Test first. Escalate second.

A coding agent that runs every fix against a locked unit test on Nemotron Nano first, and only pays for live docs or a bigger model when the test says why it failed.

Open the workbench See the 5-arm ledger

Nebius Token Factory
NVIDIA Nemotron 3 Nano, Super, Ultra
Tavily search, extract, map

SchemaGuard v3 migration Verified · exit 0

Cost of the verified fix$0.002095

Pick a failure type to see which rung of the ladder catches it.

Ten tasks, locked tests

Five routing strategies ran the same ten Python tasks. A patch only counts when the read-only test passes.

100%

of tasks verified by Mjolnir, 10 of 10

$0.0017

per verified fix, 52% below Ultra with Tavily always on

0 of 6

2026 API changes fixed by bare Nemotron Ultra

Bugs fail in two ways

A bigger model cannot guess a method that was renamed after it was trained. A web search cannot solve a graph algorithm. Mjolnir reads the error and sends each bug down the right road.

A steel arch bridge spanning still grey water

Missing knowledge

ImportError, AttributeError or an unknown keyword. The code calls an API that changed in 2026.

Fix: Nano with live Tavily docs

Pale stone columns rising into pointed vaults

Missing reasoning

AssertionError on an edge case. The API is fine; the logic is wrong. Search would only add noise.

Fix: escalate to Super, then Ultra

Four steps

Nano probeexit 1

Nemotron-3-Nano-30B $0.000045 · 210 ms

test_solution.py
mode 0444 · sha fd6cb989

First, the cheapest model tries the fix against a test it cannot edit.

Error receiptKnowledge gap

Missing symbol: dict solution.py, line 11

AttributeError: 'UserPayload'
has no attribute 'dict'

Then the traceback is cut to three lines that name the cause.

Tavily docsfound

search, extract, map 342 ms

SchemaGuard 3.0:
use .model_dump_strict()

Next, only API errors pull the current docs from the web.

Verified diffexit 0

Total $0.002095 2 rungs used

- user.dict(exclude_none=True)
+ user.model_dump_strict(
+     exclude_none=True)

Finally, the test passes and you get the diff with its cost.

The 5-arm ledger

Six tasks use SDK changes from 2026; four are pure logic. Same tasks, same locked tests, five ways of choosing a model.

Repeating modular facade of a white concrete building

Strategy	How it routes	API changes	Logic	Solved	Cost per fix
Mjolnir 2D ladder	Nano first, Tavily on API errors, Super or Ultra on logic	6 / 6	4 / 4	100%	$0.001671
Ultra + Tavily always	Biggest model, web search on every task	6 / 6	3 / 4	90%	$0.003484
Nano + Tavily	Small model with docs, no escalation	5 / 6	1 / 4	60%	$0.003396
Ultra alone	Biggest model, no docs	0 / 6	4 / 4	40%	$0.002613
Nano alone	Small model, one try	0 / 6	1 / 4	10%	$0.000330

Ten tasks is a small sample, so read these as a method check, not a market benchmark. Cost per fix divides each strategy's total spend by the tasks it actually solved. Raw receipts: live_5arm_telemetry.json · deterministic_5arm_benchmark.json.

Every fix has a receipt

Choose a task to read the diff, the docs Tavily returned and the path up the ladder.

Suspension bridge tower rising through coastal fog above dark cliffs

api-01-schema-validator Verified · exit 0

Nano → Nano + Tavily · $0.002095

- return user.dict(exclude_none=True)
+ return user.model_dump_strict(exclude_none=True)

Loading the Tavily receipt from the benchmark artifact.

Race it
yourself

Pick a task and a baseline. The workbench runs both side by side against the same locked test and shows what each one cost.