> ## Documentation Index > Fetch the complete documentation index at: https://delivery.vexa.ai/llms.txt > Use this file to discover all available pages before exploring further. # Acceptance > What must be true before a version of this software is distributed to anyone. **Nothing ships to a channel or gets announced until every line below has been executed and its output kept.** Each row names the command that proves it. "We checked" is not a proof; a pasted verdict is. Every failure this contract exists to prevent has already happened once, on 2026-08-27, in the hour before the first send. ## 1 · The stranger test The only test that matters: someone with no relationship to us follows the docs and gets a working deployment. | Prove | How | | ------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------- | | A fresh anonymous clone works | `git clone https://github.com/Vexa-ai/vexa-delivery` from a machine with no credentials | | Every command in `install.mdx` and `upgrade.mdx` runs **as written**, in order, on a cluster provisioned for the purpose | run them; keep the console | | The reporter runs read-only against a real deployment | `python3 kit/report/vexa_state_report.py --namespace --dry-run`, then for real | | Preflight passes on that cluster | `python3 kit/preflight/vexa_preflight.py --namespace ` → `VERDICT: PASS`, exit 0 | | The subscription installs and syncs | `./kit/install.sh …` with the exact flags the docs print — including every flag the docs omit at your peril | **A command that needs a flag the docs do not show has failed this section**, even if it works when you add the flag. The reader does not have your shell history. ## 2 · Docs are true of the code | Prove | How | | -------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------- | | Every flag shown exists, spelled as shown | `--help` on each tool, diffed against the docs | | Every exit code stated matches the source | grep the exit paths | | Every verdict promised is the verdict produced | run the command and read the last line — a doc that promises `FAIL` and delivers `PASS` costs more than a missing page | | Every component the docs say runs, runs after the documented install | check the cluster, not the intent | | Every number cites a measurement in the tree or a receipt | no figure survives without a source | | Generated pages regenerate clean on **CI's interpreter** | `python3 docs/gen-cli-reference.py --check` under the version `.github/workflows/ci.yml` pins | ## 3 · The channel serves what the repo says | Prove | How | | -------------------------------------------------------------------------------------------- | ---------------------------------------------- | | The published kit is built from the tag being announced | pull it, unpack it, `diff -r` against the tree | | Its digest matches what the entry names | `oras pull` and compare | | Its signature verifies **against the real channel key, by someone who is not the publisher** | `cosign verify --key channel.pub …` | | Every subscriber channel carries the version being announced | list the tags | **A stale kit in a live channel is a shipped defect**, not a housekeeping item: the docs describe one artifact and the subscriber receives another. ## 4 · Honesty | Prove | How | | --------------------------------------------------------- | ------------------------------------------------------------- | | [What's proven, and where](/tested) reflects this version | every row's evidence link resolves and says what the row says | | No claim exceeds its evidence | anything unproven carries a Rung note naming what is missing | | Every known gap is listed with its issue | the gaps table, not a private list | | Nothing describes a service that is not running | grep the docs for what we wish were true | ## 5 · Nothing leaks | Prove | How | | --------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- | | No customer names, hostnames or identifiers | `grep -riE ''` across the tree, including receipts and generated pages | | No private repository paths, no personal namespaces, no local machine paths | the same sweep — this class is separate from customer tokens and was missed by three earlier sweeps that only looked for the first | | No credentials, keys or `.env` files beyond examples | pattern sweep for PEM blocks, `AKIA`, `ghp_`, `sk-` | | No editor or build state | `git ls-files` for `.obsidian`, caches, backups | ## 6 · What we ask a subscriber for The report a subscriber sends is the only thing a bundle is built from, so its coverage is **argued, not assumed**. | Prove | How | | -------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Every input that determines a bundle is either supplied by the report, a customer secret, or a named gap | enumerate them from the values file, every provider profile, every preflight check, every installer flag, the chart's `.Values` references and the contract fields — and keep the table | | Nothing unreadable is silently missing | anything that cannot be read read-only appears in the report as a stated question, with its reason | | A gap found by one lens is not the end of the search | 2026-08-27: five gaps found against preflight, two more against the values surface an hour later, five more against a full sweep — including Argo CD in `openshift-gitops`, which would have been missed at the one OpenShift account | ## 7 · Teardown | Prove | How | | ---------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | A throwaway environment leaves nothing running | list instances, volumes and load balancers **with a token that can see them** | | Nothing was verified inside too small a scope | 2026-08-27: an LKE cluster was deleted, its two node VMs kept running, and the check said clean — because the scoped token could not see them. They were only visible under the owner token | **Deleting the parent does not reap the children, and a check that cannot see a thing is not evidence the thing is absent.** ## 8 · Gates All green, on CI's interpreter, in one run: ```bash theme={null} make test python3 docs/check-docs.py python3 docs/gen-cli-reference.py --check cd docs && npx mint validate && npx mint broken-links ``` **A red gate is a stop, not a note.** On 2026-08-27 twenty commits landed past a docs gate that had been failing since the morning, because the failure was familiar. ## What this does not cover Smoke with a human in a real meeting has never completed on any platform, so it is not yet a gate — it is the first thing that becomes one when it passes. Provider coverage beyond Linode LKE is the same: the contract requires the stranger test on *a* cluster, not on every cluster, and [what's proven, and where](/tested) carries the difference.