# ADR-0007: Editorial Pivot — AI News Aggregator → AI Enthusiast Daily

**Status:** ACCEPTED · 2026-07-15
**Supersedes:** implicit "AI news aggregation" assumption (never written down — that was the bug)
**Superseded by:** nothing yet

---

## Title

Editorial Pivot: AI News Aggregator → AI Enthusiast Daily

## Context

Athena Oracle was built as a multi-source AI research pipeline: ingest → score → pattern/falsification. Sprint 1 (pure rule-based bucket classifier + scorer) was the first time we put 200 real stories in front of a human with the question "would we proudly publish these?"

The classifier accuracy was never the discovery. The discovery was **what the stories were actually for**. Scanning the 200, the ones we'd proudly ship were uniformly *actionable by a builder*: local-model how-tos, benchmarks with real numbers, reproducible tooling, agent infra with code. The ones we'd reject were uniformly *spectator content*: lawsuits, funding rounds, CEO opinions, valuations, corporate announcements.

We had been optimizing the system for the wrong question.

## Decision

Athena Oracle is **no longer optimizing for:**

> "What happened in AI today?"

Athena Oracle is **optimizing for:**

> "What can an AI enthusiast DO after reading this?"

This changes every layer: scoring, classification, editorial review, rendering, narratives, memory. Everything.

## Before / After

| Dimension | Before (implicit) | After (locked) |
|-----------|------------------|-----------------|
| Optimized for | Recency | Actionability |
| | Volume | Reproducibility |
| | Virality | Enthusiast value |
| | | Builder usefulness |
| Unit of value | "was published" | "can be done" |
| Audience | anyone interested in AI | builders / tinkerers / local-AI users |
| Reject signal | none | "reader would do NOTHING" |

## Sprint 1 Evidence

From the 200-story manual review (athena_review_report.txt):

**High signal (would ship):**
- LOCAL AI bucket — gguf / rtx / quantization / llama.cpp stories scored highest on enthusiast + replication
- Open-source tools (Juggler GUI coding agent, open-source arXiv tool, Zer0Fit MCP server)
- Benchmarks with real numbers (GPUHedge: 117s→30s p95; Migrating to GPT-5.6: 2.2x faster, 27% cheaper)
- Infrastructure with code (Your $80 Tesla P100 llama.cpp fix)
- Production learnings (agent identity, permissioned crawlers)

**Low signal (would reject):**
- Funding rounds (PixVerse $439M, DeepSeek $7B)
- Lawsuits (Apple/OpenAI trade secret, Google training suit)
- CEO opinions (Altman, Hassabis, Mosseri takes)
- Valuations (PixVerse $2B)
- Corporate announcements (Waze features, Spotify assistant)

**Taxonomy gap found:** 45 / 200 (22.5%) were UNCATEGORIZED — mostly Reddit/HN opinion + culture-adjacent pieces with no keyword hit. This is exactly the discovery data we wanted: the editorial boundary is sharper than any keyword set yet.

## Consequences

- All future scoring work measures actionability, not recency.
- The Lens (Sprint 2) is **blocked** until manual review tallies are returned and the taxonomy is tuned by hand.
- Narratives and memory engines, when built, must encode "what can be done," not "what happened."
- This ADR is the reference anchor for every later ADR.

## Confirmation

Accepted by founder directive 2026-07-15. Locked in `/docs/vision/VISION_LOCK_V1.md`.