cf-core: pipeline crystallization engine — FPGA→ASIC LLM path promotion #30

New issue

Open

opened 2026-04-06 08:25:43 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-06 08:25:43 -07:00

Owner

Summary

The pipeline module in circuitforge-core becomes a crystallization engine that makes CF products progressively smarter and cheaper to run.

Analogy: LLMs are FPGAs (flexible, expensive per-run). After a human approves a discovered solution path, it crystallizes into a deterministic workflow (ASIC) that runs without LLM until it hits an edge case.

Design

Full doc: circuitforge-plans/circuitforge-core/2026-04-06-pipeline-crystallization-design.md
Implementation: Circuit-Forge/circuitforge-core#33

What this unlocks per product

Product	Task type	Crystallizes to
`osprey`	`ivr_navigate`	DTMF sequence + timing per agency
`falcon`	`form_fill`	Field ruleset per form schema
`peregrine`	`cover_letter`	Template + scoring rules per niche
`kiwi`	`receipt_parse`	Deterministic parser per merchant
`snipe`	`trust_score`	New signal rule added to aggregator

Why this matters

Products get cheaper to run over time (fewer LLM calls)
Crystallized paths are transparent and auditable (not black-box ML)
Human approval gates what gets promoted — critical for high-stakes tasks
Fits the CF privacy model: input hashing, local-first storage, no PII in records

## Summary The `pipeline` module in `circuitforge-core` becomes a crystallization engine that makes CF products progressively smarter and cheaper to run. **Analogy:** LLMs are FPGAs (flexible, expensive per-run). After a human approves a discovered solution path, it crystallizes into a deterministic workflow (ASIC) that runs without LLM until it hits an edge case. ## Design Full doc: `circuitforge-plans/circuitforge-core/2026-04-06-pipeline-crystallization-design.md` Implementation: Circuit-Forge/circuitforge-core#33 ## What this unlocks per product | Product | Task type | Crystallizes to | |---------|-----------|----------------| | `osprey` | `ivr_navigate` | DTMF sequence + timing per agency | | `falcon` | `form_fill` | Field ruleset per form schema | | `peregrine` | `cover_letter` | Template + scoring rules per niche | | `kiwi` | `receipt_parse` | Deterministic parser per merchant | | `snipe` | `trust_score` | New signal rule added to aggregator | ## Why this matters - Products get cheaper to run over time (fewer LLM calls) - Crystallized paths are transparent and auditable (not black-box ML) - Human approval gates what gets promoted — critical for high-stakes tasks - Fits the CF privacy model: input hashing, local-first storage, no PII in records