> ## 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. # What's proven, and where > Every claim in these docs against the run that backs it — and the runs that have not happened yet. **Each row names the run that proves it. Where nothing proves it, the row says so and links what does exist.** Every "proven" cell points at a [receipt](https://github.com/Vexa-ai/vexa-delivery/tree/main/docs/receipts) or a test target you can run yourself. One fact frames all of it: **almost everything below was run on Linode LKE, most of it on throwaway clusters, once each.** Our own production is the only estate with history that this machinery has touched. Your environment is a different one, and the gaps are where you come in — see [running it somewhere new](#running-it-somewhere-new) at the bottom. ## Install and adoption | What | Proven where | Not yet | | ---------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Fresh install on an empty cluster** | Throwaway LKE, several times: preflight PASS, pinned Argo CD v3.5.1 + Kyverno v1.19.0, subscription synced, 12–13 pods Running on digests, the production pin holding at `UNPINNED` until moved by hand — [M2](receipts/2026-08-21-m2-throwaway-test), [MVP0](receipts/2026-08-21-mvp0-implementation), [week-exit rehearsal](receipts/2026-08-24-week-exit-rehearsal) | Any cluster that is not LKE. | | **A whole production estate pulled onto an empty cluster** | Partial. All 115 objects resolved and applied from the channel's signed artifacts alone, images pulled from the channel and **0 from Docker Hub** — [clean pull](receipts/2026-08-25-throwaway-clean-pull) | The estate coming up whole: **16 of 24 workloads reached Running** (3 more Completed). Ten findings are written up in that receipt rather than fixed by it, including the published chart defaulting `database.host` to our own production host | | **Adoption of a live Helm-deployed estate** | Twice, on LKE. A throwaway simulation: 18/18 pods kept their UID, 14 Deployments → 14 ReplicaSets, diff before sync was ownership metadata only — [throwaway adoption](receipts/2026-08-25-throwaway-adoption). Then our own production: 36/36 pods kept their UID, 111 ReplicaSets before and after — [prod adoption](receipts/2026-08-25-prod-adoption) | Any estate that is not ours. The `monitoring` namespace's own three Helm releases were not adopted in either run. **No load was applied during either sync** — both numbers were measured on idle estates | | **Adoption rollback** | Pin move forward then back on the adopted throwaway: 17 s and 10 s, 17/17 pods kept their UID | The full exit — delete the Application `cascade=false`, `helm rollback` — is written down and has never been executed | ## Provider profiles `install.sh` has no per-provider branches; a profile is a small `profile.env`, and each carries a `PROFILE_TESTED` flag the installer prints on every run. | Profile | Proven where | Not yet | | ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `lke` | `PROFILE_TESTED=yes` — every run on this page | — | | `generic` · `aws-eks` · `azure-aks` · `gcp-gke` | Nothing. `PROFILE_TESTED=no`. The per-cluster deltas on [EKS](environments/eks), [AKS](environments/aks) and [GKE](environments/gke) are read from the vendor's own documentation | Never run against a real cluster of that kind. The first subscriber's preflight, smoke and station reports are what changes these lines | | `openshift` | Offline only, and grounded: SCC/PSA/LimitRange logic checked against constraints recorded from a genuine MicroShift 4.18 SCC admission rehearsal — 25 chart objects admitted stock, zero SCC rejects, random-UID contexts injected on all 21 pods — plus a subscriber's own measurements ([parity detail](engineering/openshift-parity)) | `install.sh --provider openshift` has never run on OpenShift. One known defect blocks convergence there: `HOME` is absent under a random UID and the minio-init hook fails with `mkdir /.mc: permission denied` ([OpenShift](environments/openshift)) | ## The delivered software | What | Proven where | Not yet | | -------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- | | **Preflight P1–P9** | PASS on clean clusters, and every failure class seeded deliberately then caught on a throwaway LKE — taints, LimitRange squeeze, PodSecurity, NetworkPolicy, `/dev/shm`, image pull — with the console output kept verbatim beside the [M2 receipt](receipts/2026-08-21-m2-throwaway-test) | The SCC uid-range logic is unit-tested against an audit fixture, never against live OpenShift admission | | **Smoke S1–S2** — delivered set healthy, control plane answers | Live on LKE: Synced/Healthy with 13 pods and both APIs answering over operator port-forwards ([week-exit](receipts/2026-08-24-week-exit-rehearsal)), and a full `vexa_validate.py` run producing a PASS smoke receipt on a real cluster ([channel hardening](receipts/2026-08-25-channel-hardening)) | — | | **Smoke S3 on Teams** — the human loop | Partial. Bot dispatch, the digest-pinned bot image, pod lifecycle and the `ExitReason` reporting are all proven; the join then failed at our own rig tenant's anonymous-join policy ([week-exit](receipts/2026-08-24-week-exit-rehearsal)) | **Capture. No audio has reached the pipeline in any delivery run**, on any platform, so nothing here says anything about transcription | | **Smoke S3 on Google Meet** | The link parser is unit-tested (`make test-smoke`) | A live Meet run. No receipt records one | | **Smoke S4** — flows tier | S4 correctly reports that the flows tier is not in the delivered set, instead of crashing the run as it once did | The tier itself, and the agent/model tier behind it. The question of a model credential has never arisen | ## The upgrade reporter [`kit/report/`](upgrade) is the door for a deployment that already exists. It is **fixture-tested end to end and has never read a real estate.** | Claim | Proven where | Not yet | | --------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------- | | The tool itself | 46 tests against a fake `kubectl`, where the fixture directory *is* the estate (`make test-report`) | No receipt records a run against a real cluster | | It only reads | Every command a run issues is recorded by the fake `kubectl` and asserted to be `get` or `version`. A source scan fails the build if a database client, any SQL, or a flag taking a password appears | — | | `--dry-run` touches nothing | It runs with `subprocess.run` replaced by a raising stub, and is asserted to write no files | — | | `--dry-run` tells the truth | The commands it prints are compared against the ones a real run actually executed. A drifting dry run fails the build | — | | Redaction | A withheld value that survives into the report exits `3`, naming the count and never the value. Two fields shipped named after secrets and were emptied by the blunt rule; the leak scan caught both, and a test now guards the whole class | A leak found in a real estate | ## The channel and its guarantees | What | Proven where | Not yet | | ---------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Signature admission** | Against the live channel with stock Kyverno 1.19 holding no registry credential: signed **ADMITTED**, unsigned **DENIED**, correct layout with the wrong key **DENIED** — [signature layout](receipts/2026-08-25-signature-layout), [authenticated channel](receipts/2026-08-25-kyverno-authenticated-channel). The day before, the same gate denied a correctly signed release as unsigned ([week-exit](receipts/2026-08-24-week-exit-rehearsal)); that is what these two receipts closed | The policy matches `*vexaai/*`, so the mirrored third-party images — caddy, redis, postgres, kubectl — are admitted without verification | | **The in-cluster verify gate** | Both directions on a live cluster: ELIGIBLE on a good entry, and NOT ELIGIBLE with the sync held when the station contract required an evidence kind the entry lacked ([week-exit](receipts/2026-08-24-week-exit-rehearsal)) | It was **not active for our own production adoption sync** — that entry was verified publisher-side instead. And `vexa_channel.py verify` cannot yet verify an estate entry at all ([prod adoption](receipts/2026-08-25-prod-adoption)) | | **Entry expiry, revocation, contract clauses** | Live against a scratch channel on the real registry, with the stale-entry message deliberately distinct from the signature message ([channel hardening](receipts/2026-08-25-channel-hardening)) | The deployed verifier, v0.1.1, does not check freshness — so nothing enforces expiry inside a cluster today | | **Kit bootstrap, verify-before-unpack, self-update** | From an empty directory against the live channel with a pull-only credential: the right key unpacks, the wrong key writes nothing at all, a write attempt returns 401. Self-update ran five times in one day, each time verifying before unpacking ([week-exit](receipts/2026-08-24-week-exit-rehearsal)) | — | | **Mirrored pull** | 43 images mirrored, every destination digest verified equal to its source twice, the estate then pulled with **0 images from Docker Hub**, and blob serving confirmed non-redirecting ([clean pull](receipts/2026-08-25-throwaway-clean-pull)). The pull-through-proxy lane is proven on a Harbor v2.15.2 rig: 9/9 checks through the proxy, wrong key refused ([parity detail](engineering/openshift-parity)) | An install on a **genuinely disconnected** network, as [security](security#isolated-networks) already states | | **The station bundle round trip** | Once, end to end on the live edge and a real cluster: write-scoping enforced per subscriber at the edge, and redaction exercised on a values file that really did carry live credentials — 7 removed, 0 surviving the post-write scan ([channel hardening](receipts/2026-08-25-channel-hardening)) | Concurrency, a large bundle, failure injection, a second subscriber | | **The station gate** | In the direction that matters: it **refused** the real v0.12.23 chart with 8 findings — four dependency images unpinned, two of them on `:latest` — and admits once the upstream fix is applied. The published entry carries the refusal as its evidence ([week-exit](receipts/2026-08-24-week-exit-rehearsal)) | — | ## Day-2 | What | Proven where | Not yet | | ------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- | | **Pull upgrade** | Chart 0.12.31 → 0.12.32 arrived **26 seconds after publish with zero kubectl writes** ([week-exit](receipts/2026-08-24-week-exit-rehearsal)) | One measurement, on a two-node cluster with a warm image cache. It says the transport works unattended; it is not a latency figure | | **Rollback by pin** | On the adopted throwaway: back in 10 s, every pod kept its UID ([throwaway adoption](receipts/2026-08-25-throwaway-adoption)) | On our own production, or on any estate under traffic. Two steps back is a data restore and has not been rehearsed | | **Scale, upgrade under load, failure injection, break-glass** | Nothing | Stated here once so that no row above is read as implying it | ## The from-docs run, 2026-08-27 A fresh anonymous clone, on a throwaway Linode cluster provisioned for the purpose, following [Install](install) and [Upgrade](upgrade) as written. | What | Result | | ------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------- | | Anonymous `git clone`, then the reporter's `--dry-run` | ran, exit 0 | | [Preflight](preflight) on a never-before-seen cluster (LKE, Kubernetes v1.35.7) | **PASS**, all nine checks | | `install.sh` — pinned Argo CD v3.5.1, pinned Kyverno v1.19.0, both ClusterPolicies, the ApplicationSet, both Applications | installed | | Staging synced chart `0.12.35`; production created and parked at `UNPINNED` | as documented | | Postgres, Redis and MinIO admitted and Running under the admission policy | Running | | **The state reporter against that live deployment** | one `state-report.yaml`, 270 lines, exit 0 — [the example on the upgrade page is that file](upgrade) | | Vexa's own images admitted | **denied** — see below | **Two defects the fixtures could not have found, both fixed the same day.** Node capacity printed as `8138636Ki`, and because the grouping key carried that exact figure, two identical machines appeared as two shapes — on a fifty-node cluster that is fifty rows instead of one. Capacities now round to `7.8Gi` and identical nodes group. **And one that mattered more.** Run with the install command the docs carried that morning, every Vexa image was refused by our own policy — `no signatures found` — because without `--signature-repository` the installer strips the repository line from the admission policy and Kyverno looks for signatures beside the image on Docker Hub. The cluster installs clean, Argo reports Healthy, and not one Vexa pod is admitted. The flag is now in the documented command, in `install.sh --help`, and in the onboarding template. With it set, the signatures are found; verifying them needs the channel's public key, which the operator holds. ## Known gaps, and the issue each one lives in Stated so you can pick one up, not so we look complete. | Gap | Issue | | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------- | | **The free channel does not exist.** Every channel is credentialed; the landing's step 3 describes a service that is not running. | [#9](https://github.com/Vexa-ai/vexa-delivery/issues/9) | | **Required values are documented, not enforced.** A deployment missing its transcription configuration reports Synced and Healthy and cannot dispatch a bot; the contract should refuse it instead. | [#6](https://github.com/Vexa-ai/vexa-delivery/issues/6) | | **Only Linode LKE is exercised end to end.** Every other provider profile is declared and never run. | per-provider pages | | **No audio has reached the pipeline in any delivery run.** The smoke step is implemented and never completed on any platform. | — | | **The docs gate depends on the Python patch release.** `gen-cli-reference --check` can disagree with itself between 3.12 patches; the pin is minor-only. | — | ## Running it somewhere new **Running this in an environment nobody here has seen is the contribution.** A row moves from *not yet* to *proven* when someone posts the artifacts that make it true, and most of those artifacts are things a normal install produces anyway: * **The station report.** [`vexa_validate.py`](install#step-5-send-the-station-report-back) writes `station-report.yaml` — one commented file carrying your provider profile, your values with every secret-shaped value replaced by `REDACTED`, the contract, and the preflight and smoke receipts, each with the sha256 of its own text. Read it, then send it. That report is exactly what a `PROFILE_TESTED=no` line is waiting for. * **What broke.** Open an [issue](https://github.com/Vexa-ai/vexa-delivery/issues) with the preflight output and the refusal message. Half the findings on this page arrived that way — from a run that failed, not one that passed. * **A patch.** The [reporter](https://github.com/Vexa-ai/vexa-delivery/tree/main/kit/report) takes a new collector as one function; a provider profile is one small `profile.env`. Everything is Apache-2.0, inbound and outbound. Commits carry a DCO sign-off — `git commit -s`, name and email your own. The whole of it is in [CONTRIBUTING.md](https://github.com/Vexa-ai/vexa-delivery/blob/main/CONTRIBUTING.md), and it is two things: the license and the sign-off. Next: [Install](install) · [Upgrade](upgrade) · [Environments](environments/kubernetes) · [Verify](verify)