---
title: Where each claim stops
description: >-
  The platform words a reader is most likely to over-read, what each one is
  bounded by here, and where that bound is set in the product rather than in
  prose.
tags:
  - concepts
  - honesty
  - compliance
  - orientation
---

*The platform terms a reader is most likely to over-read, the bound each one actually carries here, and where that bound is enforced.* Every claim on this page is narrower than the industry word for it.

[Glossary](/docs/glossary) says what a term means. This page says what it does **not** mean. The two are deliberately separate: a definition that carried its own disclaimer would be read as hedging, and a disclaimer filed on the page that would rather not mention it is a disclaimer nobody finds. Collected here, the bounds are citable — you can quote one into a review without trusting the page that made the promise.

For how to read a *prediction* against the bench, read [Reading an in silico result against the bench](/docs/translatability) instead. That is a different question, it has a per-tool page for every predicting tool, and this page does not duplicate them.

## The terms, and what bounds them

Each row names where the bound lives. A bound set in code or in a store is a bound you can check; a bound that exists only in a sentence is one we would have to keep remembering.

| Term | What a reader usually assumes | What it means here | Where the bound is set |
|---|---|---|---|
| **Immutable** (checkpoint) | It can never change or disappear | A saved checkpoint is not rewritten — and it is not kept indefinitely | `SupabaseSaver` prunes after each write, keeping the newest 50 per session ([Checkpoints](/docs/api/sessions-checkpoints)) |
| **Tamper-proof** (the record) | Modification is prevented | It is tamper-**evident**, not tamper-proof: it detects modification; it does not prevent it, and it is not access control. | The hash chain and the verifier ([Audit trail](/docs/audit-trail)) |
| **Audit trail** | Everything, for everyone, forever | Scoped by who ran it and from where: signed-in runs from the web app, the terminal, curl or MCP leave a tamper-evident record you can check, and so does a guest chat, from the browser tab it is open in, while the guest session lasts. | Chain writes keyed per chat or session; keyless guest `curl` calls write nothing durable |
| **Open source** | You can read the code today | FSL-1.1-Apache-2.0 licensed; source release planned. | Licence terms; the repositories are not publicly readable |
| **Zero data retention** | No prompt is ever retained, for anyone | A guest-tier routing control, not a platform-wide one: for PHI- or IP-sensitive prompts, guest LLM requests route only to endpoints that do not retain prompts (zero data retention, or ZDR), with provider data collection denied. For everyone else, signed-in OpenRouter requests deny provider data collection but are not restricted to ZDR endpoints. | Provider routing per dispatch |
| **Compliant** | Someone external certified it | Nothing here is externally certified, and these controls do not reach that far: these provider-routing controls do not cover research data stored in the platform or establish HIPAA compliance. | See below |
| **Complete** (passport) | 100% means everything knowable was found | A measure of how much data the run found across its sections. A low score marks thin sources, not failure | The completeness score, per section |
| **Reproducible** (docking) | Same inputs, same poses | DiffDock exposes no seed and samples stochastically, so re-running the same inputs is not guaranteed to return the same poses | Sampling is stochastic; raise `samples_per_complex` to cover more of the variability in one run ([DiffDock inputs](/docs/diffdock/inputs)) |
| **Larger model** | More parameters, better embeddings | Not always a different model: on this platform `esm3_lg` is byte-identical to `esm3_md` | The deployed weights ([ESM-3 translatability](/docs/esm3/translatability)) |
| **Retired** (a tool) | Its results are gone too | The tool cannot be called; existing job results stay viewable | Per-tool retirement notes, e.g. [ADMETlab 3.0](/docs/admetlab3) |
| **Erased** (GDPR request) | Every trace is deleted | Your data is deleted table by table; the hash-chained logs are retained, and the erasure report says so | Erasure procedure ([Audit trail](/docs/audit-trail)) |

## Nothing here is externally certified

No page in these docs claims a SOC 2 report, a HIPAA attestation, an ISO certificate or a third-party penetration test, and that is deliberate rather than an omission. Every security property described in this tree is **self-described**, which is why the audit trail is built to be checked by you instead of vouched for by us: you can open [/audit/verify](/audit/verify), pick one of your own chats, and recompute the chain yourself.

Read the absence of a certification claim as the claim. A compliance badge sprinkled across a documentation tree is unfalsifiable from the outside; a verifier you can run is not.

## Why the bound travels with the promise

The record claim is the clearest case, so it is worth following once. It is stated in three parts, always in this order, and never one without the next:

1. **The claim** — signed-in runs leave a tamper-evident record you can check.
2. **The failure mode** — edit a row and the chain breaks at the verifier.
3. **The boundary** — it detects modification; it does not prevent it, and it is not access control.

Part 3 is what keeps part 1 true. Detection is not prevention, and neither is access control, so a reader who takes "tamper-evident" to mean "nobody can change it" has been given the wrong thing to rely on. The three parts are stored once in the product's claims module and composed into each page by name, because a claim that is retyped on nine pages is a claim that is wrong on eight of them the day it changes.

That is also why the rows above cite a store, a licence or a sampling behaviour rather than a promise. A bound enforced where the work happens survives a rewrite of this page.

## Where this page does not reach

- **Per-tool scientific limits are not here.** Each predicting tool carries a `translatability` child page stating what its numbers were graded against, with citations. Those pages are the ones that say where a given model's figures stop, and several of them cut hard against the tool they describe.
- **Known wrong answers are a per-tool disclosure, not a vocabulary question.** Where a tool is known to return a specific class of bad result, that belongs on the tool's own page where a reader about to run it will see it — not collected here where it would read as a curiosity.
- **This page states bounds; it does not state status.** For what is live, degraded or retired, see the tool page and [Security &amp; compliance status](/help/compliance).
- **Gaps are reported, not filled.** When a run cannot complete a summary, gaps stay gaps — when a summary leaves rows out, the run says so instead of guessing them.

## What this page gives you

One place to quote from. If you are putting the platform through a review — procurement, a quality system, an IRB, or your own judgement about whether to trust an answer — the bound you need is on this page in the vendor's own words, with a pointer to the mechanism that enforces it. You do not have to find the caveat buried three pages deep, and you do not have to take our adjective for it.

## Related

- [Glossary](/docs/glossary) — what each term means
- [Reading an in silico result against the bench](/docs/translatability) — how to read a prediction
- [The audit trail of a run](/docs/audit-trail) — what the chain catches, and what it does not
- [21 CFR Part 11: what the record covers](/docs/audit-trail/part-11)
