From 631b0329b19b1b6335a174754e686eca7290c585 Mon Sep 17 00:00:00 2001 From: root Date: Mon, 27 Apr 2026 23:13:46 -0500 Subject: [PATCH] =?UTF-8?q?demo:=20proof=20=E2=80=94=20full=20architecture?= =?UTF-8?q?-page=20rewrite=20for=20current=20state?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit J: "needs a rewrite." Old version was anchored on a dual-agent mistral+qwen2.5 loop that hasn't been the model story for weeks, called the system 13 crates (it's 15), referenced "Local 7B models" in the honest-limits section, and had no mention of: - the 40-model OpenCode fleet via one sk-* key - the 9-rung cloud-first ladder - N=3 consensus + cross-architecture tie-breaker - auditor cross-lineage (Kimi K2.6 ↔ Haiku 4.5, Opus auto-promote) - distillation v1.0.0 frozen substrate (e7636f2) - pathway memory (88 traces, 11/11 replays, ADR-021) - per-staffer hot-swap index - Construction Activity Signal Engine + BAI + ticker network - the gateway as OpenAI-compat drop-in middleware Rewrote into 10 chapters: 1. Receipts — live tests + new live tile showing the Signal Engine view for THIS load (issuer count, attributed build value, contractor count, attribution edges) 2. Architecture — corrected to 15 crates with current responsibilities; ASCII diagram showing OpenAI consumers + MCP + Browser all hitting gateway /v1/*; provider fleet table with all 5 (ollama, ollama_cloud, openrouter, opencode 40-model, kimi); validator + truth + auditor crates added 3. Model fleet — REPLACED the dual-agent mistral story. Now: the 9-rung ladder (kimi-k2:1t through openrouter:free → ollama local), N=3 consensus + tie-breaker math, auditor Kimi↔Haiku alternation with Opus auto-promote on big diffs, distillation v1.0.0 freeze tag e7636f2 (145 tests · 22/22 · 16/16 · bit-identical) 4. Two memory layers — kept playbook content (Phase 19 boost math still load-bearing), added pathway memory (ADR-021) section with live counters in the page (88 / 11-11 / 100% reuse rate) 5. Per-staffer hot-swap — NEW. Pseudocode showing how staffer_id scopes state filter + playbook geo + UI relabel to MARIA'S MEMORY 6. Construction Activity Signal Engine — NEW. Three attribution flavors (direct, parent, associated), BAI math, cross-metro replication framing (NYC DOB next, then LA / Houston / Boston) 7. Architectural choices — added ADR-021 row + distillation freeze row 8. Measured at scale — kept (uses /proof.json scale data) 9. Verify or dispute — REFRESHED with current endpoints. Removed the stale "bun run tests/multi-agent/scenario.ts" recipe; added curl examples for /v1/health, pathway/stats, per-staffer scoping (3-loop bash script), late-worker triage, profiler_index, ticker_quotes, auditor verdicts, distillation acceptance gate 10. What we are NOT claiming — REFRESHED. Removed "Local 7B models" caveat; added: 12 awaiting public-data sources are placeholders, SEC name-fuzzy has rare false positives, BAI is a thesis not a backtest yet, single-metro today Live data probes added: loadPathwayLive — fills pwm-traces / pwm-replays / pwm-rate spans loadSignalLive — renders the LIVE Signal Engine tile under Ch1 Nav also gained a Profiler link to match search.html and console.html. Verified end-to-end on devop.live/lakehouse/proof: 10 chapters render, 5/5 live tests pass, pathway shows 88 traces + 100% reuse rate, live signal tile shows 11 issuers + $347M attributed + 200 contractors + 14 attribution edges. Architecture diagram + crate table accurate as of HEAD. --- mcp-server/proof.html | 425 +++++++++++++++++++++++++++++------------- 1 file changed, 297 insertions(+), 128 deletions(-) diff --git a/mcp-server/proof.html b/mcp-server/proof.html index ceb428a..deeca7f 100644 --- a/mcp-server/proof.html +++ b/mcp-server/proof.html @@ -81,6 +81,7 @@ pre{background:#161b22;border:1px solid #171d27;border-radius:8px;padding:14px 1