Server Phase 1 Validation Evidence

Validation date: 2026-07-12 to 2026-07-13

Target: Vast.ai full Ubuntu 22.04 VM, RTX 3090 24 GB, 23 vCPU, 52.8 GB usable RAM, 155.9 GB root filesystem

Package revision at initial install: 55a2a41 plus the uncommitted corrections listed below

Install mode: install.sh --software-only --tier=starter

Executive result

The software package installs and the tested Phase 1 Server workflows operate on the representative cloud VM after correcting defects found during validation. This is evidence for a controlled Server pilot only. It is not ISO boot evidence, physical hardware evidence, Workstation evidence, a customer pilot, WCAG manual approval, or general-release approval.

Results

Area	Result	Evidence
Starter installation	PASS after Docker repository fix	Ansible recap: 86 ok, 51 changed, 0 failed, 2 ignored
Source contract suite	PASS	19 tests passed
Shell/Python/static validation	PASS	`bash -n`, `py_compile`, `git diff --check`
Tier playbook syntax	PASS	starter, entry, pro, max syntax checks
Portal/backend	PASS	HTTP 200 after install and reboot
Authenticated read endpoints	PASS	51 HTTP 200 responses
Entitlement restrictions	PASS	model router and two fine-tuning routes correctly returned 403
Authentication and RBAC	PASS	admin login; viewer permitted metrics and denied admin user list
User/session lifecycle	PASS	create, login, revoke by deletion, expired-session rejection, self-delete protection
API key lifecycle	PASS	create, verify, revoke, revoked-key rejection
Guardrails	PASS	temporary keyword rule blocked matching input and was removed
Prompt library	PASS	create, update and delete
Backup/restore	PASS	ZIP created; listed; restored; pre-restore safety snapshot created
Ollama inference	PASS	Qwen 2.5 1.5B GPU inference; model persisted across reboot
Open WebUI	PASS after network correction	healthy container and model visible from WebUI
RAG	PASS after corrections	collection, TXT upload, embedding, ready status, semantic retrieval, delete/cleanup
vLLM	PASS after corrections	Phi-3 readiness and OpenAI-compatible chat completion
Monitoring	PASS	Grafana health 200, Prometheus ready 200, DCGM and node exporter running
Web console	PASS after correction	service active and `/console/` HTTP 200
Reboot persistence	PASS	all six services and five containers auto-started; models and backup persisted
Concurrent inference	PASS	vLLM used 21.3 GB and Ollama 1.8 GB on RTX 3090; both completed inference

Post-reboot vLLM cold readiness occurred at approximately 144 seconds after boot.

Defects discovered and corrected in source

Ubuntu ttyd does not support --writable, and the distribution ttyd service can occupy port 7681.
Open WebUI could not reach an Ollama process that had not reloaded the appliance service environment.
Backend targeted ChromaDB port 8000 while the service intentionally runs on 8100.
Unpinned latest ChromaDB removed the v1 API used by the backend.
Compatible ChromaDB dependencies conflicted with the shared AI environment; ChromaDB is now isolated.
Short RAG documents caused an infinite chunking loop.
Chroma query/upsert/delete data routes require collection UUIDs rather than collection names.
Starter Phi-3 was incorrectly configured as AWQ.
Blank optional vLLM arguments caused systemd command continuation into ExecStop.
vllm:latest required an incompatible CUDA/driver combination; the image is now pinned to v0.10.2.
vLLM was enabled but not explicitly started during installation.

Expected warnings / environment gaps

No final signed customer license was installed; field-staging mode was used.
TLS was not configured; access used SSH tunnels.
The VM began without a backup or models; both were created during testing.
Direct public port 80 was blocked by the provider; this is not an appliance Nginx failure.
Clean ISO boot, disk partitioning, BIOS/UEFI, USB media, physical NICs and offline ISO installation cannot be validated on this rented VM.
Workstation validation (T036), manual WCAG 2.2 AA evidence, interruption/restricted-network trials, controlled customer pilot and release approvals remain pending.

Release decision

Not ready for customer shipment yet. The corrected package is suitable for another clean Server installation validation and controlled pilot preparation. A new ISO must be built from the corrected revision and both Server and Workstation ISO paths must be tested before release consideration.

Corrected-package clean-host regression — 2026-07-13

A second independent Vast.ai Ubuntu 22.04.5 VM was installed from the corrected local package:

RTX 3090 24 GB, NVIDIA driver 580.159.03 after upgrade/reboot;
30 vCPU, 52.9 GB usable RAM, 126 GB root filesystem;
clean starter install recap: 88 ok, 63 changed, 0 failed, 3 ignored;
the ignored DCGM failure was caused by the required post-driver-upgrade reboot and passed afterward;
Grafana provisioning was retried after its startup race and the Prometheus datasource was added;
portal source permissions were normalized after clean-host validation exposed preserved macOS mode/UID values.

Regression results:

51 authenticated parameterless feature endpoints returned 200;
3 unlicensed features correctly returned 403;
RBAC, user/session, API-key, guardrail, prompt, backup and restore cases passed;
short-document RAG ingest, embedding, semantic retrieval and cleanup passed;
vLLM Phi-3 OpenAI-compatible inference passed;
Grafana, Prometheus, DCGM and node-exporter endpoints returned 200;
audit report contained the expected test actions with zero recorded failures;
after a final reboot, all six system services and six containers recovered;
portal, backend, Ollama, ChromaDB, Grafana, Prometheus and DCGM returned 200;
the embedding model and backup archive persisted;
vLLM was ready by approximately 100 seconds after boot.

This confirms the corrected software package on two Server-class cloud VMs. It does not change the remaining ISO boot, physical hardware, Workstation, manual accessibility, network-interruption, customer-pilot or release-approval gaps above.

6.2 KiB Raw Blame History