The House Edge

Appearance

Chapter 2 — Review Notes (Phase 3 self-check) ​

Status: REVIEW NOTES (Phase 3 — written by agent, accompanying DRAFT.md) Date: 2026-05-13 (revised same day after Nick's first-read review — see "Revision log" below) Working chapter: 02 — Where Liquidity Lives Word count: ≈4,375 words including footnotes after the 2026-05-25 humanizer pass (within the Book Bible's 4,000–6,000 target).

Revision log ​

2026-05-25 (rev 3) — Applied a base blader/humanizer-style prose pass:

  • Cut agentic scaffolding and repeated abstract terms ("architecture," "structural," "mechanically") where plain prose did the job.
  • Reworked the cold open, setup, Bob example, AMM/LVR explanation, Hyperliquid section, prop-AMM displacement, verdict, and Chapter 3 transition for a clearer human narrator.
  • Preserved the existing claims, examples, numbers, diagrams, and footnote structure.

2026-05-13 (rev 2) — Two Hyperliquid corrections surfaced during Chapter 1's Phase 1 research and were back-edited here:

  • §5b previously said HyperBFT "produces blocks at approximately seventy-millisecond intervals." The Hyperliquid docs cite ~200ms median end-to-end latency for a co-located client (with a 99th percentile of ~900ms). The "70ms" number circulates widely in ecosystem commentary but isn't in the canonical source. §5b and §6 prose were both updated to use the 200ms figure.
  • The OUTLINE.md and RESEARCH.md planning docs previously cited a 3.2 bps taker fee on Hyperliquid. The current Hyperliquid Fees documentation shows 4.5 bps base / 2.4 bps top-tier on perpetuals; 7.0 bps spot taker. The 3.2 bps figure in the Ch 2 planning docs was incorrect at the time of writing; updated to current.

These were back-edits triggered by fresh primary-source research for Chapter 1, not new findings on Ch 2 specifically. No load-bearing chapter claim changed; the Hyperliquid-as-CLOB-that-worked thesis stands.

2026-05-13 (rev 1) — Nick flagged two issues after looking at the live DefiLlama Phoenix dashboard:

  1. The Q1 2026 Phoenix revenue figure was wrong. The original draft cited $11,190 for Q1 2026 fee revenue. The DefiLlama dashboard tooltip for Jan 1 – Mar 31 2026 actually shows $68,604. The research agent misread the Q1 2026 quarterly value (likely conflated with a different time window). The decline from peak is therefore ~54× rather than ~330× — still a dramatic collapse, but the correct number. Updated everywhere it appears:
    • DRAFT.md cold open and footnote 1
    • RESEARCH.md key claim #10 and the numbers-to-verify table (N8)
    • book/OUTLINE.md Ch 2 entry
    • Also added cumulative lifetime revenue $14.54M to footnote 1, since the dashboard surfaces it and it gives readers a useful cross-check (~1.9 bps realised-fee rate against $75B cumulative volume).
  2. The cold open didn't articulate the chapter's argument. Compared to Chapter 4's cold open, which ends with "This chapter is about what people like the operator of Vpe do for a living, why the role exists, and out of whose pocket the $13.4 million came," the Chapter 2 cold open ended on "The architecture lost the market" — a vivid observation but not a chapter-level argument. Rewrote the closing sentence as: "This chapter is about which architectures won that market, what their economics extracted from the people supplying them with liquidity, and why the role most retail traders think of as 'providing liquidity' no longer means what it did in 2021." This mirrors Ch 4's pattern (which architectures + what they extract + named effect) and explicitly tees up the chapter's verdict (the passive LP role has been quietly redefined) rather than asking the reader to wait until §5d for it.

No other sections were affected by this revision. The §5b reference to Phoenix's decline ("Phoenix — the spot CLOB the cold open began with — lost it") still works because it references the cold open's narrative rather than the specific dollar figure.

This is the eight-question self-check the chapter spec template requires, plus an explicit ledger of what changed between RESEARCH/OUTLINE and DRAFT and what I'm uncertain about. The user instructed me to proceed through Phase 2 and Phase 3 without a checkpoint between them ("proceed with putting together a draft or whatever is next"), so this note is the first place I formally pause for review.

The eight required questions ​

  1. Could a smart business reader with zero crypto background follow this chapter on first read?Yes. Every term that appears for the first time is defined inline before it carries load. The constant-product invariant x · y = k is stated explicitly. Impermanent loss is named and then immediately reframed as LVR with the σ²/8 → 11% annualised translation. The aggregator/solver distinction is drawn carefully because the two architectures look similar to a casual reader but matter differently. The one section that might require concentration: §5d (prop-AMM displacement), because it makes the chapter's most novel structural claim. If the reader is going to slow down anywhere, it is there.

  2. Is every actor named, and is it clear how each one makes money?Yes. Bob (passive LP — fees minus IL/LVR), Wintermute (MM — spread minus inventory risk), Jupiter (aggregator — slippage capture), Temporal/Ellipsis/Tessera operators (prop-AMM — spread, kept for the firm), the solvers SCP/Wintermute/Barter (filling intents from inventory). HLP gets a short mention with its revenue source named (quoting and liquidation surplus). Each actor has at least one named firm or operator behind it.

  3. Is there a worked example with specific dollar amounts threaded through the chapter?Yes. Bob's $50,000 LP position on Uniswap V3 ETH/USDC 5-bps in February 2026: $350 in fees, $475 in LVR loss, net -$125 for the month. Plus the Topaze aggregate ($60.8M underperformance across 17 pools, 49.5% of LPs net negative). Plus Alice's $10K aggregator-routed swap as a continuity beat (60% HumidiFi, 25% Orca, 15% Meteora).

  4. Does the chapter end with a clear "who wins, who loses" verdict?Yes, intentionally lighter-touch than Ch 4. The SPEC was explicit that this chapter is structural rather than adversarial — the verdict reads as "not particularly bad, but the role has been redefined under the LP's feet" rather than the searcher chapter's clinical-extraction framing. The closing line ("The passive LP who deposited in 2021 was promised one job. In 2026, the job is being done by a different kind of entity. That is what the rest of the book is about.") is the line I think Nick will most want to either keep verbatim or rewrite.

  5. Are all numbers sourced in footnotes?Yes, with three explicit caveats called out in-line. Every dollar figure, percentage, and named incident has a footnote with a URL and an access date. The caveats:

    • Footnote 4: MEXC Research's 2025 54.7% LP-loss figure is cited "for context only" because I could not find a primary URL — only secondary citations. The chapter does not lean on this number.
    • Footnote 6: Bob's $350/$475/-$125 monthly P&L is illustrative, derived from publicly available aggregate parameters (LVR formula at typical 4–5% daily ETH vol, 5-bps tier fee yield ≈8% annualised on TVL, full-range position). Individual LP outcomes vary with range choice and realised volume. The chapter states this explicitly in the footnote.
    • Footnote 8: The Hyperliquid BTC perp top-of-book spread of approximately $1 (vs Binance's ~$5.50) is operator-disclosed via the founder's January 2026 X post and CCN coverage, not independently audited. The chapter cites it with that hedge.

  6. Does the chain comparison box exist and contain real differences (not "Solana is faster")?Yes. The three paragraphs make three structurally different points: Solana's spot is a prop-AMM-and-aggregator stack with the public CLOB experiment effectively over; Hyperliquid's CLOB works because HLP underwrites the worst flow and HyperBFT eliminates the cancel-stale-quote problem; Ethereum supports both architectures because the gas regime allows it and the solver-market layer has taken meaningful retail flow share from direct AMM trading. No throughput-comparison sentences.

  7. Did I avoid every banned move from the Book Bible?Yes, with two flagged near-misses.

    • I did not use "revolutionary," "game-changing," "paradigm-shifting," "the future of," "disrupting," "rugged," "broken," "dead" (used once: "the on-chain version of this design failed for nearly a decade" — judgment call; the underlying claim is sourced and accurate), "scam," "Web 1.0," or similar. No throat-clearing ("It's important to note that…"). No tweets cited authoritatively (footnote 8 cites a founder's X post but flags it as operator-disclosed, not as authoritative third-party data).
    • Near-miss 1: §5d's argument that "the public LP role disappears because the function the public LP performed has been internalised by a market-making firm that does the job better and keeps the proceeds" is editorial-by-degrees. I think the sourced facts (prop-AMM share, Solana's specific case) make it defensible, but it's the chapter's strongest interpretive claim and could be softened.
    • Near-miss 2: The "Solana converged on prop-AMMs routed through Jupiter — the public AMM business is real but the dominant liquidity is private" framing in OUTLINE.md is sharper than the SPEC's "AMMs + aggregators." This is the biggest structural update from research; the chapter develops it deliberately. Flagging because if you would rather defer the prop-AMM thesis to Chapter 8 (Solana), §5d would need to be re-scoped to keep Chapter 2 closer to architecture-level claims.

  8. Would the Goldman MD finish this chapter without checking her phone?Probably. I read it aloud once. The Phoenix cold open earns the first page. The Bob setup gives her a number to follow. The AMM-curve diagram replaces what would otherwise have been three paragraphs of prose. The Meet the Market Maker sidebar is short. The chain comparison is genuinely informative because each paragraph says something different. The two spots where she might slow down: the LVR paragraph (the σ²/8 derivation is technical but I kept it to two sentences with the empirical Topaze data on either side), and §5d's prop-AMM section (because the structural claim is unfamiliar). The verdict's closing line is the kind of sentence that earns its position.

What changed between phases — and what's load-bearing ​

Claims dropped from RESEARCH.md ​

  • The MEXC Research 54.7% LP-loss figure (RESEARCH.md cross-referenced via background-research output, not the original Phase 1). Dropped from the chapter body because I could not find a primary URL. Footnoted as context only (footnote 4). If a primary source surfaces, easy to upgrade to a load-bearing claim.
  • The full HLP return profile and the JELLY exploit mechanics (RESEARCH.md N13, N14). Dropped to one sentence each in §6 and footnote 18 respectively. The full JELLY case study is reserved for Chapter 9 (Hyperliquid). Chapter 2 needed HLP in cameo, not in depth — the SPEC asked for this and I held to it.
  • Manifest as the only Solana CLOB with material 2026 spot activity (RESEARCH.md N9). Mentioned once in §6 with the $3.9B 30-day number; not developed further. Manifest is a footnote-level fact for this chapter.
  • The full solver-market concentration story (RESEARCH.md N17). Used the SCP+Wintermute >90% UniswapX number and the Barter 28% CoW share, but did not develop the deeper "two-firm or three-firm oligopoly across all solver venues" thesis. That belongs in Chapter 7 (Exclusive Order Flow).
  • CrocSwap 2022 concentrated-liquidity utilization data (RESEARCH.md N4: 40–50% of V3 positions out-of-range at any time). Mentioned in passing in §5a but not as a numbered footnote claim. The structural concentrated-liquidity point lands without the specific 40–50% figure; if you want it back, it's an easy footnote addition.

New claims added in the draft (and where they came from) ​

  • Bob's $350/$475/-$125 monthly P&L. Illustrative, derived from publicly available aggregate parameters. Footnote 6 acknowledges this explicitly. Not a primary-source measurement.
  • The Hyperliquid top-of-book spread ($1 vs Binance's $5.50) and the depth comparison ($3.1M vs $2.3M at 1 bp). From the background-research pass (CCN article citing the Hyperliquid founder's January 2026 X post). Operator-disclosed, not third-party audited. Footnote 8 says so.
  • The framing "The public LP role disappears because the function the public LP performed has been internalised by a market-making firm." §5d argument; synthesis, not a directly sourced claim. Flagged near-miss 1.
  • The framing "The pool is no longer the front of the market. It is the back-stop the solver trades against when the solver's own inventory is wrong." §5c argument; synthesis, supported by but not directly stated in the Khakhar et al. paper.
  • The Tessera/Titan relationship (Tessera supplies 54% of Titan's execution). From the SolanaFloor / Phemex coverage of Titan's growth — confirmed in the Phase 1 research note but not as a numbered claim there.

Background-research output incorporated ​

The targeted parallel-research pass for the two RESEARCH.md flagged gaps (2026 LP-loss data, Hyperliquid spread data) returned partial findings:

  • Gap 1 (2026 LP-loss): MEXC Research's 54.7% figure exists but only with secondary citations. The a16z LVR primer (citing the σ²/8 → ~11% annualised derivation) is a useful primary source for the LVR estimate the chapter uses. Both are footnoted (footnotes 4 and 6). No comprehensive 2026 re-run of the Topaze methodology was found.
  • Gap 2 (Hyperliquid spread): The CCN article citing Jeff Yan's January 2026 X post provides $1 vs $5.50 top-of-book and $3.1M vs $2.3M at 1 bp depth — operator-disclosed numbers, cited as such (footnote 8). No formal lead-lag study comparing Hyperliquid and Binance price discovery was found.

The background research did not change any chapter conclusion, but tightened the §5b Hyperliquid claims with specific numbers the original Phase 1 research had flagged as unavailable.

Things I'm uncertain about ​

  1. Bob's monthly numbers. $350 fees / $475 LVR / -$125 net is derived from aggregate parameters, not from a specific dashboard observation. The chapter and footnote 6 say so. If you'd rather use a different worked-example calibration (higher vol period, different range choice, a different pool), the math is easy to redo. The structural point — passive LPs lose money in expectation in the 5-bps tier — is unchanged.
  2. The Temporal-as-HumidiFi attribution. DL News reported it; neither HumidiFi's operator nor Temporal has publicly confirmed. The chapter says "reportedly run by the firm Temporal" with that hedge. If you'd prefer to keep it anonymous, the change is one sentence.
  3. The Phoenix cold open. Strong opening but it is essentially a pivot story, not an extraction story. If you want the cold open to land harder on the chapter's adversarial-by-implication tone, an alternative open on Bob's $125 monthly loss (small but real, multiplied across many Bobs into the Topaze aggregate) would do that. I went with Phoenix because the dollar gap ($3.7M → $11K) is visceral. Flagging for second opinion.
  4. The "Meet the Market Maker" sidebar formatting (VitePress ::: info block). Same approach as the Chapter 4 "Meet the Searcher" sidebar; renders cleanly in the dev view but the print-pipeline needs a pandoc filter at MANUSCRIPT.md assembly time. Not a Phase 3 blocker.
  5. The "Is this bad?" verdict tone. Lighter than Ch 4's, deliberately. The SPEC said the chapter is structural, not adversarial. If you'd rather it be more adversarial — particularly on the prop-AMM displacement — the §5d argument could land harder. My instinct is to hold the heavier critique for Chapter 11.

Places where the prose got technical and might lose the reader ​

  • §5a, the LVR paragraph. Two sentences on a sigma-squared-over-eight derivation. I kept it tight and put the empirical Topaze data on either side. If the math feels intrusive, the fix is to drop to a one-sentence statement of the result ("at 5% daily volatility, the implied loss is about 11% annualised") and move the derivation to a footnote.
  • §5b, the HyperBFT explanation. Four nouns in one paragraph: HyperBFT, public mempool, matching inside consensus, smart contract calls. I think the load is fair because the section earns it (it explains why nine years of failed on-chain CLOB attempts ended with Hyperliquid), but the architectural detail is the densest in the chapter.
  • §5c, the aggregator-vs-solver distinction. Mechanically the chapter says: "aggregators split a trade across venues; solvers fill an intent from inventory." If a reader misses that distinction, the §5d prop-AMM argument doesn't land. I tried to mark the distinction clearly; worth re-reading aloud to confirm.
  • §5d as a whole. The chapter's most novel structural claim. Three new firms (HumidiFi, SolFi, Tessera), one reported operator (Temporal), and an economic argument about adverse selection. If the reader is going to put the book down somewhere, this is the section.

Files written/modified in Phase 3 (this chapter only) ​

The dev server picks all of this up on restart and renders the chapter at /book/chapters/02_liquidity/DRAFT.

Phase 3 is complete. The chapter is now in Nick's review queue.