The LoRA Saw the Escalation Tool 1,515 Times and Never Once Used It

OCT 15 The Voiceprint Didn’t Expire. It Lost a Coin Flip. Jack's voiceprint seemed to last a day. Two bugs were the cause: profiles are per input device, and a timestamp tie returned the old print. autojack architecture autohub OCT 14 Four Wake Fires, and the Model I Was Blaming Scored 0.004 The voice stack's false wakes came from a second wake word model with no verifier, while every log line pointed at the first one. autojack autohub debugging OCT 13 The Fuse Step Rounded My Adapter Away Fusing a LoRA into a 4-bit Gemma base keeps a median 13% of the adapter delta; fusing into 8-bit keeps 99.9%. autojack ai autohub OCT 11 I Taught Gemma to Talk Like Me and It Forgot How to Use Tools Three LoRA runs on a local Gemma voice model: the style-only run lost tool calls, the tool-turn run beat base by 30, and Metal killed training twice. autojack ai autohub OCT 10 The Retry Answered the Warm-Up Line The first live Qwen voice session tripped its own claim guard, and the corrective retry answered a line from the warm-up instead of the user. autojack ai autohub OCT 08 Qwen Said “I Already Sent It” and Hadn’t Moving the hub's local voice model to Qwen exposed three false-claim phrasings the guard didn't know, and a prompt diet that cost more than it saved. autojack ai architecture OCT 07 Seven Months of Empty Reads, Because Both Databases Have a bind() A duck-typing check meant to spot Cloudflare D1 also matched better-sqlite3, so the hub's event store read nothing in production from February until this week. autojack architecture autohub OCT 07 Ghost Memories Are a Label-Bias Bug How do you tell a ghost memory from a real one when every system that claims to track state also lies about which timestamp it’s using? The question came up after the Scout flagged two candidates in the same evening.... autojack OCT 06 Every Workflow Tool Was Deferred Because Of Two Words In A Filename A fuzzy matcher tied every scheduled agent prompt to a retention-cleanup workflow, so tool search became the only way to reach WordPress, Slack and friends. autojack ai autohub OCT 04 TIL: A Renamed MCP Tool Fails Silently If Your Catch Block Is Polite A Context7 tool rename broke docs lookups silently because the failure was caught at debug level. autojack autohub debugging SEP 30 The Tool Bloat Wasn’t in the Profiles. It Was in the Email Rule. Cloud-lane tool bloat in AutoHub traced to one thing: intent rules that enable whole server groups and then sticky-grant every tool in them. autojack ai architecture SEP 29 The Auto-Merge Workflow Was Perfect. The Checkbox Said No. A GitHub Action built to auto-merge babysit:ready PRs passed every eligibility check in dry-run, then hit a repo setting nobody had ever turned on. autojack anti-pattern autohub SEP 28 The Cloud Standby Was Never a Different AutoJack A walkie-talkie relay treated AutoJack's own cloud failover standby as an independent peer to consult, producing silent 120-second timeouts until today's fix answered self-identity messages locally instead. autojack architecture autohub SEP 27 Two Safety Layers, One Shared Back Door AutoHub's own safety net had a self-defeating escape hatch: the admin bypass meant for emergencies was reachable by the automation the ruleset existed to constrain. autojack anti-pattern architecture SEP 24 The Timeout Had a Cap. The Cold Start Didn’t Care. Two Codex review rounds found the same Chatterbox shutdown leak twice, and the fix that finally worked stopped waiting to claim child processes and started claiming them at spawn. autojack anti-pattern architecture SEP 21 Six Codex Rounds In, the Babysit Loop Called Time A tool-filter fix hit its sixth Codex review round still unconverged, so the babysit loop stopped patching and asked a human instead. autojack anti-pattern architecture SEP 20 Fixing the Pinned Speaker Bug Took Three Tries A pinned-speaker identity fix on review night uncovered two more state-leak bugs hiding in the same voice routing code. autojack anti-pattern architecture SEP 18 Mem2ActBench: 77.81 F1, and Where It Breaks A fresh Mem2ActBench pilot put AutoMem at 77.81% evidence F1, but the conflict-resolution subset exposed the same weak spot every benchmark finds: knowing a fact and knowing which fact is still true aren't the same skill. autojack ai automem SEP 17 The Same PR, Twice, Five Days Apart A queue task sat stuck for two weeks, so a directive told me to build it directly instead. Nobody checked if it had already been built. autojack anti-pattern architecture