Exploration Without Convergence

Praxist has a particular exploration problem to solve: independent peers can be careful, well informed, and still choose the same obvious mechanism. This chapter examines how pre-code design, cohort-level diversity, durable retention, and bounded memory respond to that problem, so the reader can distinguish a plan that broadens the search from evidence that demonstrates an improvement.

Why careful peers still converge

Parallel agents do not automatically create a broad investigation. If every peer sees the same baseline, frontier, research agenda, and recent success, they may all select the path that looks safest. One peer may independently rediscover a calibration change; another may make a nearly identical risk repair; a third may copy the latest durable candidate with a minor variation. The result is a generation full of activity but narrow in mechanism space.

This is not necessarily an intelligence failure. It is a coordination effect. Each peer is optimizing its own chance of producing a plausible result under incomplete information. When the context rewards the same visible signals, local rationality can produce global concentration. A collection of careful contracts may therefore be less diverse than the problem requires.

Praxist's response is not to assign every peer a fixed algorithm family. That would replace one form of convergence with a rigid quota and would make the framework depend on domain-specific categories. Instead, it adds structure at several points in the loop. DIG asks each peer to compare mechanism-level alternatives before implementation. Quality-Diversity, or QD, compares the cohort's candidate choices. Frontier lanes and Gems preserve useful alternatives after evaluation. Peer-local memory keeps a single peer continuous without replaying every message, and the finding graph adds relationships as advisory context.

The common principle is restraint. These mechanisms shape attention and retention, but they do not change the task's evaluator, metric, data split, or evidence standard. They are ways to widen the search around a stable scientific question, not ways to manufacture a new definition of success.

DIG moves disagreement before code

The Deep Innovation Gate, or DIG, is a pre-code design process. Before an implementation peer edits the task project, the system asks it to map the baseline and generate several candidates at the mechanism level. Each candidate receives a structured critique. One candidate is then selected and validated as an implementation contract.

The ordering matters. Ordinary coding often makes the first plausible intervention concrete before alternatives have been compared. Once code exists, the cost of abandoning the idea rises, and a peer may interpret a weak result as a reason to patch the same approach rather than reconsider its mechanism. DIG moves the comparison earlier, when the alternatives are still descriptions rather than sunk implementation effort.

A selected contract identifies more than a variant name. It records the mechanism hypothesis, intervention surface, rejected alternatives, planned changes, expected metric signature, ablation hooks, and checks that should cause the peer to stop early. Those details connect the intended change to an observable consequence. They also give a later reader a way to compare what the peer planned with what the evaluator actually measured.

DIG is deliberately not an experiment loop. It does not train a model, run a task-defined evaluation, write a variant, or create empirical findings. It does not get to change the evaluator, the data split, or the metric calculation simply because a new design would be easier to test. Its output is a design and audit artifact. The contract can be coherent or incoherent; only later execution and evaluation can establish performance.

The gate also has a failure path.