PayCargo · AI Platform · TLDR · 5 of 5

The Forge

Where rides get built from the Anvil's design.

The Anvil produces a design. The Forge turns the design into a working cartridge. Our blacksmith is not one person; it is a coordinated team of AI models that takes the spec, drafts the code, reviews the code, files the pull request, runs the tests, and lands the cartridge on the playground.

The Forge has two parts: the AI team that does the work, and the routing layer that picks which AI does which step. Both matter. Getting either wrong makes the Forge slow or expensive (or both).

The AI team (the blacksmiths)

Modern AI work uses multiple specialized models, not one big one. The right tool per step. A typical Forge run on a new cartridge looks like:

step 1 "Scout the codebase" cheap fast model (Claude Haiku or local LLM) find existing patterns, list affected files step 2 "Sketch the change" cheap fast model propose file paths, function signatures step 3 "Write the code" expensive smart model (Claude Sonnet 4.5) actually generate the Terraform + handler + frontend step 4 "Review the code" expensive smart model catch bugs, security issues, style inconsistencies step 5 "Document the cartridge" cheap fast model generate the README and the cartridge manifest

The cheap scout models do the rough work; the expensive writer model does the high-stakes work. Without this split, every step pays the expensive price and the bill compounds.

The routing layer (which blacksmith does which step)

Today's cheap scout model is Claude Haiku via Bedrock. Tomorrow it might be a local model running on a Mac Mini under someone's desk. Two years from now it might be an open-source model that has caught up to Claude. The Forge has to absorb each of those changes without rewriting cartridge code.

The answer is a router. Cartridges call route("scout") and route("writer"); the router decides which actual backend serves each route. Today both routes hit Bedrock. Tomorrow the scout route quietly switches to a local model and the Forge keeps working. Cartridges never know the difference.

Cost trajectory the router unlocks

Today: naive cloud-only

$0.87 per cartridge spec

What competitors who hard-coded vendor IDs are paying. Opus everywhere. Fast but expensive.

Phase 1.5: Mac Mini scout + Sonnet writer

$0.10 per cartridge spec

Scouts run free on a Mac Mini in the office. Writer is Sonnet via Bedrock. 88% cost reduction with no code change in any cartridge.

Phase 3: nearly all local

~$0 per cartridge spec

When open-source models close the gap on Sonnet, the writer route flips local too. Cost flips to electricity. Sovereignty becomes structural; competitors who hard-coded vendor IDs cannot follow.

This is why the model router doctrine (Step 6.m) is the moat doc. The Forge is not "cheap Claude calls"; it is "the right model per step, swappable as the landscape changes, structurally locked into PayCargo's substrate." That asymmetric switching cost is what we keep emphasizing in executive materials.

The whole loop in one diagram

User intent Operator decision | | v v +-----------+ +---------+ | Anvil | signed-off spec +----> | Forge | +-----------+ +----+----+ | | scouts (cheap fast) | writer (expensive smart) | docs (cheap fast) v +-------------+ | New cartridge| +------+------+ | | pull request v +--------------+ | Human review | +------+-------+ | | merge + apply v +-------------+ | Console | | mounts it | +------+------+ | v +-------------+ | Foundation | | protects it | +-------------+ | v Live cartridge

Where the Forge is today

Forge componentState today
Vibe coding workflow with ClaudeLive. Every cartridge in the repo today was built this way.
Two-turn pattern (scout, then writer)Specced in Step 6.l. Code-agents cartridge will be the first to use it end-to-end.
Model router substrateSpecced in Step 6.m. Ships behind code-agents Phase 1.5.
Mac Mini local inferencePlanned. Reachable from AWS via Tailscale. Pilot once router substrate lands.
Human review gateLive. Every Forge-generated change goes through a pull request with CODEOWNERS review.

Why humans stay in the loop

The Forge produces code fast and cheap; that does not make it correct. Every cartridge change goes through a pull request. A human reads the diff before it merges. Terraform Apply requires a literal "APPLY" confirmation typed by a human. AI scales the writing; humans gate the landing. Both halves are non-negotiable.

Want the technical version?

Forge engineering depth lives in Step 6.l (code-agents and the two-turn pattern), Step 6.m (model router substrate and the moat argument), and the executive narrative in the AI infrastructure moat memo.