savarin

Question:
Is there a reason you switched from "interpretable" when there were 4 groups to
mean "there is only one way to do things and thus interpretable to humans", but
when there were 8 groups you mean "interpretable by the machine"?

Answer:
I apologize for any confusion caused. The term "interpretable" can be used in
different contexts and with different meanings. In the first categorization with
four groups, I used "interpretable" to refer to languages that are human
-readable and allow for clear understanding of the code's behavior. In the

savarin

import sys


DETAILS = {
    2: [7, 5, "Introduction to Operating Systems", "https://pages.cs.wisc.edu/~remzi/OSTEP/intro.pdf"],
    4: [6, 5, "The Abstraction: The Process", "https://pages.cs.wisc.edu/~remzi/OSTEP/cpu-intro.pdf"],
    5: [7, 6, "Interlude: Process API", "https://pages.cs.wisc.edu/~remzi/OSTEP/cpu-api.pdf"],
    6: [5, 12 ,"Mechanism: Limited Direct Execution", "https://pages.cs.wisc.edu/~remzi/OSTEP/cpu-mechanisms.pdf"],
    7: [10, 5, "Scheduling: Introduction", "https://pages.cs.wisc.edu/~remzi/OSTEP/cpu-sched.pdf"],
    8: [6, 5, "Scheduling: The Multi-Level Feedback Queue", "https://pages.cs.wisc.edu/~remzi/OSTEP/cpu-sched-mlfq.pdf"],

savarin

Review: PR #418 - Training Pipeline Hardening

This is a substantial and high-quality piece of engineering. It makes several critical improvements to the training pipeline that address real-world operational concerns. My review is strongly positive.

Summary

This PR hardens the nbcore training pipeline by tackling two fundamental issues: the risk of OOM errors from loading large training windows into memory, and the temporal information leakage from using a random train/test split.

It introduces:

1. A week-by-week streaming pattern (stream_training_weeks) to cap peak memory usage.

savarin

ALICE

Review: Broaden _requires_full_refresh to check entire query tree

Summary

This PR simplifies _requires_full_refresh by replacing a targeted SELECT-projection-only check with a whole-query parsed.find(expression) search. The motivation is that Snowflake dynamic tables don't support change tracking for CURRENT_TIMESTAMP()/CURRENT_DATE/CURRENT_TIME anywhere in the query, not just in SELECT projections. The old code incorrectly allowed INCREMENTAL mode when these functions appeared in WHERE clauses.

This is a correctness fix — the old behavior could produce dynamic tables that Snowflake silently downgrades or errors on.

savarin

ENG-407: Before/After Report Comparisons

Three experiments tested against the local endpoint. Before = current production (no pre-pass). After = with LLM pre-pass generating hypotheses from template content.

---

1. Activity Nudges Email: At Risk (Gen 1 — zero metadata)

This is the strongest case. 34 variants, zero hypothesis or description in Postgres.

savarin

Report Insights: v4 Model Comparison — Sonnet 4.6 vs Opus 4.6 vs Opus 4.7

Scored 3 models × 12 experiments against the v4 rubric (10 criteria, 3 tiers). Responses frozen from May 21, 2026. Judge model: Opus 4.7.

Results

#	Criterion	Tier	Method	S46	O46	O47	Notes
1	Summary names test dimensions	1	Codebook extract-then-score	12/12	12/12	12/12	All models perfect
2	Business impact vs holdout control	1	LLM extraction	10/12	10/12	6/12	O47 states lifts vs peer variants

savarin

"Pruned" jargon leak — analysis and fix

Context

The report-insights endpoint generates LLM-powered experiment reports for exec stakeholders. Reports are evaluated against a 10-criterion rubric (v4) across 12 Suno experiments. The v4 baseline for v2 is 88/112 (79%).

Finding

All C4 (no pipeline jargon) failures in the v4 baseline come from a single term: "pruned."

savarin

Experiment Categorization Dimensions

Categorization pipeline output for all 22 active experiments.

Activity Nudges Email: At Risk — 34 variants · 3 cohorts

• Hero headline framing (h1): 20 values across 5 themes — urgency & readiness prompts, creation & originality invitations, lyric & idea transformation, content & backdrop utility, and anticipation & discovery — with CTA, persona, and subject line each bundled per headline
• Variant cohort: original standard HTML batch (w/ control) → JTBD persona-named variants (standard HTML) → Simple HTML re-run of select variants

Activity Nudges Email: Dormant — 20 variants · 2 cohorts

• User persona segment (persona): 10 values across 3 themes — creative hobbyists & experimenters, social & sharing-oriented creators, and professional & utility-driven musicians — with CTA, cta url, headline, and subject line each bundled per persona
• Variant cohort: standard HTML (w/ control) → Simple HTML

savarin

from __future__ import annotations

from collections.abc import AsyncIterator
from contextlib import asynccontextmanager

from fastapi import FastAPI
from fastapi.routing import APIRoute
from fastapi_mcp import FastApiMCP
from nbcore.partner import PartnerContextServiceType, partner_context

savarin

Agents

Domain model

An experiment tests messaging variations (email or push) against a target audience. Each experiment has variants (individual messages with creative copy and a cohort_date), cohorts (date-based batches), and evaluations (per-variant lift, rank, and significance from Snowflake).

Tool pipeline

Six tools run in order: categorize → interpret → audience (S1) → design (S2) → result (S3) → takeaway (S4). All accept experiment_id plus optional --variant-ids and --cohort-date.