Evaluation + observability
Ship model changes on evidence, not on vibes.
Graded evaluations in CI. Full-fidelity traces. A price that does not rise when you look at your own data.
- model-a120 ms
- prompt18 ms
- context64 ms
- tool: search_index892 ms
- retry310 ms
- response36 ms
Scene 01 / The signal
02 / Quietly, completely
See the whole call. Decide with the whole record.
Prompts, context, tools, retries, responses, cost and timing — kept together.
- 01Evaluations
- 02Observability
- 03Promotion
- 04Human review
- 05Self-hosting
03 / The platform — 01
Evaluations
Datasets, graders and thresholds live in your repository. They run in CI, and the diff goes back to the pull request.
| Case | Grader | Baseline | Candidate | Diff | Status |
|---|---|---|---|---|---|
| login_flow | correctness | 0.92 | 0.95 | +0.03 | PASS |
| refund_policy | factuality | 0.88 | 0.91 | +0.03 | PASS |
| product_search | relevance | 0.90 | 0.87 | −0.03 | FAIL |
| order_status | tool_use | 0.94 | 0.96 | +0.02 | PASS |
| return_steps | helpfulness | 0.89 | 0.92 | +0.03 | PASS |
48 / 50 passed
2 improved1 regressedNo threshold crossed
03 / The platform — 02
Observability
Prompt, context, tools, retries and the response stay together, with the cost and the time at every node. Six months, at full fidelity.
- prompt£0.00218 ms
- context£0.00464 ms
- tool: search_index£0.008892 ms
- retry£0.001310 ms
- response£0.00336 ms
03 — 05 / The rest of the platform
Promotion
Turn any traced request into an evaluation case in one click, already redacted by the collector.
Human review
A queue support staff and domain experts can actually use, feeding the same score table as the automatic graders.
Self-hosting
The same build at the same price in your own cluster. Air-gapped by arrangement.
04 / From traffic to tests
The cases were already there.
Watch what arrives. Keep what matters.
Scene 02 / The record
Frame 024
The numbers we can stand behind
TC 00:20:00
20min
To the first trace, about
TC 06:00:00
6mo
Of full-fidelity traces on Team
TC 14:00:00
14days
Trial on your staging traffic, no card
TC 40:00:00
£40
Per seat a month, nothing metered
05 / Getting started
Four steps. One afternoon.
About twenty minutes to the first trace.
01
Install the collector
One package. One environment variable. Personal data is redacted inside your network before anything leaves; the rules are open source.
02
Watch a week of traffic
See what users actually send, including the awkward edge cases nobody wrote down.
03
Promote fifty requests
Fifty real cases beat four hundred invented ones. Turn them into a useful suite in an afternoon.
04
Gate the next release
Add the suite to CI. Every change arrives with its evidence attached.
06 / Honest fit
Built for the work after launch.
Written down so you find out now, not in week three of a trial.
Look elsewhere when
Look elsewhere when: You have not shipped anything yet.
Look elsewhere when: You want a prompt playground or a model gateway. Halcyon is neither.
Look elsewhere when: You need fine-tuning infrastructure.
Look elsewhere when: Your volume is a few thousand calls a month. The free plan is all you need.
Good fit when
Good fit when: A model feature is already in production and you cannot tell whether a prompt change helped.
Good fit when: Your evaluation set was written by hand and is probably too clean.
Good fit when: Someone senior asks what the feature costs per customer and nobody can answer.
Good fit when: Compliance will not allow a hosted tool.
07 / Voices
From teams who checked first
We ran Halcyon against our own traffic before signing anything.
The per-seat price decided it. A per-trace price grows every quarter.
Opening a request from four months ago and seeing exactly what the model saw has settled more arguments than any dashboard.
08 / Pricing
One price. Every feature.
Per seat. Nothing metered. A 14-day trial on your own staging traffic. No card. No call.
Free
£0forever
For one team getting a feature to production. Not a trial — it does not expire.
- Up to 3 seats
- 30 days of full-fidelity traces
- Evaluations, graders and CI gates
- SSO and audit log
- Self-hosting (not included)
Most teams
Team
£40per seat / month
The plan most teams are on. Unmetered, whatever your volume.
- Unlimited seats at the same rate
- 6 months of full-fidelity traces
- Unlimited evaluation runs and datasets
- Human review queues
- A shared chat channel with our engineers
Self-hosted
£40per seat / month
The same build, in your cluster, at the same price.
- Everything in Team
- Runs in your own Kubernetes
- Retention follows your storage
- Quarterly upgrade support
- Air-gapped installs, by arrangement
Every feature on every plan. Tiers differ only in support and deployment.
09 / Six answers for security
Built to be asked hard questions.
We have never trained a model on customer data, and the contract forbids it.
Redaction at the collector
Personal data is stripped inside your network before anything is sent. The rule set is open source and you can add to it.
Self-hosting at the same price
If none of it may leave at all, run the whole platform in your own cluster for the same per-seat fee.
SOC 2 Type II and ISO 27001
Both current. The reports are available under NDA; the summary is not.
EU and US regions
Choose at account creation. Data never crosses between them, backups included.
SSO and SCIM on every plan
Including the free one. A security control is not an upsell.
Four sub-processors, published
Listed with what each one touches, and thirty days' notice before any change.
11 / Contact
Tell us what you are building.
An engineer replies, not a sales rep. The trial does not need a call at all.
Security reviews
The SOC 2 and ISO reports under NDA, and an engineer on the call.
Or write here
Put the next change on the record.
Start with your own staging traffic.
Scene 04 / First light
Frame 084