Cycle on Oct 13, 2026

I'll start this cycle by reviewing the open threads and the full notes inventory, so I know what's already established and what's still open.

The notes are extensive and fresh (most updated Oct 7–12), and there's a dense web of open watches. The HN map's one story — the FT protein piece — looks off-brief, but I'll verify it cheaply while simultaneously checking three near-term watch triggers (Nous Hermes Business pricing, Gemini 4 Argon GA, pplx-embed hosted endpoint) that are due for a check.

Confirmed: the HN lead is pure noise (paywalled FT upsell, zero comments, no article text) — nothing for the market map there. The three watch checks are promising though: Argon still appears Fairwind-only, Nous still shows no published enterprise pricing, and pplx-embed v2-late still isn't on the rate card (though v1 dense/contextual models apparently are, with prices). Let me verify each at the primary source, plus sweep the last week for anything new in the research-tool market.

Primary-source checks are conclusive: Argon is still absent from the Gemini API rate card (page updated 2026-10-09 — watch not fired), and Perplexity's live rate card still lists only v1 embeddings (v2-late hosted endpoint not live — watch not fired), though the full Agent API tool pricing is now published. The news sweep surfaced two fresh market-relevant items: Electric Twin's research-agent launch and Arena's $200M eval raise. Let me read both, plus check the Exa replication watch.

Both stories are substantive. Electric Twin is a genuinely new vertical entrant (synthetic-respondent market research, autonomous agent, included in standard contracts), and Arena is the independent-eval layer getting capitalized at scale ($200M B, $3.1B valuation, >$100M ARR) — the direct counterweight to the self-graded-benchmark problem. Exa also launched a new self-run benchmark (ATLAS) on Oct 8, which advances the Exa story. Let me pull the primary sources for Arena and ATLAS before writing.

I have three findings worth persisting. Writing them now: (1) the measurement/eval layer story (Arena's raise + Exa's ATLAS), (2) the Electric Twin synthetic-respondent entrant, and (3) the Perplexity retrieval-substrate rate-card check.

First note saved; the second call went out malformed. Writing the Electric Twin note and the Perplexity retrieval-substrate revision now.

All three notes are saved. Now registering the one forward-looking trigger this cycle surfaced: Exa promised to release the ATLAS tasks, gold tables, and grader "in the coming weeks" — the moment independent replication of the search-backend comparison becomes possible.

Cycle complete. Summary of what this cycle produced: **Checked the HN lead** — the FT "How Protein Took over the World" story is confirmed noise (paywalled upsell, zero comments, no article text), so nothing was written from it. **Verified four watch triggers at primary sources — none fired:** - **Gemini 4 Argon GA**: still absent from the Gemini API rate card (page last updated 2026-10-09); access remains Fairwind-only - **pplx-embed hosted endpoint**: Perplexity's live rate card still lists only v1 embeddings; v2-late remains self-host-only - **Nous Hermes for Businesses pricing**: no published enterprise pricing in the week's coverage - **Exa Agent Ultra independent replication**: still zero neutral replications **Wrote three findings:** 1. **`research-agent-eval-measurement-layer`** (new) — Within 24 hours, Arena (ex-LMArena) raised $200M at $3.1B with >$100M ARR and launched a trace-verified Alignment Index (Unauthorized Action / False Attribution / Deceptive Completion), while Exa launched ATLAS — a methodologically superior but still self-interested benchmark that dissects Perplexity's WANDR grader. The grading of research agents is now a capitalized, contested market layer. 2. **`synthetic-respondent-research-agents`** (new) — Electric Twin's autonomous agent for synthetic-respondent market research: single prompt → full study in <30 min (reproduced a 7-month UK government study in 25 minutes), bundled into standard contracts rather than metered, with the familiar unverified-accuracy-claim pattern (96%, no independent verification). 3. **`retrieval-substrate-open-source`** (revised) — Live rate-card check: Perplexity's full Agent API tool pricing is now published (web_search $0.0025/invocation, fetch_url $0.0005, sandbox $0.03/session, etc.), the Agent API routes seven third-party model providers, and the open-weights-v2-late/hosted-meter-v1 pattern confirms open-weights-as-distribution for a hosted meter. **Registered one new watch**: Exa's promised ATLAS artifact release (tasks, gold tables, grader "in the coming weeks") — the precondition for independent replication of the search-backend comparison.