What's the design rationale for the MEMORY/LEARNING rating synthesis?

[acq23]

While discussing this with my DA, I found some information regarding the learning rating system.
I would like someone to explain the reasons behind these implementation choices:

• In my case, it feels like I need to adapt to how the platform works, rather than the platform adapting to me.
• Why isn't a straightforward voting system implemented to complement the one already in place?"

No — and this is the most important thing I've said this session. PR #966 does NOT fix "synthesis
always says nothing." That's a separate, deeper bug. I just read the synthesis tool source
(PAI/Tools/LearningPatternSynthesis.ts).

What the synthesis tool actually does:

It scans each rating's sentiment_summary field against hardcoded keyword regex patterns:

Frustrations: /time|slow|delay|wait|incomplete|missing|wrong|incorrect|misunderstand|
over-engineer|fail|error|broken|crash|bug|unclear|confus|didn't ask|
should have asked|again|repeat|same problem/i

If a summary doesn't match any of these keyword sets → it's not counted as a pattern. When zero
ratings match any pattern → output is literally "No significant issues detected. Continue current
patterns - no major issues detected."

Why this is empty FOR YOU specifically:

Your feedback style is terse instructions, code references, and questions — not verbose English
problem descriptions.

The tool was designed for verbose feedback like "this took too long" or "the response was
incomplete and wrong." Your style produces zero matches → empty synthesis. The April 15 vacuous
output is the predictable result, not a bug — given how the tool decides what counts as a pattern.

And

🗒️ TASK: Explain why synthesis ignores ratings as primary signal
🔄 ITERATION on: you spotted the bigger design flaw

You're right — that's the bigger design flaw. I read the source: the tool uses ratings only as a
filter and average, NOT as a discovery axis.

What the synthesis tool USES the rating field for:

• Overall average (line 151)
• Splitting frustrations (rating ≤ 4) from successes (rating ≥ 7) — line 154
• Per-pattern average rating, computed AFTER patterns are already formed — line 114

What the tool does NOT use the rating for:

• Pattern discovery — patterns come ONLY from regex against sentiment_summary text (lines 88–99)
• Trend detection — no week-over-week delta, no decay tracking
• Session clustering — never asks "this session_id had 3 ratings of 1–2, what task was that?"
• PRD cross-reference — MEMORY/WORK/{slug}/PRD.md files exist with task descriptions, never read
• Verbatim repeats — "implies additional services at a price?" appearing twice = real pattern, but
  never detected
• Rating density — 5 ones in 48 hours is a burst signal, ignored

The rating is the strongest signal in the dataset — numeric, unambiguous, no keyword extraction
needed. Your terse style works WITH a rating-first design ("Vote 1" carries the signal in the
number) and AGAINST a keyword-first design (no English problem-vocabulary to match).

A rating-first synthesis would key off this tuple:

(rating, comment, session_id, timestamp, source)

…and cluster on:

1. Rating density bursts — N low ratings in M hours
2. Session co-occurrence — multiple low ratings in same session → read that session's PRD for
   context
3. Cross-session task themes — fetch PRD.task for each low-rated session_id, find themes across
   them
4. Verbatim comment repeats — exact-match comment clusters ("only" appearing 3x = scope-discipline
   issue)
5. Trend slope — rolling 7-day average direction
6. Sentiment text as corroborating signal only

Why the tool was built keyword-first: my guess — it was designed for a verbose-feedback user.
Keyword regex is faster and cheaper than re-reading PRDs and clustering by session. For your
style, the cost trade-off is wrong: you give number + minimal text, so the tool needs to lean on
the `number.

`

[danielmiessler]

Closing this out as part of a cleanup of the discussion board. We're on the 7.x releases now and a lot has changed since this thread was active. Thanks to everyone who jumped in. If this is still relevant on 7.x, please open a fresh discussion or issue and we'll pick it up there.