Support the public infrastructure / 2026 program

Help us measure the systems Latin America is asked to trust.

Crafter Research publishes independent model evaluations, reproducible legal-AI benchmarks, and source-backed public corpora. We are looking for compute, research funding, and expert review tied to specific open milestones.

Offer compute Fund a milestone

Shipped proof

The proposed work starts from artifacts that already exist.

These are inspectable baselines, not projections. Proposed programs appear separately below.

11models evaluated

LatamBench

Trueque-500 and CHOCLO-500, with 37 published run records and response sets, plus 32 completed judge transcripts.
21,244public legal norms

Legalize Perú

Approximately 98% full text across the national tier and 26 regional jurisdictions, versioned as a public corpus.
N=35firm gold set

Amicus Eval

Six retrieval configurations with independently recomputable metrics and explicit limitations.

Proposed research

Three programs waiting on real bottlenecks, not invented ones.

Each program has a useful 90-day outcome, a smaller fallback, and an open evidence trail. These projects are proposed and compute-gated. They are not presented as current output.

Proposed / compute-gated

LatamBench Open Weights Observatory

The models Latin America can actually run.

Evaluate open-weight and regional models under controlled serving. Measure accuracy, abstention, hallucination, quantization effects, serving variance, and benchmark-reference defects.

Resource
300–600 A100-equivalent GPU-hours
90-day gate
6–8 open models across two benchmark families, with controlled repetitions and human validation.
Public return
Raw responses, exact revisions, serving recipes, judge transcripts, exclusions, cost, and measured utilization.
Boundary
Core scope stops at 70B. Larger models follow measured utilization, not ambition alone.
Proposed / compute-gated

Spanish Legal Model Reliability Lab

Can a model cite the law that was valid then?

Compare retrieval-augmented, long-context, and bounded local-model approaches over Peruvian law. Test citation correctness, temporal validity, groundedness, abstention, and hallucination.

Resource
$2k–$5k API credits + expert review
90-day gate
Expand Amicus Eval from N=35 to at least N=150 and compare 4–6 fixed configurations.
Public return
Per-query evidence, corpus versions, failure taxonomy, and one-command metric verification.
Boundary
The benchmark measures system behavior. It does not replace qualified legal judgment.
Proposed / compute-gated

Public-Service Agent Reliability Benchmark

The trámite gauntlet.

Replay realistic public-service workflows across models and interface perturbations. Measure task completion, source fidelity, unsafe actions, recovery, and portal drift.

Resource
$3k–$8k multimodal model credits
90-day gate
20–25 non-destructive workflows, four configurations, and at least three repetitions per task.
Public return
Redacted trajectories, task fixtures, scorer definitions, replay format, and failure taxonomy.
Boundary
No payments, identity verification, submissions, or irreversible government actions in the pilot.

Ways to support

Match the resource to the work.

The strongest offer is not necessarily the largest. It is the one that unlocks a named release with enough time to use it responsibly.

API

Model access

$5k–$10k in creditsValid for 12 months

Frontier controls, independent judges, long-context legal evaluation, embeddings, reranking, and multimodal agent runs.

Fallback: $1,000 for one bounded release.
GPU

Accelerator capacity

1,000 A100-equivalent hoursOr $5,000 in cloud credits

Controlled 70B serving, open models unavailable through APIs, repetitions, and Spanish retrieval experiments.

First pilot: 300–600 hours over 90 days.
REV

Human validation

$25,000 milestone grantResearch and expert review

Native-speaker adjudication, qualified legal review, dataset curation, recurring maintenance, and publication work.

Fallback: $5,000 for one reviewed legal milestone.
HW

Measured hardware pilot

90-day loanDedicated or equivalent remote capacity

Repeatable high-memory inference with a utilization report, reproducibility notes, and a return, extend, or convert decision.

Hardware follows a measurable workload. Cloud is always acceptable.

Research independence

Support buys a release. It never buys the result.

Every awarded resource is disclosed. The same protocol applies to supported and unsupported models. Providers receive no editorial veto, ranking control, or right to suppress negative results or infrastructure failures.

  1. 01Resources and approximate value are disclosed.
  2. 02Model inclusion follows the study design.
  3. 03Methods, failures, and limitations stay publishable.
  4. 04Sponsored case studies remain separate from research conclusions.

Start with one measurable release

Tell us what resource you can offer.

We will reply with the matching workload, budget, milestones, stewardship plan, and public deliverables. Compute, credits, expert time, institutional collaboration, and project grants are all useful.

Write to Railly railly@crafterstation.com / Crafter Research is a public-interest research lab by Crafter Station.