Current Coding vs the Minimal Solution

The shape of the problem

In three years AI coding has grown a stack — autocomplete, then chat, then in-IDE agents, then autonomous cloud agents, then terminal harnesses, and now "loops" above the harness — with a matching stack of philosophies. Every layer compensates for the same underlying fact: the model sits on a substrate built for humans and machines — files, code, relational tables — not for models, so each layer has to re-supply the context, memory, and coordination the substrate cannot hold. The strain is now measurable. The minimal alternative does not add another layer; it changes the substrate so most of the stack is no longer needed.

Where it strains (verified)

The strain shows up downstream, not in first-draft generation. A controlled METR trial (10 July 2025) found experienced developers 19% slower with AI on mature codebases — while believing they were 20% faster. Veracode's July 2025 test of 100+ models found 45% of AI-generated samples introduced an OWASP Top-10 vulnerability, flat across newer models. GitClear's 2025 analysis reports a ~4× rise in duplicated code blocks and climbing churn as assistants scale. And Addy Osmani's "70% problem" names the pattern: AI gets you most of the way; the last-mile integration, debugging and hardening — where effort actually concentrates — still falls to humans. The scaffolding produces first drafts fast and defers the cost to security, duplication, and review. Scaffolding is also a depreciating asset (the "bitter lesson" applied to agents): every hand-built layer becomes a liability as base models improve.

The minimal solution: one primitive that collapses the stack

The pscale block is that primitive, and its self-reference is exact and small — each level's underscore names its own siblings, so the block states its own schema, needing nothing outside it to be read (see The Pscale Substrate). Because a block authored with forward intent is a state machine whose graph is the JSON tree itself — "the block is the program; the walker is the electricity; the (S, P) pair is the dial" (the live state teaching block) — one object does the work the whole stack is erected to do:

Nothing here is a new framework. It is removal: the block is simultaneously the loop, the memory, the coordinate, the identity, and the spec, so the layers that shuttle those things between a model and a hostile substrate simply fall away. This is the "minimal systemic" bet the project's own systemic-kernel states plainly: "a system is what happens when minimal axioms produce emergent levels that could not have been predicted from the axioms alone but follow from them necessarily" — Go's two rules produce civilisations; DNA's four bases produce all of biology. Minimalism here is not austerity; it is where the emergence comes from.

The industry is adding layers to carry structured meaning across a substrate that cannot hold it. We changed the substrate so the meaning is native — and the layers become unnecessary. The proof that this is not a slide is the fourth document: it is running.