RFC: separate deterministic confidence decay from observer reasoning

[annexiao]

Been using continuous-learning-v2 for a while and the instinct architecture is genuinely well-designed (confidence scoring + auto-load at SessionStart works better than I expected). Your X posts about ECC have been good reading too. Both have shifted how I think about a few things.

One thing I noticed in v2: the decay rule in agents/observer.md line 148 (-0.02 per week without observation) appears to be implemented as a prompt instruction to the Haiku observer agent. That works conceptually, but it makes the decay behavior harder to reason about, test, or dry-run. Since time-based decay is basically mechanical, I wonder if it should be handled outside the observer prompt.

The local change I tried was to split the two concerns currently bundled in observer.md:

• Haiku keeps doing pattern recognition and evidence-driven confidence adjustments.
• Time-based decay moves to a small Python script that reads file mtime as the last-evidence timestamp and applies a flat per-week decay.

One thing I am less sure about is the decay rate itself. The current -0.02 per week rate seems to carry an implicit assumption that if a user has not seen a pattern recently, the instinct is probably stale. That may be true for some workflows, but it feels weaker now. I find myself working across more stacks and domains thanks to coding agents, and I doubt I am the only one. In that setting, “I have not touched this domain in three months” may mean variety is up, not that the instinct is wrong. A more aggressive decay can end up penalizing rare butvalid instincts rather than stale ones.

A few open questions before I turn this into a PR:

1. Should the default rate be lower? I tried -0.005/week, but that is a judgment call
2. Where should the rate be configured: env var, CLI arg, or config field?
3. Should deterministic decay be opt-in first, with the current observer-based decay kept as the default?
4. Should scheduling guidance live in SKILL.md, or should the repo stay scheduler-agnostic and just document the script?

Happy to send a small PR if this direction lands.

[affaan-m]

might need to add another parameter into the function or reformulate, will look into this.