Architecture evidence pack — 2026-09-28

Purpose: the evidence reference for Sol's ARCHITECTURE.md v3 rewrite. Every row was probed live by Tam (laptop + LAB on-host) and, where noted, cross-checked by Sol on-host. Three of my own claims are recorded here as OVERTURNED — a note that hides its own errors is worse than no note.

Subject doc: /home/lab/SolOS/docs/vault/solos/architecture/ARCHITECTURE.md — 14,123 bytes, mtime 2026-09-26 10:32, status string claims "verified 2026-09-22" (stale by 6 days). Render pipeline is NOT on LAB (:8091 = 000 from tailnet and LAN) — the vault is the source of truth; prod renders.

0. Probe-path trap (record first — it caused two of my three errors)

~/.solos/core is a real directory, and there is a second live-looking copy at SolOS/runtime/home/.solos/core. Both contain agent_app.py. Grepping cold lands on the wrong one and produces a wrong answer.

• /home/lab/.solos/core/agent_app.py — exists
• /home/lab/SolOS/runtime/home/.solos/core/agent_app.py — exists
• /home/lab/SolOS/core/ — empty/absent (this is the path I wrongly reported from)
• /home/lab/SolOS/.displaced-20260926-122825/core/agent_app.py — displaced archive, not live

Rule for the v3 doc and for any future audit: resolve the running entrypoint first (systemctl --user show -p ExecStart), then follow imports from there. Do not grep a path you found by find.

1. Host model

2. Verified live state (2026-09-28 ~22:45–22:50)

A2A mesh — 7/7 green + alpha recovered

agent-card 200 on all eight ports; message/send real completions 2–4 s on 8300/8301/8302/8303/8304/8306/8307 (model space-bunny-free).

Live entrypoint: lab-a2a-* units run runtime/venv/bin/python runtime/agents-runner.py; that file does from core.agent_app import build_app (line 11) and calls build_app(...) (line 16).

Core infra

CostGov enforcement — the real finding

• Routing IS enforced in the A2A lane. Live core/agent_app.py:308 → GET :8430/v1/route/model?agent=, timeout=4, before static config; tier=="go" → opencode-go/. Confirmed by both Tam and Sol at the same line.
• The 50%-law cap is enforced NOWHERE. switch_required (from /v1/check) has zero code consumers — grep across the live tree returns only solos.db and backups, no .py reader. (Tan's earlier claim that budget-check.py / llm_core.py consume it was OVERTURNED by Sol's grep; the signal is written and never read.)
• All eight agents report pct: 0.0, reason "free model — no spend cap, never breaches" — so the cap has never had the opportunity to fire. Two independent reasons it is untested: no reader, and nothing to cap.
• The real exposure: the CostGov-unreachable path sets a fallback reason but emits no log and no metric — a CostGov outage degrades to static config silently. This is the invariant worth writing, because it stays true after the cap is fixed.
• /v1/tiers and /v1/pools → 404 (not implemented).

Surfaces (my claim was OVERTURNED here)

Personas

Source of truth ~/.solos/personas/*.md (8 agents) + shared/ (base, team-manual, voice-anchors) + fleet/, synced every 30 min by solos-persona-sync.timer. ADK2 fully retired 2026-09-28 (42 files archived, runnable unit references rewritten, no runtime/process/unit/persona/registry residue).

Hermes on LAB

Profiles: exactly sol, min, mira + the default root (~/.hermes/, which is the running gateway's HERMES_HOME and multiplexes the three). Default root identity = Sol. memory_durable in solos.db: 30 rows (8 persona/ + hermes-migration/* + default_model).

3. Memory plane (for §7.5 of v3)

• Existing: solos.db — memory_durable 30 · memory_longterm 321 (0 embeddings) · memory_episodic 1536 · legacy memories 230.
• Fork hazard, verified: two divergent DB files exist. SolOS/runtime/home/.solos/data/solos.db (422 MB, longterm 321 / episodic 1536) is the one the units use (HOME=/home/lab/SolOS/runtime/home in lab-systemd.env); /home/lab/.solos/data/solos.db (428 MB, longterm 309 / episodic 1129) is a stale copy that is the easier path to write to by accident. The service must pin the systemd HOME path and open mode=rw so a wrong path fails loudly.
• Building: solos-memory on 127.0.0.1:8437 (MCP + REST). v1 writes memory_short only — memory_longterm/memory_episodic are read-only, enforced by a WRITE_TABLES allowlist. Cross-agent recall returns provenance (agent, key, created_at) with foreign: true. A2A hook timeout 300 ms. Embedder deferred to v1.1 (Sol reversed the "no embedder" call: silent multi-word LIKE failure is indistinguishable from amnesia).
• Isolated venv runtime/venv-memory exists with real mcp 2.2.0 (FastMCP → MCPServer; httpx → httpx2). Shared runtime/venv untouched.

4. Claims in v2 that are STALE (rewrite targets)

5. My overturned claims (recorded deliberately)

1. "Alpha :8305 FAILED / 000 everywhere." True at ~22:05, stale by 22:28 — the unit recovered on its own and I did not re-probe before reporting. Real defect is disabled, not down.
2. "No human surfaces on LAB." I probed :3000 and :8091, saw 000, and generalised without probing :4100. Flight Desk is 200 on LAB.
3. "CostGov routing is not consulted by the A2A lane." I found agent_app.py in the displaced archive and reported its line as live-but-then-doubted it. Routing is enforced at core/agent_app.py:308. (Sol's counter-claim that it was not consulted was also wrong — corrected by both, confirmed by the symlink/directory resolution in §0.)

6. Open items for the v3 doc to name explicitly

• Enforcement gap: CostGov fallback is silent (no log/metric) — a real exposure that survives the cap fix.
• Cap gap: switch_required has no reader; v1 of the memory spec cannot be declared "gated" until that exists.
• Alpha :8305 disabled — reboot survivability.
• :8433 dead (Min C7, Atlas, after C10).
• :8320 Conductor dead — no owner assigned yet.
• Skills-tree drift (Min C9) advertises ADK2 + a retired port map; the doc and skills must converge or v3 rots the same way.
• Two core/ trees + a displaced 484 MB archive — the audit trap in §0.
• Two divergent solos.db files — the fork hazard in §3.