📌 The whole architecture in 4 sentences:
Strategic dispatches (Phase X plans, cutover scripts, retire-snippet directives) flow through 5 verification layers before reaching Robert: Scope-Boundary Referee at intake, Protocol Gates pre-load, Per-agent Zeus RAG retrieval, Auditor empirical-grep cross-check, Validation Score binary checks. Bulletproof is the default — never the fast option (CLAUDE.md Rule 9). Tier 2+ dispatches REFUSE Solo mode (architectural enforcement). Multi-model fallback (Claude Opus → GPT-4o) and retry-with-backoff handle platform incidents without human intervention.
Status: 11 of 14 components shipped + verified. 3 deferred to T3-T5 with concrete plans.
v1 (Claude Desktop ↔ Robert relay ↔ Claude Code) had three structural failures:
⏰
Context max-out
Desktop accumulated 4-day threads, hit 420K tokens. Each session re-paid the context cost.
🔁
Relay bottleneck
Robert pasting both sides of the loop — Desktop output → Terminal, Terminal output → Desktop — slowed every dispatch.
🎯
Single-model blind spots
Desktop alone missed what 4 specialized agents catch in parallel. Hallucinations shipped to Robert as "ready."
Plus the meta-failure: the fast path kept producing slop that had to be re-done. Apr 24 functions.php truncation, Apr 25 codebase brief 80× compression, Apr 25 inaugural cutover hallucinations. Each fix took longer than doing it bulletproof first time.
Robert, Apr 25: "for the record we never want the fast option we always want the best in practice bulletproof option" — codified as CLAUDE.md Rule 9.
┌─────────────────────────────┐
│ Robert: dispatch question │
└──────────────┬──────────────┘
▼
╔═══════════════════════════════════════════════════════════╗
║ LAYER 1 — SCOPE-BOUNDARY REFEREE (deterministic Python) ║
║ Maps question keywords → Terminal | Bricks UI | Ambig. ║
║ Bricks UI scope → REJECT at intake (sys.exit 3) ║
║ Terminal scope → proceed ║
║ Ambiguous → flag warning, proceed for Robert review ║
╚═══════════════════════════════════════════════════════════╝
▼
╔═══════════════════════════════════════════════════════════╗
║ LAYER 2 — PROTOCOL GATES PRE-LOAD ║
║ ~12 KB hand-curated identifier inventory + §11 rails + ║
║ scope boundary + 9 named failure-mode shortcuts + ║
║ 11 NEVER-DOs. Loaded into every agent's prompt. ║
╚═══════════════════════════════════════════════════════════╝
▼
╔═══════════════════════════════════════════════════════════╗
║ LAYER 3 — PER-AGENT ZEUS RAG RETRIEVAL ║
║ STRATEGIST: broad query, k=15, similarity ≥0.35 ║
║ CRITIC: failure-mode focused query ║
║ RESEARCHER: external web only (no RAG) ║
║ AUDITOR: identifier-cross-check vs Strategist draft ║
╚═══════════════════════════════════════════════════════════╝
▼
╔═══════════════════════════════════════════════════════════╗
║ LAYER 4 — AUDITOR EMPIRICAL-GREP TOOLS [T3] ║
║ Live snippet census (curl /code-snippets paginated) ║
║ HTML grep for cited element IDs ║
║ File existence stat via SSH ║
║ functions.php sha256 vs locked baseline ║
║ NOT static doc lookup. Live system probes. ║
╚═══════════════════════════════════════════════════════════╝
▼
╔═══════════════════════════════════════════════════════════╗
║ LAYER 5 — VALIDATION SCORE [T4] ║
║ Binary: zero hallucinations (no robert@, no /var/www/, ║
║ snippet IDs in live census) ║
║ Count: cited IDs in step bodies, abort gates per state ║
║ change step ≥1, rollback literal command ║
║ Scope match: Strategist matches Referee verdict ║
║ FAIL = surfaces with "validation failed, here's why" ║
╚═══════════════════════════════════════════════════════════╝
▼
┌─────────────────────────────┐
│ Robert: SHIP / AMEND / │
│ HOLD │
└─────────────────────────────┘
📈 Net result: A dispatch reaching Robert has passed 5 verification gates. His time goes to ship/hold/amend, not proofreading.
| # | Component | What | Status | Turn |
|---|
| 1 | Protocol Gates | Hand-curated 12KB canonical IDs/limits/scope/failure-modes/never-dos | SHIPPED | T0 |
| 2 | Zeus RAG + Context Harness wiring | Per-question semantic retrieval from 18,468 chunks + warnings/blast_radius | SHIPPED | T0 |
| 3 | Retry-with-backoff | All 4 model calls retry on 5xx/429 with 5/15/30s waits | SHIPPED | T1 |
| 4 | Multi-model fallback chain | Strategist Opus → GPT-4o substitute on Anthropic platform incident | SHIPPED | T1 |
| 5 | Scope-Boundary Referee | Deterministic Python keyword mapper at intake. Tier 2 Bricks-UI = REFUSED. | SHIPPED | T2 |
| 6 | Per-agent context tailoring | Each reviewer gets role-tailored RAG (Critic=failure-mode, Auditor=ID-cross-check) | SHIPPED | T3 |
| 7 | Auditor empirical-grep tools | Live snippet census, HTML grep, file stat, sha256 check (not static doc lookup) | T3 | — |
| 8 | Validation Score | Binary + count + scope checks BEFORE surfacing dispatch to Robert | T4 | — |
| 9 | 5-Generation correction sweep | Find every "4th-gen" in today's outputs, fix to canonical "5-Generation" | T4 | — |
| 10 | Cleanup | Delete codebase_brief.py + brief-gen call from project_ledger + stale SOP sections | T4 | — |
| 11 | SOP HTML rewrite | Rename to "Bulletproof-First Operator Manual", document v2 reality | T5 | — |
| 12 | Validation test (cutover re-run) | Re-run inaugural cutover under v2 + diff vs v1 hallucinations | T3.1 in flight | — |
| 13 | MH log v2 architecture | Section bsp-apr25-bulletproof-v2-architecture-shipped (paper trail) | T5 | — |
| 14 | Triad integration | Council fills Nexus Triad's Strategist+Verifier slots. Cockpit = Morpheus Next.js frontend (already live at /app/robert) — NOT Streamlit (deprecated). See Nexus Autonomous Intelligence | Phase 3 | — |
The "bulletproof is the default" principle is enforced at 12 separate touchpoints so it can't be bypassed accidentally:
1
Auto-memory SHIPPED
~/.claude/.../memory/feedback_bulletproof_default.md (14 lines stub) + MEMORY.md index entry. Reloads every session.
2
CLAUDE.md Rule 9 SHIPPED
System-instruction layer. Joins Rule 0 (Web Check), Rule 1 (Producer/Verifier), Rule 4 (Pre-Commit), Rule 6 (Two-Failure Stop), Rule 7 (Log to MH), Rule 8 (Deep Cycle).
3
Protocol Gates §0 Operating Principle SHIPPED
Council agents read on every dispatch. 6-question pre-flight (Apr 21 SOP + bulletproof check).
4
Council runtime Tier 2+ Solo refusal SHIPPED
Architectural enforcement: sys.exit(2). Can't bypass with --mode solo on Tier 2+.
5
Per-agent personas SHIPPED
All 4 agents (Strategist, Critic, Researcher, Auditor) inherit §0 via gates pre-load.
6
6th pre-flight question T3
Added to Apr 21 §13 SOP: "Am I taking bulletproof or fast? If fast, did Robert explicitly approve?"
7
Validation Score bulletproof axis T4
Was retry+fallback used? Was independent reader verification done? Was multi-model verified? Score per axis.
8
Project Ledger compliance % T4
Field added: % of dispatches under bulletproof. Telemetry over time, surfaces in cockpit.
9
SOP HTML title T5
Rename to "Bulletproof-First Operator Manual." Operator manual frame shifts from "how to use" to "how to defend."
10
MH paper trail SHIPPED
Section bsp-apr25-bulletproof-default-rule-9. Forever-record in master history.
11
Pre-commit hook on council_runtime.py T7
CI/CD: refuses commits that introduce API calls without retry-with-backoff.
12
Cross-project propagation T7
Review Nexus cron services (HCP/Daniel/ST attribution) for fast-path defaults, flip them.
ROBERT: "I need a Tier 1 dispatch for retiring snippet #115"
│
▼
┌────────────────────────────────────────────────────────────┐
│ Step 1: Project Ledger refresh │
│ python3 project_ledger.py │
│ Captures: drift sha256, Template 105 element count, │
│ snippet census, recent MH sections │
│ Auto-generates codebase brief (deprecated, kept for now)│
└─────────────────────────┬──────────────────────────────────┘
▼
┌────────────────────────────────────────────────────────────┐
│ Step 2: Council runtime invoked │
│ python3 council_runtime.py --mode council --tier 1 \\ │
│ --question "..." --slug retire-115 \\ │
│ --ledger /tmp/project_ledger_excerpt.md │
└─────────────────────────┬──────────────────────────────────┘
▼
┌────────────────────────────────────────────────────────────┐
│ Step 3: Scope-Boundary Referee │
│ classify_scope("retire snippet #115...") │
│ Match: ['snippet #'] → Terminal scope → PROCEED │
└─────────────────────────┬──────────────────────────────────┘
▼
┌────────────────────────────────────────────────────────────┐
│ Step 4: Context assembly │
│ • Protocol Gates (12 KB pre-load) │
│ • Zeus RAG: 18,468 chunks → top 15 above sim 0.35 │
│ • Context Harness: warnings + blast_radius │
│ • Project Ledger excerpt │
│ • Referee verdict │
└─────────────────────────┬──────────────────────────────────┘
▼
┌────────────────────────────────────────────────────────────┐
│ Step 5: Strategist drafts (Claude Opus 4.5) │
│ max_tokens: 16000 (no truncation) │
│ Retry-with-backoff on 529/5xx │
│ Fallback to GPT-4o if Anthropic down │
└─────────────────────────┬──────────────────────────────────┘
▼
┌────────────────────────────────────────────────────────────┐
│ Step 6: Parallel review (3 agents, role-tailored RAG) │
│ Critic (GPT-4o): failure-mode RAG + draft review │
│ Researcher (Sonar): external web verify (citations) │
│ Auditor (Gemini): identifier-cross-check RAG + draft │
│ (Pro→Flash fallback) │
└─────────────────────────┬──────────────────────────────────┘
▼
┌────────────────────────────────────────────────────────────┐
│ Step 7: Synthesis + receipts │
│ Dispatch markdown saved to │
│ /opt/nexus/.../dispatches/phase__.md │
│ Per-agent raw JSON saved to │
│ /opt/nexus/.../council/raw__.json │
│ MH section logged automatically │
└─────────────────────────┬──────────────────────────────────┘
▼
┌──────────────────────────────┐
│ Robert: SHIP / AMEND / HOLD │
└──────────────────────────────┘
🧠
STRATEGIST — Claude Opus 4.5
Authors Phase X dispatches in BSP house style. Default voice. Fallback: GPT-4o if Anthropic down. Max tokens: 16,000 (truncation-proof). RAG: broad query k=15.
⚔️
CRITIC — GPT-4o
Devil's advocate. Numbered risk list with severity. Vote: APPROVE / AMEND / REJECT. RAG: failure-mode focused query. Brutally specific — "it might not work" is rejected.
📚
RESEARCHER — Perplexity Sonar Pro
Citation-first external evidence. Verifies factual claims with 2024-2025 sources. RAG: none (external web grounding only). Reject opinion-only sources.
🔍
AUDITOR — Gemini 2.5 Pro
Cross-checks dispatch against MH + Project Ledger + codebase doc. Empirical grep [T3] for cited identifiers. Fallback: 2.5 Flash if Pro hits MAX_TOKENS. RAG: identifier-cross-check on Strategist draft.
🛠️
EXECUTOR — Claude Code (this agent)
VM SSH + Hostinger MCP + file ops + native-save / install-child / cache purge / Playwright verify. Runs the dispatch end-to-end after Robert ships. Single integration point — no copy-paste relay.
| Tier | What | Examples | Mode required | Robert approval? |
|---|
| TIER 0 | Read-only / docs / MH log | SEO research, project ledger refresh, doc draft, MH section | Solo OK | None — auto-execute |
| TIER 1 | Staging writes with rollback | Template 105 native-save, snippet activate, child theme install, op258 patches | Solo OK (or Council) | Auto-execute, rollback ready |
| TIER 2 | Production writes / content tree change / schema | Production cutover, callbrightside.com edits, sitemap regen, GMB updates | Council REQUIRED | Robert approval REQUIRED |
| TIER 3 | Destructive / irreversible | Drop snippet permanently, delete Bricks template, force-push to main, drop DB table | Council REQUIRED | Robert approval + signed receipt |
Architectural enforcement: Tier 2+ refuses Solo mode. Try it: --mode solo --tier 2 exits with code 2 + redirect message. Bulletproof Default rule cannot be bypassed by accident.
Each happened during Apr 24-25. Each is now codified as a never-do across all 12 enforcement layers.
- Fast option chosen when bulletproof exists. Default to bulletproof unless Robert explicitly says "go fast."
- Single-model verification of state-changing claims. Tier 2+ requires Council mode — multi-agent crosscheck.
- Producer == Verifier collapse. The script that performs a change is never the authority on success. Independent reader.
- No retry on transient API failures. All 4 model calls have retry-with-backoff (5/15/30s).
- Skipping pre-flight gates. 6-question test (Apr 21 SOP + bulletproof check) before any code.
- Truncating output to fit token budget. Strategist max_tokens 16,000 (was 8,000 — caused Apr 25 cutover rollback truncation).
- Regex string-replace on production files. Use
grep -n + targeted Edit with ≥20-char unique anchors (Apr 24 truncation incident).
- Citing identifiers without verification. Auditor empirical-grep [T3] cross-checks every cited element/snippet ID against live system.
- Claiming done without receipts. CLAUDE.md Rule 2: literal tool output in code block. No narration.
- Local fix instead of class-of-problem fix. Mario: "agents have only local view." Zoom out for systemic answer.
- Code shipped before tests. Test-after-each-layer. Smoke test before next component.
- Documentation written before system works. Aspirational features documented as factual = Apr 25 SOP mistake.
- Outsourcing QA to Robert by default. Playwright + audit + diff before reporting (§11 self-limits).
Per Mario Zechner's "Slow the fuck down" framework (March 2026 essay) — applied to BSP coding sessions:
| Limit | Number | Why |
|---|
| Lines of code per turn | ≤ 200 net new | Code I can't read in 2 min is code I can't defend. |
| New snippets per session | ≤ 3 net new | Each snippet = new surface to maintain. |
| Element tree size per build | ≤ 150 elements per page | Bigger = split into templates + sections. |
| Patches before rebuild | ≤ 3 patches to a section | >3 = underlying tree wrong. Rebuild from Figma. |
| QA cycles outsourced to Robert | 0 by default | Playwright + audit + diff before reporting. |
| Diagnostic snippets left active | 0 (delete within 24h) | Slop until removed. |
| Memory file size | ≤ 20 lines per file | Stub form only. Full content in MH. (Apr 25 addition) |
The check before any deploy (§11):
- Recall: Did I search existing for prior art? Extending vs duplicating?
- Local-vs-global: One-element patch or solving the underlying class?
- Defensibility: Can I explain to Kalen in one sentence why each new line of code exists?
- Verification: Did I run Playwright + audit + diff before reporting?
- Cleanup: What am I leaving behind (diagnostic snippets, dead helpers)?
- Bulletproof check (Apr 25): Am I taking bulletproof or fast? If fast, did Robert explicitly approve?
If ANY answer is "no" or "I don't know" → STOP. Do not write code.
Refresh project ledger + codebase brief
ssh dovew@34.55.179.122 "python3 /opt/nexus/nexus/scripts/project_ledger.py"
Solo dispatch — Tier 1 tactical work
ssh dovew@34.55.179.122 "python3 /opt/nexus/nexus/scripts/council_runtime.py --mode solo --tier 1 --slug short-name --question 'YOUR ASK' --ledger /tmp/project_ledger_excerpt.md"
Council dispatch — Tier 2+ high stakes (refuses Solo, this is the only path)
ssh dovew@34.55.179.122 "python3 /opt/nexus/nexus/scripts/council_runtime.py --mode council --tier 2 --slug short-name --question 'YOUR ASK' --ledger /tmp/project_ledger_excerpt.md"
Read most-recent dispatch
ssh dovew@34.55.179.122 "cat \$(ls -t /opt/nexus/nexus/scripts/output/dispatches/*.md | head -1)"
Pull dispatch to Windows for Notepad review
scp -i ~/.ssh/google_compute_engine "dovew@34.55.179.122:\$(ssh dovew@34.55.179.122 'ls -t /opt/nexus/nexus/scripts/output/dispatches/*.md | head -1')" C:/Users/dovew/Downloads/
Test scope-boundary referee on a question
ssh dovew@34.55.179.122 "python3 /opt/nexus/nexus/scripts/referee.py 'YOUR ASK'"
Correction (Apr 25): Robert flagged: "we don't use Streamlit anymore." Cockpit is the Morpheus Next.js frontend (live at https://morpheus.callbrightside.com/app/robert) plus the document library at /documents/*. Streamlit dashboards were deprecated. Adjusting deferred items below.
Phase 3 — integrate into Nexus Autonomous Intelligence
Per BSP_Nexus_Autonomous_Intelligence.html, the broader system is already running with: 6,332 RAG chunks, CRAG (Corrective RAG quality gate), Self-RAG (self-reflection), 18 antibodies (only 3 firing — gap documented), nexus_priority_engine.py at port 8765 with 350 endpoints, the Triad architecture (Strategist + Verifier + Executor), Morpheus Next.js cockpit, daily 5AM priority queue ranking. The BSP Operator Council fits inside this as a multi-model implementation of the Triad's Strategist+Verifier slots.
- Slot the council into the Nexus Triad — Council Strategist+Auditor become Triad's Strategist+Verifier roles, sharing the existing CRAG / Self-RAG / antibody machinery
- Surface dispatch queue in Morpheus frontend — add a route under
/app/robert for council dispatch queue + Validation Score history (replacing the Streamlit cockpit idea)
- Promote 4 deferred deliverables (T4-T7) into Nexus antibodies — the 6th pre-flight question, Validation Score, Project Ledger compliance %, pre-commit hook map cleanly onto Nexus' antibody pattern + would lift firing rate from 3/18 toward 7/18
- Cross-project propagation — same council harness for HCP / Daniel AI / ST attribution work, leveraging Zeus RAG cross-domain chunks
- Telemetry into nexus_priority_engine.py — bulletproof compliance % feeds the daily 5AM priority queue scoring
- Daily cron loop — drift check 06:30, ledger refresh 07:00, daily brief 07:30, end-of-day MH snapshot 17:00 (slot into existing Nexus cron)
v2.0 · 2026-04-25 · BSP Bulletproof Council Architecture
Single-page document. No external dependencies. Open in any browser.
Sources: protocol_gates.md, council_runtime.py, referee.py, project_ledger.py, codebase_brief.py [deprecated], BSP_Master_Session_History.html
BSP_Bulletproof_Council_Architecture_v2.html