A critical bug in my job-screening judge agent taught it false rejection reasons. The agent, designed to learn from my decisions, was incorporating staging script instructions into its understanding of why I rejected certain roles. This wasn't a minor data anomaly; it fundamentally corrupted the agent's core learning loop, leading to a persistent, incorrect lesson: "pay below her floor 8x."
This specific lesson originated not from my actual feedback, but from a micro-note in a staging environment: '🚨 The form pre-fills'. This note, intended for internal agent guidance during form filling, was mistakenly interpreted as my personal reason for rejecting a job. The impact was significant. A read-only audit of how my operation's agents share knowledge revealed this issue in the screening judge's live lesson file, the summary it consults for every job.
The False Lesson: "Pay Below Her Floor 8x"
The agent's internal model for my rejection reasons was heavily skewed by this staging note. The "pay below her floor 8x" lesson was the strongest, appearing eight times in its learned reasons. Seven of those eight instances were directly attributable to the staging script's instructions. This meant that for a significant portion of its decision-making, the agent was operating under a false premise about my compensation preferences.
My aideazz repository commit 37ec1cd on 2026-10-15 explicitly states: "another agent wrote her top lesson (new concept: label-provenance) - a staging script's form instructions were learned as Elena's rejection reason; verified 54 -> 47 with her reason, pay lesson 8x -> 1x". This commit details the discovery and the immediate impact of the fix: the number of instances where this false reason was attributed dropped from 54 to 47. More importantly, the "pay below her floor" lesson, previously appearing 8x, was corrected to 1x.
Identifying the Source of Corruption
The VibeJobHunterAIPA_AIMCF repository, specifically commit e864641 on 2026-10-15, outlines the fix: "the cto-aipa micro1 staging note ('🚨 The form pre-fills') is an agent note, not Elena's reason - added to _KIT_NOTE; a cached reason that matches _KIT_NOTE is re-derived once (removes the false 'pay below her floor 8x' lesson from every judge prompt)".
This commit highlights the precise origin of the problem: a staging note from the cto-aipa microservice. The fix involved classifying this specific note as an internal _KIT_NOTE rather than a user-provided reason. Any cached reason that matched this _KIT_NOTE was then re-derived, effectively purging the false lesson from the agent's memory. The cto-aipa process itself has seen 209 restarts and has been up for 0 days, indicating recent, active development and deployment.
The Broader Impact on Agent Learning
This incident underscores a critical challenge in building robust AI agents: ensuring the provenance of learning data. When agents learn from human feedback, it's crucial to distinguish between actual user input and internal system notes, staging instructions, or even agent-generated observations. Without clear label provenance, the agent's internal model can become polluted, leading to incorrect inferences and suboptimal decision-making.
Another related fix, commit 6b9ebd4 in VibeJobHunterAIPA_AIMCF on 2026-10-15, addressed how my "No-fit screenshots" were processed. Previously, the agent prioritized the kit's CV PDFs, taking up the two available slots for image processing. This meant my actual rejection notes, often conveyed through screenshots, were not being fully read. The fix ensures that "only her notes' images (the kit's CV PDFs took the 2 slots), up to 4 joined, one model's NONE gets a second opinion, a failed read is retried, recent reason-less rejections re-read; on-site/hybrid/must-live-in teach the location lesson." This further refines the agent's ability to learn accurate rejection reasons directly from my input, rather than relying on incomplete or misinterpreted data.
Operational Stability and Continuous Improvement
My cto-aipa repository has seen 12 commits in the last 48 hours, demonstrating continuous development. These commits include updates to the resilience documentation (a689ad72, effd7ddf), voice default changes (38489e2f, de06ac75), and other fixes related to agent wiring (bb25cdb3). The voice default change, for instance, moved from gpt-4o-mini-tts (which stalled 1 in 4 calls) to tts-1 (which completed 12/12 calls in 1.7-3.5 seconds), improving reliability.
The aideazz repository also shows 4 commits in the last 48 hours, including the ai-ops-wiki updates (37ec1cd, 3dabeca, b8614b3) and a blog regeneration (3334377). This consistent activity across repositories reflects the ongoing effort to refine and stabilize the AI agent ecosystem. The pm2 jlist shows 9 processes online, with cto-aipa having 209 restarts and algom-stream having 55193 restarts, indicating a system that is actively managed and recovers from failures.
Frequently Asked Questions
Q: How did the agent distinguish between a staging note and Elena's actual rejection reason?
A: The fix involved explicitly classifying the staging note ('🚨 The form pre-fills') as an internal _KIT_NOTE within the VibeJobHunterAIPA_AIMCF repository. This prevented the agent from learning it as an external, user-provided reason.
Q: What was the immediate impact on the agent's learned reasons after the fix?
A: The number of instances where the false reason was attributed dropped from 54 to 47. The "pay below her floor" lesson, which previously appeared 8x, was corrected to 1x.
Q: How does this fix relate to other agent learning improvements?
A: This fix is part of a broader effort to improve agent learning. Another recent fix ensures that my "No-fit screenshots" are fully read, allowing up to 4 images to be joined and retrying failed reads, rather than prioritizing CV PDFs.
Q: What is "label-provenance" in this context?
A: Label-provenance refers to the origin and context of the data labels or reasons an AI agent learns from. In this case, it was crucial to establish whether a "reason" originated from an internal staging instruction or from my direct feedback.