Field note · Decision compilation
Crafted Decision Model: Beating Jev without a GPU.
With Decision Compilation, we turn your experts' past decisions into rules and concept vectors, encoded directly into neural-network weights with no training run. On the Banking77 benchmark, rules and concept vectors together score 89.6%, 9.7 points ahead of Jev 1.14. Rules alone still win by 7.2 points. Both versions run on an ordinary CPU.
Benchmark • Banking77 · vs Jev 1.14 • Nov 2026
+9.7
points ahead of Jev 1.14, with rules and concept vectors
+7.2
points ahead of Jev 1.14, with rules alone
0
GPUs needed to run either version
0.26 ms
per message with rules alone, 28 ms with concept vectors, on 4 CPU cores
Head to head
Both versions beat Jev.
All three were scored on the same 3,080 official Banking77 test messages. The test set stayed sealed until our versions were final, and each was scored once.
How it works
Decision Compilation
Most models learn their weights in a long training run. Decision Compilation encodes them directly instead, from examples of the decisions your experts already make. Rules capture what a message says, and concept vectors on an off-the-shelf encoder capture what it means. Every weight comes from a rule or a concept vector, so every decision traces back to the ones behind it.
No training run
Nothing is trained, so there is no training job to run or pay for.
Every decision traceable
Each answer comes with the rules and concept vectors that produced it.
Fixed by editing, not retraining
When the model gets something wrong, we change the rule or concept vector behind it.
Example 6 in the demo
A miss, traced and fixed
When I tried to pay in a store my card din't work.
The typo din't slips past the rules, so the rules-only version answers card acceptance. The concept vectors read the meaning and get it right.
​
One added line in the rules fixes the rules-only version too. There is nothing to retrain.
Rules only, before the fix
card acceptance ✗
Rules only, after a one-line fix
declined card payment ✓
Rules + concept vectors
declined card payment ✓
Speed
No GPU needed.
Both versions run on an ordinary CPU. The rules take about a quarter of a millisecond per message, and the encoder behind the concept vectors adds the rest.
Method
How expert decisions become a model
Step 1
Seal the test set
A share of your experts' decisions is fingerprinted and locked away before any work on the model starts.
Step 2
Encode rules and concept vectors
Rules and concept vectors are built from the rest of your expert decisions and encoded into the network's weights, checked against a separate validation set.
Step 3
Score it once
Each finished version is scored a single time against the sealed decisions, and the whole build is audited.
Step 4
Run it on a CPU
The finished network runs on ordinary hardware and comes with a full evidence report.
Sources
Banking77 by PolyAI (Casanueva et al., 2020), CC BY 4.0. Jev 1.14 score from the Decision Index leaderboard. Timings measured 10 November 2026.
More field notes
Keep reading
Experiments from the rhino.ai research team, each with the result, the method, the checks and the sources.
MODEL MEMORY · Pre-registered experiment
15.6×
less compute per update
Cutting Training Costs 16x with Universal Engram Memory
A 1.7B model writes new facts into a shared memory table, and a frozen 27B model answers questions about them in wordings and formats it has never seen. Only the small model runs during the update.
Qwen3-1.7B writes, Qwen3.8-27B reads Nov 12, 2026