Most decisions your business makes aren’t new.
Approve the loan or refuse it. Route the ticket to the team that owns it. Escalate the alert or close it. Each one has been made hundreds of times, and the outcome is on record. Glyphh Ada decides the next one from that record.
What you want
A decision your software can act on: the right call where a mistake costs money, a confidence you can set a bar on, the same answer every time, and a reason you can show.
What you get today
A frontier model reasons from scratch on every call. It pays for every token, can answer differently on a rerun, and answers anyway when it doesn’t know. Training your own means a dataset, a pipeline and a retrain for every change.
The approach we took
Precedent. Ada answers a new case from the decisions you have already made, the way case law does. It runs on hyperdimensional computing, not a transformer: no training run, no language model in the path, and a decision counts the moment it’s recorded.
A language model is built to continue text. Ada is built to repeat a decision the way your own history made it.
Typed decisions
An outcome, not a paragraph
Ada answers with one of the outcomes your spec allows and a calibrated probability for each. Your software acts on it with nothing to parse.
Calibrated confidence
It knows when to ask
Set the bar. Above it your software acts; below it the case goes to a person or a model, and their answer becomes the next precedent.
Same answer twice
As reliable as code
No language model in the path, so the same case gets the same answer, in milliseconds. On 200 loans, a rerun changed none of Ada’s answers; two runs of the same LLM disagreed on 12.
The same question, with and without Glyphh.
An expense claim: business class to London, $2,400, over the limit and not pre-approved. Without Glyphh, Claude reads the 40 past decisions and answers. With Glyphh, your software asks Ada, which learned from the same 40.
without Glyphh · claude · Claude Fable 5.113.3 s
19,752tokens in
678tokens out
$0.28at list price
13.3 sto answer
with Glyphh · node · your code326 ms
0LLM tokens
1Ada read
$0.0001at most, past your plan
326 msto answer
Both are terminal recordings from one laptop on 26 October 2026. They play in real time, lined up so the two questions go in together, and the Claude one ends by scrolling up to the start of its answer. Claude Code ran Claude Fable 5.1 with nothing but the file of past decisions. Its tokens come from the session’s own transcript, priced at Anthropic’s list price. Your code asked Ada Cloud through the Glyphh SDK, with no language model. The 326 ms is the whole round trip from a fresh process, connection setup included.
Most cases never need the language model.
Claude reads every past decision on every question, so its bill grows with your history. On 200 bank messages with 10,003 past ones, Ada answered the ones it was sure of and handed the rest to Claude Fable. Together they were more accurate than Claude Fable alone, and cost cents instead of hundreds of dollars.
Banking77, 200 held-out messages, 77 intents. The experiment and every run are public.
Ask the questions you already answer every day.
A question is the fields a case has and the outcomes it can lead to. Describe it once, record the decisions you have already made, and ask. Claude can write it with you.
Every answer comes with its receipts.
Ada finds the precedents nearest the case, weighs them and reports what they say: a typed outcome, how sure it is, and the precedents that voted. When nothing is close enough, it says so and asks.
The caseexample
- Class
- TimeoutError
- Message
- payment gateway timed out
- Status
- 504
- Service
- checkout
Asked by your software, from the incidents model.
The answer
Raise the circuit breaker.
82% sure
From 6 precedents on record.
- Last time
- Cause
- gateway rate limit hit on a retry storm
- Owner
- payments
- Runbook
- RB-12, back off retries and raise the circuit breaker
A spec can put layers behind a clearance: a status bot gets the fix, an engineer also gets the cause, the owner and the runbook.
When Ada is unsure the answer is ask. Your software hands the case to a model or a person, records the result, and Ada has one more precedent.
A model is a JSON spec and the cases you have labelled. Each graded case becomes a precedent, and Ada learns its own weights from them.
Reach Ada from Claude or from your own code.
Your models live in Ada Cloud, isolated per organisation, with nothing to install or run. We never train on your data. Ada is in every plan, Free included.