# SPRINT 1 FINDINGS — Reference Material

Raw evidence from the Sprint 1 manual review (200 stories, athena_review_report.txt). This document is **reference material for future decisions** — it records what the data actually showed so later tuning has a baseline.

---

## What We Learned

Sprint 1 was pure rules (no embeddings, no LLM, no semantic similarity) by founder directive: *discover the taxonomy before adding intelligence*. The classifier's job was not accuracy — it was to surface misclassifications as signal. It did.

## High Signal (ship)
Stories the editorial test would pass — reader can DO something:

- **Local AI** — gguf / rtx / quantization / llama.cpp. Highest enthusiast + replication scores.
  - e.g. "I benched quad 5060Tis for code generation with Qwen3.6-27B" (final 0.50)
  - "Voodoo Quant beats Unsloth Dynamic 2.0 KLD by 95%" (0.45)
- **Open-source tools** — things to install/run.
  - "Show HN: Juggler – open-source GUI coding agent" (0.57, highest in set)
  - "Open Source Local LLM Training Tool (consumer hardware)" (0.46)
  - "Zer0Fit — MCP server for zero-shot ML, 100% local" (0.30–0.46)
- **Benchmarks with real numbers** — reproducible, comparable.
  - "GPUHedge: cold-start p95 117s→30s" (0.41)
  - "Migrating to GPT-5.6: 2.2x faster, 27% cheaper" (0.41)
- **Infrastructure with code** — fixable, runnable.
  - "Your $80 Tesla P100 silently noisy math in llama.cpp — 3-line fix, free" (0.32)
- **Production learnings** — transferable lessons.
  - "The real bottleneck for AI agents may be proving who they are" (0.19)
  - "AI agent crawlers now need permission" (0.25)

## Low Signal (reject)
Stories the editorial test would fail — reader does NOTHING:

- **Funding rounds** — PixVerse $439M, DeepSeek $7B. Money moved, no capability changed.
- **Lawsuits** — Apple/OpenAI trade secret, Google training suit. Spectator content.
- **CEO opinions** — Altman, Hassabis, Mosseri takes. Someone said something.
- **Valuations** — PixVerse $2B. Number, no action.
- **Corporate announcements** — Waze features, Spotify assistant, Siri beta. No reproducible substance.

## Taxonomy Gap (the real discovery)
- **45 / 200 (22.5%) UNCATEGORIZED.** Mostly Reddit/HN opinion + culture-adjacent pieces with no keyword hit (e.g. "Stop Telling Me to Ask an LLM", "Mesh LLM", "Ghost Font").
- This is **not a failure** — it's the editorial boundary made visible. The keyword set is narrower than taste. Tuning Sprint 2 (Lens) must start from these 45, not from classifier precision.

## Bucket distribution (200 reviewed)
| Bucket | n |
|--------|---|
| RESEARCH | 40 |
| INFRASTRUCTURE | 23 |
| UNCATEGORIZED | 45 |
| MODEL RELEASE | 24 |
| PROBLEM SOLVED | 19 |
| SHIPPING | 12 |
| CULTURE | 15 |
| BUSINESS | 13 |
| LOCAL AI | 9 |

## Open question for Sprint 2
When the Lens arrives, its first job is **not** to classify better. It's to explain *why* a story is or isn't actionable — and to propose a section (§SECTION_DEFINITIONS) for human confirmation. The 45 UNCATEGORIZED are the training set for that judgment, not the 155 that already matched.