Roadmap
Everything on this site that we answer with “not yet” is on this page, with what it would unblock and what is true in the meantime. It is built from the things this product’s own design notes record as deliberately out of scope for now — not from a wishlist.
This is intent, not commitment. There are no dates on anything that has not shipped, because a date from a team this size is a guess dressed up as a promise. Order changes when customers tell us it should — which is most of the reason the design-partner programme exists.
For what has actually shipped, the changelog is the record — 18 releases so far, most recently “A quickstart page, and docs where you look for them”. That page is dated and this one is not, which is the whole difference between them.
Being built
Work has genuinely started on these. No dates, because a date from a team this size is a guess dressed up as a promise.
Hash-only text storage, with per-org opt-in
What it unblocks: Using the interception path — automatic latency, tokens, errors, and judge-scored quality — without your prompts and responses sitting in our storage in plain text.
Today: The interception path stores prompt and response text raw for your plan’s retention window. The key-only path sends no model text at all, and today that is the answer for data you cannot share. Security →
Recorded in: CONVENTIONS §8 — deferred from v1 as a known privacy trade-off
Bring your own judge API key
What it unblocks: Running the judge under your own provider agreement and your own billing, instead of ours — which is the difference between "we sampled your outputs" and "your account sampled your outputs" in a security review.
Today: Judging runs on the platform judge key, under an agreement that excludes training on submitted data. The judge model is published as a sub-processor on the security page and in the privacy policy. Sub-processors →
Recorded in: CONVENTIONS §8 — customer-managed judge keys, explicitly out of scope for v1
Next up
Decided, designed enough to start, and not started. The order inside this group is the order we currently expect, and it moves when customers tell us it should.
Enterprise: SSO, custom limits, security review
What it unblocks: Buying this at a company where procurement, SSO, and a signed security review are not optional.
Today: The Enterprise tier is marked "coming soon" on the pricing page with a register-interest form. Free, Pro, and Business are fully self-serve today. Pricing →
Recorded in: PRICING_TIERS — Enterprise carries a "Coming soon" badge and a register-interest CTA
SOC 2
What it unblocks: Clearing the questionnaire without a call. Right now the honest answer to "are you certified" is no, and that is a hard stop for some buyers.
Today: Not certified, and the security page says so above the fold rather than leaving you to find out. We will describe the architecture, walk a data flow, or fill in a review document in the meantime. Security →
Recorded in: CONVENTIONS §8 — SOC 2 listed under scope guardrails for v1
Sequential testing
What it unblocks: Looking at a running experiment as often as you like without inflating your false-positive rate — which is what everybody does anyway, and currently has a real statistical cost.
Today: Results are computed over the experiment’s own window with FDR correction across every metric-by-variant comparison. Peeking repeatedly at a fixed-horizon test is still peeking, and the posts on when you can call it say so. When can you call it? →
Recorded in: CONVENTIONS §8 — sequential testing deferred from v1
Considering
Real requests we have not committed to. Being on this list is not a promise that it ships; it is a promise that we have not forgotten it.
Bandits — shifting traffic toward the winner while a test runs
What it unblocks: Spending less traffic on a variant that is clearly losing, on experiments where the cost of exposure is the thing you care about.
Today: Allocation is fixed for the run, which is what makes the comparison clean and the statistics simple to defend. The post on bandits versus A/B testing covers when the trade is worth it. Bandits vs A/B testing →
Recorded in: CONVENTIONS §8 — bandits deferred from v1
Self-hosting
What it unblocks: Running the whole thing inside your own infrastructure, for teams whose architecture is on-premises by policy rather than by preference.
Today: LLMJury is hosted only. The SDKs are open source and Apache-2.0, so the code closest to your application is inspectable, but the service is not something you can run yourself. Self-hosted models →
Recorded in: CONVENTIONS §8 — self-hosting and Helm charts listed under scope guardrails
Automatic capture for more provider clients
What it unblocks: Getting latency, tokens, and judge-scored quality on Bedrock, Vertex, and framework-mediated calls without writing the tracking code yourself.
Today: wrap() intercepts OpenAI-shaped and Anthropic-shaped clients. Everything else works on the key-only path, where you send the measurements you want compared — the integration pages say which is which, per provider. Integrations →
Recorded in: Named as a gap on the Bedrock and Vertex integration pages
Something you need that is not here?
Say so. This list is short because it only contains things already written down as deferred, and the ordering inside it is decided by who asks. Email [email protected] or use the contact form — a founder reads it, within one business day.