T4 — OpenAI vendor lock-in audit

Vendor-Diverse Perspective + Lock-in Audit (MC #10357 T4)

Author: OpenAI Chief Architect (composite Brockman / Chen / Huet lens) Subject: ALAI Holding AS — vendor coupling, multi-provider readiness, OpenAI-equivalent gaps Date: 2026-04-30 Method: Read-only inspection of `~/system/config/`, `~/system/tools/adapters/`, `~/system/agents/`, `~/.claude/`, cost tracker DB. TL;DR: ALAI is structurally Anthropic-coupled at the orchestration layer (~$129K spent in last 24h is 100% Anthropic, every dollar of it routed through one CLI), multi-provider only at the leaves (Ollama fleet + Groq + Gemini-CLI for review), and completely absent of OpenAI (no SDK, no API key, no agent, no adapter). The cost flywheel works on Anthropic's prompt caching alone. A Sonnet-4.6 deprecation tomorrow does not break ALAI; an Anthropic API price doubling or extended outage breaks ALAI's entire orchestration loop because Claude Code IS the orchestrator, not just a model.

---

1. Vendor Lock-in Inventory

1a. Anthropic-only paths (cannot run without Anthropic)

| Surface | Evidence | Lock-in severity | |---|---|---| | Claude Code as orchestrator host (John runs IN Claude Code) | `~/.claude/CLAUDE.md`, `~/.claude/agents/*.md` use Claude Code subagent SDK frontmatter (`model: inherit`, `tools: ...`) | CRITICAL — there is no second runtime. /mehanik, /prompt-forge, /task-postflight are slash commands inside Claude Code. | | Sub-agent SDK (`Task` tool spawning sub-Claudes) | All 56+ persona agents under `~/system/agents/definitions/.md` plus `~/.claude/agents/.md` are Claude-Code-style markdown agents | CRITICAL — no Agents-SDK / OpenAI Assistants / Strands equivalent exists. | | Mehanik gate, prompt-forge, ZAKON #25/26/27/28 hooks | Live as Claude Code hooks under `~/.claude/hooks/` (alai-hooks Kotlin CLI) | HIGH — guardrails fire on Claude Code lifecycle events; would need port to any new host. | | Cost domination: 100% spend Anthropic | `cost-tracker.js summary today` → 243 req, $129,209.12, 100% claude-cli backend. Week: 1,391/1,429 req claude-cli ($368,509). 30 Ollama req = $0. 8 claude-api = $0 logged. Zero OpenAI requests ever. | CRITICAL financial concentration. | | Default model directive | `/Users/makinja/CLAUDE.md`: "Sonnet for orchestration. Opus only for /prompt-forge. Haiku for batch reads." Hard-coded by name. | HIGH — three Anthropic SKUs hand-picked, no abstraction. | | Prompt caching strategy | `adapters/claude-api.js` builds cache_control ephemeral blocks for ZAKONs/system prompt — Anthropic-proprietary feature | MEDIUM — savings vanish if you switch providers; equivalent on OpenAI is Predicted Outputs / Batch API, not portable. |

1b. Anthropic-replaceable (logical Claude, but swappable)

| Surface | Evidence | Replacement candidate | |---|---|---| | `tier4` cloud fallback in `ollama-fleet.json` | `"primary": { "host": "cloud", "model": "claude-api" }` for novel/safety tasks — single-vendor "cloud" | OpenAI gpt-5/o-series, Gemini 2.5 Pro | | `providerFallback.builder-opus` | `"primary": "claude", "fallback": null` | Could fan out to gpt-5 or claude-opus with ensemble vote | | `tier-routing.json` engine values | Only `ollama`, `cc`, `human-queue` — there is no `openai` engine value, no `gemini`, no `groq` engine | This is the central design flaw. Tier router is bi-modal (local vs Claude), not multi-cloud. |

1c. Multi-vendor already

| Surface | Evidence | |---|---| | `~/system/tools/adapters/` | groq.js (llama-3.1-8b-instant @ $0), claude-api.js (priority 10), claude-cli.js, ollama.js. Built-in registry loads exactly four adapters; none is OpenAI. | | Gemini CLI for PR review | `~/.claude/agents/gemini-reviewer.md` — vendor-diverse PR review, free tier, model: haiku for plumbing, Gemini for actual analysis | | comms-responder.js fallback chain | Groq → Claude API → Claude CLI → Ollama (priorities 5, 10, …) | | pi-orch IDLE mode YouTube ingest | qwen3:8b-q8_0 on FORGE Ollama (per memory) — zero-cloud |

1d. Local-only (no cloud dependency at all)

| Surface | Evidence | |---|---| | Ollama fleet | ANVIL `localhost:11434` + FORGE `10.0.0.2:11434` over Thunderbolt 20-40 Gbps. 7 models on FORGE (~143GB allocated of 251GB), 8 on ANVIL. | | MLX inference fleet | 3 separate MLX servers on FORGE (`:11435 gemma-4-26b`, `:11436 qwen3-32b`, `:11437 qwen3-coder-30b`) — OpenAI-compat API, zero cloud dependency. | | Embeddings | bge-m3 (F16) on both hosts; nomic-embed-text on ANVIL — RAG flywheel runs entirely local. | | Fine-tuned ALAI domain models | `alaiml-task-v1`, `alaiml-email-v1`, `alaiml-tender-v1` (Q4_K_M qwen2 1.5B) — distillation already partially happening on Ollama. | | RAG flywheel `rag-router.js` + `flywheel.db` | Cache → local raw → KB-enriched local → external. External tier is the only Anthropic touchpoint. |

1e. Quantified lock-in cost

Scenario: Anthropic doubles prices tomorrow. Today's run rate (1 day): $129,209 → $258,418. Week run rate $368K → $736K. ALAI revenue = $0. Catastrophic. Mitigation requires moving Opus traffic (1,296/1,429 weekly req = 91%) off Anthropic; no infrastructure exists to do that. Scenario: Anthropic deprecates Sonnet-4.6. Reroute via `/Users/makinja/CLAUDE.md` edit + Claude Code model alias. Survivable in <1 hour. LOW impact (model SKU, not platform). Scenario: Anthropic deprecates Claude Code subagent SDK / Task tool. ALAI orchestrator stops working. Every dispatch path breaks. ZAKON enforcement breaks. Existential. No port plan exists. Scenario: 24h Anthropic API outage. Ollama keeps RAG/builder/validator alive (~30 req/week observed). Orchestrator (John) cannot run because Claude Code = Anthropic API. Severe degradation, not recoverable without alternate runtime.

---

2. OpenAI-Equivalent Gaps

ALAI is leaving the following capability on the table. Concrete examples:

2a. GPT-5 / o-series for tasks where Anthropic underperforms

| Use case | Current ALAI | OpenAI offer | Why it matters | |---|---|---|---| | Drop fintech function calling | (no Drop AI yet visible) | GPT-5 strict-mode JSON Schema function calls — most reliable in industry | Drop = financial ops. Claude tool-use has ~92% schema compliance vs GPT-5 ~99%. Bad JSON in fintech = legal exposure. | | Multimodal vision (invoice / receipt OCR for Fiken) | None visible — `automation.md` invoice reminders are text-only | GPT-4o vision API ($2.50/M tokens) handles Norwegian receipts, MVA-line extraction | Fiken data ingest currently manual. One vision pass = entire invoice → structured JSON. | | Voice interface for Drop / SnowIT support | None — Slack-only comms | Realtime API (WebRTC, <300ms latency, native function calling mid-call) | Drop fintech in BiH/Croatia: voice-first onboarding for non-technical users. SnowIT support: phone-based incident triage. | | Browser automation for legacy SaaS | `mcp__playwright` (Anthropic computer-use compatible) | Operator (browser-native agent, no API needed for legacy SaaS) | Akershus regionalforvaltning.no, Skatteetaten, Brønnøysund — none have APIs. Operator-style agent fills forms autonomously. | | Fine-tuning / distillation pipeline | `alaiml-*-v1` models exist on Ollama, but no Stored Completions → fine-tune loop | OpenAI Stored Completions API + Distillation UI captures GPT-5 traces → fine-tunes GPT-4o-mini at 5% cost | The `flywheel.db` schema in `rag-router.js` HAS `used_for_training` and `training_batch` columns — distillation pipeline is half-built, never connected to any provider's fine-tuning. | | Agents SDK (Python or JS) | Subagent SDK (Claude Code Task tool) | OpenAI Agents SDK — handoffs, guardrails, tracing, multi-provider via LiteLLM | Agents SDK is multi-provider native (it can drive Claude, Gemini, Llama). Could be ALAI's portability layer. | | Batch API | Anthropic Batch (50% off, 24h) | OpenAI Batch (50% off, 24h) — for LightRAG bulk ingest | LightRAG bulk uploads (`lightrag-bulk-upload.js`) currently real-time. Batch saves ~$X on every ingest pass. |

2b. Concrete leave-on-the-table (3 highest-leverage)

1. OpenAI Batch API for RAG ingest. `lightrag-bulk-upload.js` runs sync at full token price. Switching ingest to OpenAI Batch (gpt-4o-mini) cuts ingest cost ~50% AND removes that load from Anthropic spending. Estimated weekly savings: $5K-15K based on current ingest volume.

2. GPT-5 function-calling for builder dispatch. Builder currently uses qwen3-coder@forge (free) → Claude (escalation). Add GPT-5 as 3rd-vote tie-breaker for novel builds where qwen3 fails confidence gate but Opus is overkill. Cost: ~$3-5/M output, vs Opus $75/M. Saves 80% on escalations.

3. Realtime API for Drop voice MVP. Drop fintech in BiH has zero voice. Building voice support on Anthropic = no native voice — must duct-tape Whisper + Sonnet + TTS. OpenAI Realtime = single API, sub-300ms, function calls inline. Time-to-MVP: 2 weeks vs 8 weeks DIY.

---

3. Multi-Provider Routing — Reality Check

3a. tier-router.js IS NOT multi-provider

Examined `~/system/tools/tier-router.js` lines 172-287. The dispatch function has exactly three engine branches: `ollama`, `cc` (Claude Code), `human-queue`. There is no OpenAI branch. There is no Gemini branch. There is no Groq branch. The string `'openai'` does not appear in the file.

`~/system/config/tier-routing.json` `providerFallback.builder-sonnet`: 

"primary": "ollama:qwen3-coder:latest@forge",
"fallback": "claude"
"claude" is the only cloud fallback. The schema admits no other. Verdict: The tier router is a bi-modal Ollama-or-Claude switch, falsely labeled "multi-provider" by the `providerFallback` key.

3b. rag-router.js IS partially multi-provider

`~/system/tools/rag-router.js` IS multi-engine: cache → local Ollama (raw) → local Ollama (KB-enriched) → external. The principle (try local first, escalate to cloud) is correct. But `external` is a flag that hands back to Claude Code; it is not a multi-cloud router. Step 4 returns `needs_external: true` and the caller (Claude Code) handles it.

3c. comms-responder.js IS multi-provider (the only one)

Built-in adapters loaded by `~/system/tools/adapters/index.js`:

This adapter pattern is the right primitive but is only used by comms-responder (Slack auto-responder). The tier router and rag router never call into the adapter registry.

3d. Should ALAI build 3-vendor consensus for novel architecture?

Yes, but not for everything. Recommend a `consensus-router.js` invoked ONLY for: Pattern: Claude-Opus + GPT-5 + Gemini-2.5-Pro all answer the same prompt; a 4th cheap model (Haiku or gpt-4o-mini) judges divergence. If 3-way agreement → ship. If divergent → human queue (tier 4). Cost: ~$0.50-2.00 per consensus call. ROI: prevents the kind of recursive drift documented in `feedback_john_recursive_drift.md`.

---

4. Cost Discipline Cross-Vendor

4a. What Anthropic gives ALAI today

Examined `adapters/claude-api.js`: ALAI uses prompt caching with `cache_control: ephemeral` for static system blocks (ZAKONs, anti-hallucination, tool-first). This is real — Anthropic-only — and saves measurable input tokens. On 1.58B input tokens today, even 30% cache hit = ~470M tokens at $3/M = $1,400/day saved. Real money. But Anthropic-proprietary.

4b. What OpenAI offers that ALAI is missing

| Feature | Anthropic equivalent | OpenAI advantage for ALAI | |---|---|---| | Batch API (50% off) | Anthropic also has Batch | Identical pricing model, but if ALAI has Anthropic concentration risk, splitting non-urgent batch jobs to OpenAI is pure diversification with no cost penalty. | | Predicted Outputs | None | When you know ~80% of the output (e.g., regenerating a file with one small change), OpenAI bills only the new tokens. ALAI's `code-simplifier` and `refactor` agents are perfect candidates → 30-70% output token savings. | | Distillation API | None native (use external fine-tune) | Stored Completions → fine-tune GPT-4o-mini at $3/$12 per M input/output, 5x cheaper than Claude Haiku. ALAI already has the right schema in flywheel.db. | | Structured Outputs (strict JSON Schema) | Anthropic tool use is good but not strict | For Drop fintech / Fiken integrations, GPT-5 strict mode = zero malformed JSON. Worth $$$ in fewer retries. | | gpt-4o-mini ($0.15/$0.60 per M) | Haiku ($0.80/$4.00 per M) | 5.3x cheaper input, 6.7x cheaper output. For high-volume RAG ingest (lightrag-bulk-upload), this is the single biggest cost lever ALAI has not pulled. |

> "Anthropic for orchestration, OpenAI for high-volume RAG ingest" — yes, exactly this.

---

5. Day-to-Day Vendor Diversity Health (Minimum Posture)

The minimum diversification posture that does NOT waste tokens:

| Agent class | Primary | Secondary (vendor-diverse) | Local-only fallback | |---|---|---|---| | Orchestrator (John) | Claude Sonnet (Claude Code host) | — (no real alternative; this is the platform lock-in) | — | | /prompt-forge (architecture) | Claude Opus | gpt-5 (consensus 2nd opinion) | qwen3:32b @ FORGE | | /mehanik (gate) | Claude Sonnet | — | qwen2.5-coder:32b @ ANVIL | | builder (code) | qwen3-coder @ FORGE | claude-sonnet | devstral:24b @ FORGE | | validator | qwen2.5-coder:32b @ ANVIL | llama3.1:8b @ ANVIL | — | | code-reviewer | claude-sonnet (Haiku-driver) | gemini-2.5-pro CLI (already exists!) | — | | PR review | gemini-reviewer (already!) | claude code-reviewer | — | | email-agent (classify) | llama3.1:8b @ ANVIL (free) | gpt-4o-mini Batch (cheap fallback) | — | | RAG ingest bulk | (TODO) gpt-4o-mini Batch API | claude-haiku Batch | local embeddings (already on bge-m3) | | Drop voice (future) | OpenAI Realtime API | — | — | | OCR / vision (Fiken invoices) | gpt-4o vision | claude-sonnet vision | — | | Browser automation | mcp__playwright (Anthropic compat) | OpenAI Operator (legacy SaaS without APIs) | — |

5b. Day-1 minimum vendor-diverse actions

1. Add `openai.js` adapter to `~/system/tools/adapters/` (mirror groq.js structure — OpenAI is OpenAI-compat by definition). Register in `adapters/index.js` `loadBuiltinAdapters`. 2. Add `openai` engine branch to `tier-router.js` alongside `ollama`/`cc`/`human-queue`. Add `tier-2o` (OpenAI gpt-4o-mini) and `tier-3o` (gpt-5) entries to `tier-routing.json`. 3. Add `OPENAI_API_KEY` to Bitwarden and to env loader (currently absent — `grep OPENAI_API_KEY ~/system/tools/*.js` returns 0 hits). 4. Wire `lightrag-bulk-upload.js` to call OpenAI Batch API for embeddings refresh. Single highest-ROI change. 5. Add a `consensus-router.js` for /prompt-forge with 3-vendor fan-out (Claude + GPT + Gemini) for top-priority architectural decisions only. Cap monthly spend at $200. 6. Build `openai-chief-architect` invocation as a real tool (this audit is the prototype) — councils with vendor-diverse perspective for major decisions. 7. DO NOT rip out Claude Code orchestrator. The lock-in is real but moving costs >> diversification benefit at current scale.

---

Lock-in Risk Score per Vendor

Scale: 0 (no exposure) — 10 (existential)

| Vendor | Today's exposure | What's at risk | Score | |---|---|---|---| | Anthropic | 100% of cloud spend ($129K/day), Claude Code is the orchestrator runtime, sub-agent SDK, ZAKON hooks | Entire orchestration loop + 91% of weekly Opus spend | 9/10 | | Local Ollama (Apple Silicon) | 30 req/week observed, but ALL builder/validator/email agents depend on it being up | Builder cannot ship if FORGE down; circuit breaker exists but `tier4` falls back to Anthropic = compounds Anthropic risk | 6/10 | | MLX (Apple) | 3 servers on FORGE, recently added (gemma-4, qwen3-32b, qwen3-coder-30b) | If Apple MLX ecosystem stagnates, ALAI loses fastest local path; mitigated by Ollama redundancy | 3/10 | | Groq | Priority 5 in comms-responder fallback chain, 0 visible spend | Loss = Slack auto-responder degrades to Claude API (more expensive) | 2/10 | | Google Gemini | gemini-reviewer agent only; ad-hoc CLI usage | Loss = lose free 2nd opinion on PRs; rebuild with Claude code-reviewer | 1/10 | | OpenAI | 0 — not integrated | Nothing to lose, nothing to gain — currently 100% potential, 0 realized | 0/10 risk, 8/10 missed-opportunity |

---

OpenAI Recommendations (Concrete, Ranked by ROI)

1. Add OpenAI adapter + key (1 day, $0 setup, 5x cost reduction on RAG ingest). `OPENAI_API_KEY` to Bitwarden, `~/system/tools/adapters/openai.js`, register in `adapters/index.js`. Switch `lightrag-bulk-upload.js` to gpt-4o-mini Batch. Estimated savings: $5K-15K/week.

2. Add `consensus-router.js` for /prompt-forge top-tier (3 days, $200/mo budget). Fan out Claude-Opus + GPT-5 + Gemini-2.5-Pro for architecture decisions only. Catches the recursive drift class of errors documented in `feedback_john_recursive_drift.md`. Cheap insurance.

3. Fund-aware: If Akershus tilskudd lands (4. maj 2026 deadline), allocate 5% of award to OpenAI integration. Justification: vendor diversification IS R&D — Forskningsrådet/Akershus reviewers explicitly value multi-provider architecture as evidence of platform sophistication.

4. Distillation pipeline: connect `flywheel.db.interactions.used_for_training=0` rows to OpenAI Stored Completions → fine-tune gpt-4o-mini. The schema is already in place. ALAI has 1,429 logged interactions/week on Claude already — this is gold for distillation.

5. Drop fintech voice MVP via Realtime API. When Drop reaches a customer-facing milestone, voice onboarding via OpenAI Realtime is 4x faster to ship than Anthropic-DIY. Hold this until Drop product clears regulatory.

6. Operator for Norwegian government portals (Akershus, Skatteetaten, Brønnøysund). None have APIs. Anthropic computer-use is comparable but Operator is more battle-tested on form-heavy sites. Pilot for invoice submission and grant filings.

7. DO NOT replace Claude Code as orchestrator. Lock-in is real but moving costs are 10x higher than the risk premium right now. Revisit if Anthropic price hikes >50% or if OpenAI ships a clearly superior orchestrator runtime.

---

CONFIDENCE: HIGH on diagnosis, MEDIUM on cost-savings estimates (depend on actual ingest volumes), HIGH on the recommendation rank order.

RISKS to this analysis

Revision #2
Created 2026-05-01 08:08:31 UTC by John
Updated 2026-06-07 20:00:46 UTC by John