Skip to content
LocoLab LocoLab

What We've Discovered

LocoLab claims five principles. Below are the artifacts — built, measured, or in progress — that those claims rest on. Each entry has a status marker borrowed from research (where papers carry similar markers):

  • claim — position taken, not yet built
  • architecture — designed and specced, implementation pending
  • built — implementation shipped, not yet measured at scale
  • measured — empirical results, not yet published
  • published — paper or report public

This page evolves. Items move down the status ladder as work matures. Failed claims stay on the page with their evidence — that’s the point.

Engineer before hardware

Section titled “Engineer before hardware”

Qwen3-30B-A3B at reading speed on a GTX 1060 6GB

Section titled “Qwen3-30B-A3B at reading speed on a GTX 1060 6GB”
  • Status: architecture
  • Claim: five llama.cpp flags (--n-cpu-moe, --no-mmap, --mlock, --cache-type-k q4_0, --cache-type-v q3_0) plus 24-64GB of system RAM let a $80 used GTX 1060 run Qwen3-30B-A3B at ~17 tokens/sec
  • Artifacts: LocoBench MoE-on-a-budget design spec; implementation plan; Python harness library (33+ passing tests covering config loading, server boot, llama-bench wrapper, hardware fingerprinting); 17 cell configs across 4 VRAM tiers ready to run
  • What would invalidate: if measured tg128 < 10 tok/s on the 1060 with the optimised preset, or if the article’s claim doesn’t replicate

Specialize and harness

Section titled “Specialize and harness”

LocoAgente conversational harness — Phase 1

Section titled “LocoAgente conversational harness — Phase 1”
  • Status: built
  • Claim: a routed harness around small models (four-subsystem architecture: Orchestration / Context / Tools / Inference) makes a Qwen3-4B genuinely useful as a thinking partner for tasks where frontier models converge to the average
  • Artifacts: LocoAgente design spec; Phase 1 library (27 commits, 63 tests at 98% coverage on harness.core); E primitive generate_variants enforcing n >= 2; four orchestration patterns (SinglePass, DebatePattern, SynthesisPattern, IterativeRefinement); three Context profile bundles (business, academic, writing)
  • What would invalidate: if a blind-test reader cannot reliably distinguish the three frames in a Perspective Debate output (frame collapse on small models)

LocoLLM routing thesis

Section titled “LocoLLM routing thesis”
  • Status: architecture
  • Claim: a routed swarm of task-specific Qwen3-4B adapters outperforms a single undifferentiated Qwen3-4B on focused tasks (math, code, summarisation, etc.)
  • Artifacts: LocoLLM repository; QLoRA training pipeline; routing layer; partial benchmark suite
  • What would invalidate: if the router’s task-classification accuracy drops below the gain from specialisation, the system is net-negative versus the unrouted baseline

Vary, don’t average

Section titled “Vary, don’t average”

FrameStrategy deliberate variance engineering

Section titled “FrameStrategy deliberate variance engineering”
  • Status: built
  • Claim: N variants engineered to differ (identity frames, discipline frames, constraint inversions) produce more useful divergent output than N samples hoping to differ
  • Artifacts: four FrameStrategy implementations in harness/frames.py (IdentityFrames, DisciplineFrames, TemperatureLadder, ConstraintInversion); test suite verifying each strategy produces structurally distinct prompts; documented variance-collapse warning on TemperatureLadder for narrow temperature spreads
  • What would invalidate: if the four strategies all collapse to the same output distribution on small models — i.e., engineered framing doesn’t actually channel variance through the model — the principle is just an aesthetic claim

Conversation, not delegation

Section titled “Conversation, not delegation”

Variant + Uncertainty verification contract

Section titled “Variant + Uncertainty verification contract”
  • Status: built
  • Claim: every harness output carries a rationale and surfaced uncertainty (load-bearing flags + verification_hooks; auxiliary confidence); singular outputs are forbidden at the primitive level; the human’s verification loop is cheap because the harness does the prep work
  • Artifacts: Variant and Uncertainty dataclasses in harness/core.py; XML tag parser enforcing required <text> and <rationale>; CalibrationLog recording user picks/rejects/edits as JSONL for downstream analysis
  • What would invalidate: if users routinely ignore the verification_hooks field — i.e., the harness produces hooks but they’re not the right hooks — the principle is performative not functional

Keep Asking — Study 1: Does the Nudge Work?

Section titled “Keep Asking — Study 1: Does the Nudge Work?”
  • Status: architecture
  • Claim: a conversational nudge shifts students from passive delegation to active conversation and improves task outcomes (using frontier models, to isolate the nudge effect from model quality)
  • Artifacts: research design under active development; see research for the full thread (Cognitive Strategy Transfer, DSR AI Education Simulation, Keep Asking Studies 1 and 2)
  • What would invalidate: if nudged students don’t differ from un-nudged controls on either conversation patterns or outcomes

Bridge experience, locally

Section titled “Bridge experience, locally”

LocoPuente PoC

Section titled “LocoPuente PoC”
  • Status: running (= measured in operational terms)
  • Claim: a single mid-range machine (Ryzen 5 2600 + RTX 3090 24GB) running an integrated stack (primary LLM, cited search, image generation, voice, research tooling) delivers frontier-equivalent UX for everyday user tasks at zero cloud cost
  • Artifacts: LocoPuente service running on the Puente machine; browser-accessible LAN deployment; usable today by lab members and visitors
  • What would invalidate: if users with full LocoPuente access still prefer the cloud equivalent for non-privacy-sensitive work, the bridge isn’t bridging