Who Ratchet is for, what it saves, how it could earn money, why now, how it differs from other approaches, and what is not proven yet.
A shop rents a master craftsperson by the hour for every job, including the same simple repair a hundred times a week. Ratchet trains an apprentice from the craftsperson’s finished work, hands over the easy repairs once the apprentice is reliably good, and keeps calling the master for the hard ones. The shop ends up owning the apprentice, and a signed logbook shows whose work the apprentice learned from and that copying it was allowed.
The problem
Section titled “The problem”- LLM bills grow with traffic. Enterprise spend on model APIs was $12.5B in 2025 (Menlo Ventures, 2025 report). Much of what teams send is narrow and repetitive: classification, extraction, ticket triage, steps inside agents.
- The obvious fix is often not allowed. Training a cheaper model on a closed model’s answers breaks the terms of the largest vendors: OpenAI Business Terms §3.3(e) and Anthropic Commercial Terms D.4(a) both forbid using outputs to train competing models. A tool that distills that traffic puts its users at risk.
- Doing it by hand is slow. Logging, cleaning, fine-tuning, evaluating, deploying and watching for regressions are separate chores, and most tools cover one of them.
Who it is for
Section titled “Who it is for”Teams with steady, repetitive LLM traffic on tasks a small model can learn, for example support triage, field extraction, routing and classification, or the many small steps inside an agent. It is a poor fit for open-ended work where every request is different, and for very low volume (see Cost and savings for the break-even).
The value
Section titled “The value”- Cost falls as the student takes over. The student answers when it is confident and the open teacher answers the rest. Every request records what it cost and what the teacher alone would have cost.
- You own the result. The product is a LoRA adapter on an open base model, plus a signed bill of materials that lists the teacher and license of every training example.
- License-clean by design. Closed teachers are refused in code, and each example’s teacher and license are recorded when it is captured. See Licenses.
- Releases are gated. A new student must beat the current one on held-out data, then pass shadow and canary stages, and rolls back automatically if it slips. See The release gate.
- Savings are measured, not promised. In the reference simulation (mock models, ticket triage), the all-teacher cost was about $1.09 per 1,000 requests and Ratchet’s steady-state cost about $0.15 to $0.16, a saving of 85 to 87% (simulation, see Cost and savings). Real savings depend on the traffic mix and on GPU utilization, and Ratchet’s cost report says when a GPU endpoint would cost more than it saves.
How it could make money
Section titled “How it could make money”The plan is an open-source core (Apache-2.0: gateway, ledger, Night School, release gate, dashboard) that anyone can self-host, plus a managed or enterprise offering for teams that do not want to run the nightly loop and GPUs themselves. The idea for the managed tier is to charge a share of savings that Ratchet can verify, with no fee when there are no savings; the code already produces signed daily savings receipts for this (see Evidence). This is a plan, not a product: nothing is sold today.
Why now
Section titled “Why now”- A deadline for fine-tuning customers. OpenAI lists no new self-serve fine-tuning jobs after Jan 6, 2027 (OpenAI deprecations, checked Oct 21, 2026). Teams with fine-tunes there need somewhere else to run them.
- Small models are being used for narrow tasks. NVIDIA Research estimates that 40 to 70% of LLM calls in agent systems could go to small models (arXiv 2506.02153, an estimate, not a measurement).
How it differs
Section titled “How it differs”| Approach | What it does well | What it does not do | Ratchet |
|---|---|---|---|
| Routers and gateways | Pick a model per request, log and govern traffic | Learn from the traffic; every call stays rented | Learns from the traffic and trains a student |
| Fine-tuning and distillation services | Train a model when you start a job | Run by themselves every night, or record each example’s teacher license | Retrains nightly and signs the lineage into the adapter |
| Reference blueprints and tutorials | Show the pattern and its parts | Run unattended; promotion is left to a person | Gates, canaries and rolls back automatically |
Each of these exists in good form; the pieces are prior art. What Ratchet adds is the combination in one loop that runs by itself, with license-checked teachers and a statistical release gate. Sources are listed in What is Ratchet.