Find the expensive habit
Most teams use the strongest model because it is easy, not because the task deserves it. We separate the work that needs judgment from the work that just needs reps.
Gate 00Mission
Stop boiling oceans for a single line of code.
RARE Labs is a bootstrapped AI foundry. Most AI work does not need the biggest model in the room. RARE Labs builds the lanes, checks, and training loops that let local agents carry the repeatable grind while the best models stay focused on planning, judgment, and rescue.
repeat work goes localbest models judgeproof before trustone account goes further
RARE control / story_gate_001
Expensive model
reserved
Repeat work
goes local
Big brain
judges only
Proof
required
Verification status
Clear to spend the big model only when the work deserves judgment.
Gate 01Control
Use the expensive brain only where it earns its keep.
RARE Labs exists because brute-forcing every task through the strongest model is lazy economics. The better system routes volume work to trained local lanes and saves the frontier call for judgment.
Train the cheaper lane
Local agents can carry a lot of the load when the job is narrow, the examples are clear, and the work is checked before anyone trusts it.
Save the best model for judgment
The top model plans, reviews, catches edge cases, and rescues failures. It stops acting like an expensive intern doing repeatable chores.
Gate 02Routing
Stop spending frontier intelligence on chores.
Most AI systems waste their best model on bounded work: file edits, retries, comparisons, formatting, checks, summaries, and follow-through. RARE straps cheap and local models into tight harnesses, gives managers the job of repair and verification, and saves the expensive brain for the call that actually matters.
RARE routing layerPremium tokens are for judgment, not errands
Premium-token burn100% default → ~15% routed target
Old default
One big model gets asked to do everything.
The expensive brain burns through repeatable work that should have been boxed, routed, and checked somewhere cheaper.
100%premium calls
file editsretry failed patchescompare outputsformat cleanuplog summariesbasic QA
RARE reroutes the choresStop spending frontier intelligence on chores.
Small moves go to small models. Managers handle repair. Harnesses prove the work. The main brain gives the final OK.
RARE routed
The expensive model stops doing errands. It becomes the judge.
bounded edits, cleanup, extraction, comparisons
retry, repair, follow through, keep context
tests, logs, diffs, screenshots, fixtures
main brain only when the decision matters
cheap/local workers handle bounded laborcheap managers handle follow-throughmain brain approves only when it matters
Gate 03Principles
How we keep AI work from getting expensive and useless.
The rules are simple: use the best model when judgment matters, use cheaper lanes when the work repeats, and never call anything done until the proof is visible.
01Save premium calls
Do not waste the best model on busywork.
If the task is repeatable, do not default to the most expensive model. Build a cheaper lane and let the big model supervise it.
02Check outside the worker
Never let the worker grade itself.
Local agents are allowed to do the work, but they are not allowed to grade themselves. The check has to live outside the worker.
03Stretch the account
Separate cheap labor from real judgment.
One AI account goes much further when routine labor and high-value judgment are separated instead of mashed together.
04Prove the result
Savings only count when the work holds up.
Savings only count when the result works. If the output cannot be tested, opened, deployed, or reviewed, it did not actually help.
Gate 04Systems
The branded pieces of the RARE system.
This is not a service menu full of jargon. These are the pieces we build when a workflow is wasting premium model calls: a lane for repeat work, a proof loop to keep it honest, and a spend map that shows what should stay expensive.
RARE system 01
RARE Lanes
A RARE Lane is a trained path for repeat work: small code changes, cleanup, QA passes, handoffs, extraction, and follow-up that should not burn premium calls every time.
ProblemRepeat work
RARE methodRARE Lane
OutcomePremium calls saved
Use it when the work keeps showing up, looks similar each time, and can be checked without needing the biggest model to do every step.
RARE system 02
Proof Loops
A Proof Loop is the safety system around the local lane: examples, tests, review steps, and scorecards that show whether the cheaper worker is actually ready.
ProblemCan we trust it?
RARE methodProof Loop
OutcomeChecks people can read
No more “the model said it worked.” The work has to pass checks that a normal person can understand and a stronger model can audit.
RARE system 03
Spend Maps
A Spend Map shows where time, attention, and model calls are leaking out of a workflow, then marks what should stay premium and what can move local.
ProblemWhere is spend leaking?
RARE methodSpend Map
OutcomeClear routing plan
The output is practical: where to save, where not to cut corners, and what proof has to exist before the workflow changes.
Gate 05Transmission
Bring us the workflows bleeding money
Bring the process where the best model is doing the same dull thing over and over. Bring the workflow where people keep paying premium prices for work that could be trained, checked, and routed locally. If the leak is real, we will find the lane, prove it works, and keep the expensive brain focused on decisions that actually deserve it.
Transmission window
Open channelpremium calls wastedrepeat work candidatelocal lane possibleproof before trust