to your goals, evaluated by something other than itself, and acting only inside permissions you granted.
Today, Padua's goal and permission qualification software and its observer interface are implemented and tested. Padua is not a deployed autonomous learning agent, and no learning loop is running for customers. Policy foundations are tested; the learning loop is planned.
Nothing starts until
the boundaries are written down.
Objectives
What outcome this project is actually for, in your words.
Measurable outcomes
The measures that would show progress, declared before work begins.
Non-goals
The things this project must not chase, however tempting they look.
Allowed resources
Which systems, repositories, providers and data are in scope — and nothing else.
Budgets
A ceiling on collection and compute spend that Padua cannot lift by itself.
Stop conditions
The signals that halt work, including regression, uncertainty and cost triggers.
Observe, propose, prove,
then earn the next step.
- 01
Observe
Collect evidence about the site, the market and how answer engines describe you — with the surface, method and time recorded.
- 02
Hypothesise
State a testable hypothesis tied to a declared objective, with the outcome that would confirm or refute it.
- 03
Propose
Prepare a specific, bounded intervention with its expected effect, cost and risk written down in advance.
- 04
Evaluate independently
A separate check assesses factuality, usefulness, brand and privacy fit, and non-regression on held-out questions.
- 05
Act within scope
Only an explicitly authorised action runs, inside its permitted resources and budget, as a reviewable change.
- 06
Verify
Re-observe the outcome against the predeclared measure, and record regressions as honestly as wins.
- 07
Retain qualified lessons
Only lessons that survive verification are retained, and they are re-qualified before reuse elsewhere.
This is the intended loop. Its policy and permission foundations are tested; the learning behaviour that runs the loop end to end is planned development scope, not a live service.
Your website — and the engine underneath it.
Padua is designed to improve two very different things. The first is your public presence: content structure, technical foundations, structured data, internal linking and the evidence that answer engines draw on. The second is SearchIntel itself — scheduling, caching, collection costs and adapter performance. A platform that gets cheaper and faster to run is a platform that can afford to observe more for you.
Core-engine optimisation is held to the same standard as customer-facing work: a hypothesis, an independent test, and a verified outcome before anything is retained.
Authority is granted narrowly, and proven repeatedly.
Autonomy is earned per project and per capability. Trust in one repository does not transfer to another, and a capability that performs well for one customer must be qualified again before it is reused.
- Padua cannot rewrite the goals it was given.
- Padua cannot approve its own evidence.
- Padua cannot raise its own spending allowance.
- Padua cannot broaden its own permissions.
- Padua does not merge, publish or deploy anything by itself.
We make no guarantee of recursive self-improvement, of SEO uplift, or of revenue. What we commit to is that every claim is traceable to evidence you can inspect.
No single provider gets to be the platform.
Padua treats models and providers as replaceable components behind a shared evidence and policy layer. Work can be routed across providers, or to local models, under explicit policy — so pricing changes, regional requirements or a better model never force a rebuild.
Built to be talked to.
Scoped APIs and CLI paths already exist for import and exchange, and the architecture is designed for tool-calling and MCP-style interoperability. To be precise: a full MCP server is not currently deployed — it is part of the planned extensibility work, not something you can point a client at today.
See the reasoning before
you grant the permission.
Walk through the goal model, the evidence layer and the catalogue with us.