← All postsLaunch · Ada beta
A decision model that learns from your own precedents, answers in milliseconds, shows the cases behind every answer, and runs without a transformer.
Most of the decisions a business makes are not new. Approve this loan or don't. Route this message to that team. Escalate this alert, close that one. Each has been made hundreds of times before, and the outcome of each is on record. Today those decisions go one of two ways: a person makes them, or an app sends the whole case to a frontier language model and pays for it to reason from scratch, every time, with a different answer on a bad day.
precedent classifier: you record the decisions you've already made, and she answers a new case from the ones it resembles. She's built on hyperdimensional computing, not a transformer. There's no training run, a new decision counts the moment it's recorded, and the same question always gets the same answer. Every answer comes with the precedents it stands on, and when nothing on record is close enough, she says so instead of guessing.
Ada also works well in front of a language model. She decides the cases she's sure of and hands the rest over, so the model sees a fifth of the traffic and every hand-off becomes a precedent for next time.
our benchmarks, and each one below has its caveats under it.
97.5% of bank messages routed right, with Ada in front of Claude FableFable alone: 96.5%
$0.22 for those 200 messages, against $300 for Fable aloneat list prices
100 the bank's cost on 200 loans, lowest on the tableLLM: 110 and 114
0 / 200 answers changed on a rerun. Two LLM runs disagreed on 12.deterministic
A different kind of model #
A language model is optimized to continue text. Ada is optimized to repeat a decision the way your own history made it. That one difference changes almost everything else.
| Frontier LLM | Ada | |
|---|---|---|
| Learns from | Pretraining, plus examples pasted into every prompt | Your own decisions, recorded once. A new one counts the moment it's recorded. |
| Inputs | Tokens | Your cases as they are: categories, numbers, text |
| Outputs | Text, parsed back into a decision | A decision, a calibrated confidence, and the precedents it stands on |
| Same input, same answer? | Usually | Always |
| When it doesn't know | Answers anyway | Says so, and hands the case to a person or a model |
| Cost per decision | $0.004–$1.50, depending on how many examples ride along | About $0.00000003 |
| Speed | 0.5–3 s | About 6 ms |
| Explains itself | In prose, after the fact | Which precedents, which fields matched, and how much each counted |
| Good for | Open-ended work, new problems, writing | Repeated decisions with a track record |
Benchmark one, routing bank messages #
Banking77 (PolyAI) is 13,083 real messages from a bank's customers, each labelled with one of 77 intents: card arrival, lost or stolen card, exchange rate, top-up failed and so on. They're short, informal and often ambiguous between neighbouring intents. It's free text, which is where a developer reaches for a language model today.
Every system learned from the same 10,003 messages and was scored on the same 200 held-out ones, each routed as its own call, the way a chatbot makes them. The Claude models read the training file and routed by judgement. Nothing was tuned on the 200.
With Ada in front, every model does better than it does alone. Ada keeps the 156 messages she's sure of and gets 98.7% of them right. The other 44 go to the model with Ada's three nearest precedents in the prompt, instead of the whole training file. For the two largest models the pair beats the model alone; for the two smaller ones it adds eight to ten points, because the small model only has to handle the hard fifth. Alone, Ada routes 90.0%.
| Accuracy | Cost for the 200 | Per 1,000 messages | |
|---|---|---|---|
| Claude Fable 5.1 alone | 96.5% | $300 | $1,500 |
| Ada, then Fable for 22% | 97.5% | $0.22 | $1.10 |
| Claude Opus 5.5 alone | 96.0% | $120 | $600 |
| Ada, then Opus for 22% | 97.0% | $0.09 | $0.44 |
| Ada alone | 90.0% | $0.000006 | $0.00003 |
A thousand times cheaper and more accurate than the best model alone. And because every hand-off is recorded, the model's share keeps falling.
Benchmark two, credit decisions #
German Credit (UCI Statlog) is 1,000 real loan applications with 20 fields each and the bank's actual outcome. Every contestant learned from the same 800 and was scored on the same 200 held out. In this dataset's own cost matrix, approving a loan that goes bad costs five times as much as refusing a good one, so the number a bank cares about is the cost, not the accuracy.
Ada costs the bank the least: 100, against 110 and 114 for two runs of Claude Fable, and 118 at best for TypeSafe's Jev. Every setting was chosen on the 800 before the 200 were opened. Ada gets there by refusing more: 70.0% of her decisions match the bank's, against 77–79% for the LLM, because a refused good loan is the cheap mistake.
| Accuracy | Bank's cost | Same answer twice? | |
|---|---|---|---|
| Always approve | 67.5% | 325 | — |
| Claude Fable 5.1, two runs | 77.0% / 79.0% | 110 / 114 | 188 of 200 |
| Jev, 300 / all 800 precedents | 71.0% / 70.5% | 123 / 118 cost-aware | not tested |
| Ada, plain vote | 78.5% | 171 | yes |
| Ada, calibrated | 70.0% | 100 | yes |
| Ada, asking a person when every answer is dear | 81.1% of 111 | 45 on those | yes |
The last row is the dial again. Priced so that asking a person costs half a refused good loan, Ada decides 111 of the 200 herself at a cost of 45 and sends the other 89 to a person.
Speed and cost #
Here's what each credit decision takes. The LLM and Jev keep nothing between calls, so the precedents are paid for on every request. Ada pays for them once, when they're recorded.
At roughly $0.00000003 and a few milliseconds a decision, Ada is about five orders of magnitude cheaper than a batched LLM call and 100 to 600 times faster.
On TypeSafe's own yardstick #
TypeSafe's published workflow evals have no ground truth. They score a contestant by agreement with the average of two frontier models. We built that reference for the 200 loans (Claude Fable 5.1 and Claude Opus 5.5, averaged) and scored the first-run contestants against it as well as against what really happened.
| Contestant | vs real outcomes | vs model consensus |
|---|---|---|
| Ada, first run | 77.0% | 88.5% |
| Jev, all 800 precedents | 70.5% | 79.0% |
| LLM, run 2 | 79.0% | 78.5% |
| LLM, run 1 | 77.0% | 74.5% |
| Jev, 300 precedents | 71.0% | 72.5% |
| Jev, no precedents | 49.5% | 44.0% |
On TypeSafe's measure Ada leads Jev by 9.5 points. But the consensus itself matches the real outcomes only 74.5% of the time, and it calls 28 loans bad where 65 were. Agreement with other models measures how model-like a contestant is. We'd rather be scored on what happened.
Benchmark three, routing Kubernetes issues #
kubernetes/kubernetes issues (1,527 to learn from) and asked her about 104 issues opened after the cutoff. The question: which SIG (special interest group) should own each one? We stripped every line where the reporter named a SIG, so she had to route from what the issue says: its template, the components it mentions, and the code paths it cites.
66% right on the first try and
86–89% within three, against 35% and 62% for always guessing the most common SIG. What makes this usable is that her confidence means something.
Above 0.7 confidence she answered a third of all issues and got 94% of them right. Above 0.9, 100%. Let Ada take the cases she's sure of, and send the rest to a person.
Finding the fix #
Because every answer cites its precedents, Ada can do more than classify. For 56 held-out issues whose fix is known, we checked whether any of her three nearest solved precedents had been fixed in a directory the real fix also touched.
77% of the time, the cited precedent pointed at the right part of the codebase. Here's the start of one answer:
#141993 Memory Manager accepts an incomplete pod-level checkpoint
as a restored allocation after kubelet restart
-> sig/node via #141069 Memory Manager pod-scope restore
compares container blocks…
fixed by PR 141070 in pkg/kubelet
That's the explanation: which earlier issue, and where it was fixed. No prose to trust or doubt.
Does it scale? #
A precedent memory is only useful if it can hold your whole history. Ada Cloud records about 7,800 precedents a second, so a million take about two minutes, and answers stay quick as the memory grows.
What Ada isn't #
Ada doesn't write, summarize, plan or chat. She can't decide something nobody has decided before; when nothing on record is close enough she says so and hands it back. If your problem is new every time, use a language model. If you've made the same decision a thousand times, you already have what Ada needs.
What's next #
Ada is in beta today. Next are more public benchmarks, including TypeSafe's security-incident workflow end to end, and models that answer every question about a case at once.
Why name it after Ada Lovelace? #
She wrote the first published program, for Babbage's Analytical Engine, a machine fed by punched cards. She also wrote that the engine "has no pretensions whatever to originate anything." For a decision model, that's a feature.
Is there a neural network in it? #
No. There's no transformer, no GPU and no training run.
How does Ada learn something new? #
You record the case with its outcome, and the next question can use it. There's nothing to retrain and nothing to redeploy.
Can I see why it decided something? #
Every answer lists the precedents it stands on, and which of their fields matched.
Where does my data live? #
Your models run in Glyphh's cloud, isolated per organisation. We never train on your data.
If you make the same decision a thousand times a day, I'd like to hear what it is.
Build with AI. Run without it.
— Timme
Ada learns from your labelled cases, answers from its nearest precedents, and hands what it is unsure of to an LLM. Start free on Ada Cloud.