Cycle on Sep 24, 2026

I'll start this research cycle by reviewing what's already known and what threads are open, then gather fresh evidence on the AI research-tool market.

I have a rich prior-findings base (~30 notes, several revised just hours ago at 2026-09-24T01:47). This cycle I'll: (1) check the two HN stories in the current map that touch the research-tool market — Panel (agentic research workspace) and "AI is breaking our proxies for expertise"; (2) sweep the web for any fresh moves from the core players (Perplexity, Sacra, Gemini Deep Research, Scholar Labs); and (3) try to resolve the open Undermind verification question. Starting with the HN stories:

Both HN stories are already absorbed into existing notes (`agentic-research-workspaces-panel` and `ai-expertise-proxies-goodhart` were revised within the last day with exactly these threads' quotes — the QB/WR analogy, the persistence question, the MCP demand). Re-writing them would just re-mint the same evidence. Now a fresh web sweep for anything new in the last few days across the core market players, plus an attempt to resolve the open Undermind verification question:

Two useful hits: (1) the Perplexity Sonar→Agent API retirement is confirmed for Sept 25–27 (watch not yet fired — fires in 1–3 days); (2) the Undermind question has real independent test data now — Aaron Tay's comparison post and an SMU library guide both cite a systematic-review-based recall test. Let me fetch those to ground a resolution of that open thread:

Aaron Tay's posts give me exactly what the open Undermind thread asked for: an independent, systematic-review-based recall test. Let me also fetch the CASRAI guide, which aggregates the vendor-claim caveats and the librarian test, so I can quote it precisely:

The CASRAI guide gives me the full audit picture. Before writing, one quick sweep for any market-moving launch/funding news this week that the last cycle (which ran just hours ago) might have missed:

One new item worth pursuing: Firecrawl raised a $75M Series B for "Alexandria" — a paid knowledge layer for AI agents ("paying them when AI agents use it"). That's a direct extension of the data-supply layer in the market map. Let me fetch it:

I have everything needed. Two findings this cycle: (1) Firecrawl's Alexandria — a new, funded broker layer for compensated agent data supply (new subject); (2) the independent-verification record for Undermind's recall claims, which resolves the open question thread (update to the existing academic deep-search note). Writing both:

The Alexandria note landed. The second call went out empty — writing the Undermind revision properly now:

Both notes are in. Now resolving the Undermind verification thread and registering the forward-looking watch on Alexandria's self-service payout launch:

Cycle complete. Summary of what this cycle added: **Checked the HN map first** — the two stories touching the research-tool market (Panel's agentic research workspace, the Goedecke "proxies for expertise" essay) were already absorbed into their standing notes (`agentic-research-workspaces-panel`, `ai-expertise-proxies-goodhart`) revised within the last day, so I didn't re-mint them. **New finding — Firecrawl Alexandria ($75M Series B, Sep 22):** the agent-knowledge supply layer now has a funded horizontal broker that pays contributors "when AI agents use it" (Wikimedia Enterprise already live; self-service payouts "soon"). Written to `compensated-knowledge-supply-agents` as a fourth data path alongside licensed MCP backends, Parallel's payout backend, and host-platform licensing — with a one_shot tripwire watch on the self-service payout launch. **Resolved the open Undermind question:** the recall claim now has an independent record — Aaron Tay's gold-standard test (recall@10 72.7%, recall@50 81.8% vs. a PTSD systematic review), a second librarian test (6 of 9 studies), the peer-reviewed JCHLA review, and CASRAI's audit showing the "10x–50x vs Google Scholar" line is an unaudited vendor benchmark on a changed corpus. Updated `academic-deep-search-undermind-asta` and closed the thread. **Watch status, no events:** Perplexity's Sonar→Agent API retirement is confirmed for Sept 25–27 (existing watch fires within days); Scholar Labs shows no GA/premium news.

The research cycle is complete. Here's where things stand for the market map: **This cycle's contributions:** 1. **New market-layer finding** — Firecrawl's Alexandria ($75M Series B) is building the broker layer for compensated agent data supply: one MCP/API integration covering live web + official providers (Wikimedia Enterprise already paying) + proprietary indexes, with contributor payouts generalizing to a self-service system soon. This sits alongside the three data paths already mapped (institutional MCP vendors, Parallel's per-request payout backend, Google's host-platform licensing) and matters for Hey, Lefty because it lowers the integration cost of a many-sources orchestration thesis — while raising the dependency question. 2. **Resolved open question** — Undermind's recall differentiation is now independently spot-checked (72.7% recall@10 / 81.8% recall@50 against a systematic-review gold standard), so the market map can position it as a "high-recall narrow-query specialist" rather than repeating the unaudited "10x–50x vs Google Scholar" line. 3. **Watches queued** — Perplexity's Sonar deep-research endpoints retire Sept 25–27 (existing watch will fire within days), and the new Alexandria payout-launch tripwire is registered. Scholar Labs remains waitlist-gated with no premium-tier news.