Two pipelines, running in sequence on weekday post-close.

• Quant feed runs at 6:00 PM ET. Prices, fundamentals, 5-year P/E history, benchmark regression, news + social tagging.
• Narrative feed (ADK agent) runs at 6:30 PM ET, after the quant feed. It reads today's quant snapshot, the last 72h of news + social, and the prior narrative record, and produces one narrative record per date.

The same narrative record powers the TLDR hero, Catalysts tab, and Risks tab. Only the TLDR is time-scrubbable; Catalysts / Risks / Basket / News all read the latest snapshot.

The ADK agent answers these four questions every day, in this order:

1. What happened since the last snapshot?
2. How did it affect basket alpha and individual names?
3. Does it reinforce or weaken the structural thesis that power, compute, and deployment are constrained?
4. What's the next catalyst to watch?

This is a tracker, not a theme-discovery tool. The thesis is fixed; the TLDR reports whether today made it more or less likely to be right.

• Power — what AI has to plug into. GEV, CEG, VST, ETN, HUBB.
• Compute — what AI runs on. NVDA, AVGO, MU, TSM, ASML.
• Deployment — what puts AI in the physical world. PWR, FIX, STRL, EME, EQIX.

Names sit in a layer because of what they supply to an AI buildout, not because of GICS. Equal-weight within each layer. The Basket change-log in the Basket tab auto-diffs composition on every run.

• Equity curve: equal-weighted basket, rebased to 100 at Jan 1 2026. Benchmarks: PAVE (US Infrastructure, primary), SMH (Semiconductors), SPY (S&P 500).
• Horizon returns: 1D / 7D / 1M / 3M / YTD / 1Y, equal-weight basket total return.
• Alpha vs PAVE: Basket YTD minus PAVE YTD, in points.
• Valuation tag: forward P/E percentile vs the name's own 5-year history. Cheap ≤30, Fair 30–70, Rich ≥70.
• Benchmark attribution: Pearson regression of basket daily returns on benchmark daily returns since Jan 1. Alpha is annualized from the average daily excess return; Beta is the regression coefficient; R² is the correlation squared.

• Prices & fundamentals: Alva equities module — OHLCV, forward estimates, 5-year P/E time series.
• Benchmarks: PAVE, SMH, SPY.
• News: Brave search (trade press, Reuters, Bloomberg, Nikkei, Utility Dive).
• Social: X / Twitter via Grok search, engagement-filtered.
• Narrative: ADK agent output written to ~/feeds/ai-bottleneck-narrative/v1/data/narrative/records, one record per date.

Every number in the TLDR and the push line is verified against the input snapshot. If any number fails verification, the record is flagged source: "fallback", the TLDR becomes a short deterministic summary, and push notifications are suppressed unless there are non-numeric signals (delta, catalyst flip, risk change).

• Private-market AI capex. Sovereign AI funds or private data-center developers moving faster than public vendors show up late here.
• Bilateral PPAs that stay confidential. If Meta signs another Crane-style deal that never hits an 8-K, we miss it until earnings.
• Intra-quarter hyperscaler guidance. Public commentary is the input; private roadmap shifts aren't.
• Chinese fab / turbine capacity. Export data is lagged; direction only, not magnitude.

• Power layer — supply of electrons to an AI workload (generation + T&D + site power equipment).
• Compute layer — supply of AI-capable silicon (GPU / ASIC / HBM / lithography).
• Deployment layer — supply of physical AI deployment (site, construction, interconnect, data-center real estate).
• Bottleneck — a supply constraint whose backlog or lead-time is visibly setting pricing.
• PAVE — Global X US Infrastructure Development ETF.
• SMH — VanEck Semiconductor ETF.
• Grounding — the rule that every number on the page comes from the day's data snapshot.