Loading intelligence...
Every error caught. Every correction documented. We play good, clean ball.
No consulting firm shows error rates. We do.
4 times a validation corrected something — and a later pass found THAT correction was wrong.
5 claims are permanently unverifiable. We don't hide them — we label them D.
| Claim | Was | Corrected To | Caught | Source |
|---|---|---|---|---|
| "Gen X" fraud model at BofA | BofA | American Express | V1 | NVIDIA, Tearsheet, Emerj |
| Cognition/Devin pilot at BofA | Confirmed | Goldman Sachs (not BofA) | V1 | CNBC Jul 2025 |
| BofA NOT among agentic patent filers | Not in list | IS among top 3 of 8 banks | V1 | CFOTech Asia / Evident |
| Citi stock 2025 performance | +67% | +47-53% | V11 | Yahoo Finance, Nasdaq |
| Morgan Stanley total client assets | $6.5T | $9.3T (total W+IM) | V17 | MS Q4 2025 earnings |
| Wells Fargo consent orders | 3 remaining | ALL CLEARED (2025-2026) | V16 | Federal Reserve |
| Goldman Marcus consumer losses | $7B+ | ~$4B documented | V16 | American Banker |
| Wells Fargo "#1-2 mortgage lender" | #1 or #2 | #14 ($20.4B vs UWM $137.8B) | V15 | HousingWire/HMDA |
| Capital One "no money market products" | No money market | 360 Money Market Account EXISTS | V15 | Capital One website |
| Claim | V1 Said | V2 Corrected | Final (V3+) |
|---|---|---|---|
| BofA tech spend | $13.5B | $14B | $13.5B (V2 was wrong, reverted) |
| BofA AI patents | 1,200+ | 1,400+ | 1,500+ (V2 outdated, updated) |
| Capital One AI patents | 1,700+ unverifiable | Use 967 | 1,700+ IS verified via Evident (V2 was wrong) |
| BofA Devin pilot | Not piloting | Confirmed (cross-ref) | Reported only (V2 overcorrected; Cognition doesn't list BofA) |
These prove multi-pass validation works. A single pass would have published incorrect corrections.
13 bugs identified by 3 parallel Opus review agents. All critical and high-severity issues resolved.
| Area | Issue | Fix Applied | Root Cause | Date |
|---|---|---|---|---|
| Voice Agent | Gemini API deprecated clientContent.turns format | Migrated to realtimeInput.text | Google API breaking change — undocumented | 2026-03-28 |
| Voice Agent | Gemini API deprecated realtimeInput.mediaChunks | Migrated to realtimeInput.audio | Google API breaking change | 2026-03-28 |
| Voice Agent | Audio mimeType "audio/pcm;rate=16000" rejected | Changed to "audio/pcm" | Rate parameter no longer accepted in mimeType string | 2026-03-28 |
| Voice Agent | Echo/feedback loop — mic audio played through speakers | Connected processor to silent gain node instead of destination | ScriptProcessorNode requires output connection; was connected to speakers | 2026-03-28 |
| Voice Agent | Choppy robotic speech playback | Pre-scheduled gapless buffer timing + 80ms chunk batching + 2ms crossfade | onended callback chaining caused 5-20ms gaps between audio chunks | 2026-03-28 |
| Voice Agent | Memory leaks on page navigation | Added useEffect cleanup for WebSocket, AudioContext, MediaStream | No cleanup on component unmount | 2026-03-28 |
| Voice Agent | "hashtag 2" instead of "number 2" in speech | Added TTS text cleanup: #N→"number N", tickers→full names, acronyms→phrases | Gemini TTS reads # as "hashtag" | 2026-03-28 |
| Chat API | Scraping detection ran AFTER burning Opus tokens | Moved scraping check before Anthropic API call | Ordering oversight — expensive API call before cheap check | 2026-03-28 |
| Chat API | Invalid JSON body returned 500 instead of 400 | Separate try/catch for JSON parse | request.json() throw caught by outer catch returning 500 | 2026-03-28 |
| Chat API | No timeout on Anthropic API call | Added 45s timeout | Missing timeout parameter | 2026-03-28 |
| Chat API | Client-supplied history could inject fake assistant messages | Strict role filtering — only accepts "user" or "assistant" | No validation on history role field | 2026-03-28 |
| OpenClaw | 3 critical security vulnerabilities | iMessage allowlist, sandbox all agents, workspace-only FS | Default open group policy with unsandboxed small models | 2026-03-28 |
| OpenClaw | All 10 cron jobs in error state | Restarted Ollama server + switched default to Opus | Ollama server was down — no model available for cron execution | 2026-03-28 |
| Database | WFC evident field had consent order text instead of ranking | Corrected to "#6 (down from #4)" | Accidental overwrite during earlier UPDATE that set evident instead of a notes field | 2026-03-28 |
| Database | WFC tech_spend showed "$4B" (new initiatives only) | Corrected to "~$9B total (~$4B new initiatives)" | Original DB entry truncated — only captured new initiative portion | 2026-03-28 |
| Database | Missing regulatory data for GS, MS (CFPB) and COF, GS, C, MS (CourtListener) | Inserted 6 missing rows — all 7 banks now have complete coverage | Initial migration only inserted banks with high counts — missed low/zero entries | 2026-03-28 |
For the complete build history including all features, infrastructure changes, and security updates:
View Full Changelog