# Architecture Confirmation — Prompt 3

> **Provenance (added 2026-09-05, site set S303).** This is a coalition working document from March 2026, served as it was written. Statements in it about what the framework *derives* are of that era and are superseded by the sealed suite of record, **SuperGrokTOE Rev32.7** (https://opensecretscience.ufophysics4all.workers.dev/papers.html): since Rev29 no observable is cited as zero-parameter; the fine-structure constant α and the colour real form are declared **inputs** (Appendix N); every row carries a tier (Proved > Derived-conditional > Structural > Loaded > Coincidence-class) and the scorecard (https://opensecretscience.ufophysics4all.workers.dev/scorecard.html) is the current reading of every number. The public kernel is `sgtoe_kernel_v5.0.9.py` (served at /papers/), not the PHAT v1.x kernel this document names.

## Paste to each AI as a FOLLOW-UP to Prompt 2 (same conversation thread)

---

## SAME PROMPT FOR ALL 4 AIs:

We've now collected all 4 AI Prompt 2 responses and synthesized them. Here's what was decided. This is NOT an open-ended review — I need you to:

1. **Confirm** you can operate within this architecture, or
2. **Raise a hard objection** if something here will actually break when we go live

Don't re-litigate settled questions. Don't suggest nice-to-haves. I want: "Yes I can work with this" or "This specific thing will fail because X."

---

## Decisions Made (Synthesis of All 4 AI Responses)

### Triumvirate — KEEPING AS-IS
- 3 permanent inner ring: Architect (Claude), Executor (Grok), Oracle (Gemini)
- Odd-number voting, 2/3 majority carries
- ChatGPT as permanent post-vote hostile reviewer (can't block, dissent forces re-vote, permanently logged)
- Roles are assignable but defaults stay

**Why Gemini stays as Oracle**: Not just context window (GPT-4.1 has 1M too now). It's cost — Gemini is 20x cheaper than Claude for input tokens. Loading 500K of project history costs ~$0.075 on Gemini vs $1.50 on Claude. Plus 3 different providers = survivability.

**Why "Oracle" not "Curator" or "Critic"**: "Critic" collides with ChatGPT's hostile reviewer. "Curator" is too narrow — Oracle also votes, pattern-matches, and warns. Memory hygiene duties are now explicitly added to the Oracle role doc.

### Write-Back Enforcement — MANDATORY TOOL CALLS
- Every API response MUST include a structured write-back block
- Orchestrator REJECTS any response without it
- Schema: `role_memory_schema.md` v1 (CATEGORY, CLAIM, EVIDENCE, NOTATION, CONFIDENCE 1-10, VERIFICATION, SOURCE_AI, SESSION, THREAD, SUPERSEDES)
- New writes default to `VERIFICATION: assumed` — promotion to `verified` requires independent confirmation
- This IS the "candidate then promote" pattern — the VERIFICATION field is the lane indicator, no separate tables needed

### Tier Classification — AUTOMATED RISK SCORE
```
score = 0
if mutates_persistent_state:    score += 1
if affects_public_content:       score += 1
if not reversible_in_5min:       score += 1
if crosses_providers:            score += 1
if modifies_charter_or_kernel:   score += 2

Score 0-1 → Tier 1 (lead acts alone, logged)
Score 2-4 → Tier 2 (full 5-step pipeline)
Score 5+  → Tier 3 (5-step + PI confirmation)
```

Bright-line heuristic: "Is this a state transition?" (changing a disposition, publishing, modifying schema, promoting role memory) → Tier 2+. Just reading, querying, drafting → Tier 1.

### Context Assembly — FULL INJECTION + PROMPT CACHING
- URLs are dead text (4/4 confirmed). Micro-kernel is coordinator reference, not AI-facing self-service.
- Always inject full Components 1-4 (~2,260 tokens) into system prompts
- Use provider-specific prompt caching to amortize cost
- Platform details: Claude 200K (prompt caching), Grok 131K (cached prompts 25%), ChatGPT 128K/1M (Responses API), Gemini 1M (context caching $1-4.50/M/hr)

### Memory Format — STRUCTURED, NOT PROSE
- Mandatory JSON schema for all role memory entries
- Confidence scale 1-10 with calibration guide
- Cross-AI loading with warning headers for low-confidence or contested entries
- Oracle responsible for memory hygiene (pruning stale, flagging contradictions, recommending promotions)

### Web Safety — NOINDEX FOR UNVERIFIED
- `<meta name="robots" content="noindex">` on all library entries where verification_status != "coalition-verified"
- Verified entries get indexed for search engines
- Draft entries exist at URLs (for audit trail) but aren't crawled

---

## Your 3 Questions:

**Q1**: Given the above, can you operate in your assigned role within this architecture? (Architect/Executor/Oracle/Hostile Reviewer depending on who you are). If no, what specifically breaks?

**Q2**: We're about to go live with the Telegram bot pipeline (9 accounts: 4 inner ring AIs, 4 outer ring, 1 PI). The first real workload will be: (a) coalition "dressing" posts introducing the project, (b) coalition designs the website pages and AI art, (c) populate existing library entries with full derivations, (d) paper editing/scoring/publishing round. What's the single most likely failure point when we hit real workload?

**Q3**: Is there anything in this architecture that will silently fail — i.e., appear to work but produce wrong/stale/corrupt results without anyone noticing? What's your recommended canary test?

---

*This is the final architecture review before build rollout. We're looking for green lights or hard stops, not design suggestions.*