TradFi Dark Pools & Crossing Engines

Scope

Workstream 1 of the decentralized crossing-engine research. Surveys how traditional-finance dark pools and crossing/matching engines work — matching mechanics, order types, anti-gaming controls, periodic-auction design — across four reference markets:

1. US equities ATS — Alternative Trading Systems regulated under SEC Regulation ATS, disclosed in detail via Form ATS-N (POSIT/Virtu, Liquidnet, Instinet, UBS ATS, IBKR ATS, IEX, Luminex).
2. EU/UK MiFID II — systematic internalisers, periodic-auction venues, the double-volume-cap, reference-price and large-in-scale waivers.
3. Asia-Pacific — Hong Kong (HKEx ALP regime), Japan (PTS / ToSTNeT), Australia (ASX Centre Point, Chi-X hidden orders).
4. Crypto OTC desks / RFQ block trading — the de-facto dark liquidity of crypto today (Cumberland, B2C2, Wintermute, Galaxy; Paradigm as the multi-dealer RFQ network).

Vocabulary and the 7-axis rubric are inherited from 00-framing.md. This doc notes how the cross price is set but defers deep treatment to WS5; it does not cover regulatory/legal analysis (out of scope per framing).

Key questions

This doc primarily informs Q1 (matching mechanism) and feeds Q5 (reference price). Specifically:

• Continuous crossing vs scheduled crosses vs periodic auctions — how each works and what each defends against.
• Matching priority: time, size, pro-rata, random/uniform midpoint, broker preferencing / anti-internalization.
• Order types: midpoint peg, conditional orders + the firm-up handshake, IOIs (incl. actionable/targeted), minimum quantity / MAQ, MELO and per-order speed bumps.
• Anti-gaming / anti-leakage: speed bumps, minimum execution life, pinging detection, toxicity-based counterparty segmentation, fill-size floors.
• Periodic-auction mechanics (frequent batch auctions; EU periodic-auction venues).
• How the cross price is imported (reference/midpoint) — noted, deferred to WS5.

Findings

A. The three matching regimes

1. Continuous crossing. Orders rest in a hidden book and match the instant a contra arrives, in continuous time, priced off an imported reference (typically the NBBO midpoint in the US, EBBO/PBBO midpoint in the EU). This is the dominant ATS model — POSIT, UBS ATS, Instinet, IEX dark order types. The fill quantity is min(buy, sell) subject to any minimum-quantity floor, rounded down to a round lot [1]. Continuous crossing is operationally simple and gives immediacy, but recreates a continuous-time speed race: whoever reacts fastest to a reference move trades against stale resting interest ("latency arbitrage" / "picking off" the midpoint). Every anti-gaming control below exists to patch this.

2. Scheduled / periodic crosses. Matching happens only at fixed times — e.g. ITG/POSIT's original "Match" product ran point-in-time crosses several times a day; exchange open/close auctions are the canonical lit example. Orders are collected, then crossed in one event at a single reference price. This removes the continuous speed race but at the cost of immediacy — interest waits for the next cross. Volume is concentrated, which deepens each cross but telegraphs its timing.

3. Periodic / frequent batch auctions. A short collection window (often sub-second to a few hundred ms, randomised duration) followed by a uniform-price batch match. This is the modern compromise: near-continuous availability but discrete-time matching. The theoretical foundation is Budish–Cramton–Shim, The High-Frequency Trading Arms Race: Frequent Batch Auctions as a Market Design Response (QJE 2015): a continuous limit-order book builds in mechanical arbitrage rents and a never-ending speed arms race; batching to discrete intervals processes all orders arriving in the same window together, so a speed advantage within the window is worthless [10]. EU "periodic auction" venues (Cboe Periodic Auctions Book, Aquis, Goldman SIGMA X periodic, Turquoise Plato) are the live deployment of this idea.

Maturity: continuous crossing — live & proven, decades (POSIT since 1987). Scheduled crosses — live & proven but declining share. Frequent batch auctions — live & proven in EU since ~2017; FBA-as-primary-market remains paper/proposed (no major exchange runs its main book this way).

B. Matching priority rules

• Price–time priority. Default in continuous books: best price first, then earliest arrival. Time priority is what creates the speed race; it rewards the fastest reactor.
• Pro-rata. At a single match price, an incoming order is split across resting orders in proportion to size. POSIT uses pro-rata for resting peg orders executable at the same price [9]. Pro-rata dilutes the time-priority speed incentive but invites quote-inflation (oversizing to grab a bigger share).
• Size priority. Larger orders ranked ahead — explicitly rewards block interest, the buy-side's stated goal for dark venues. Often a tie-breaker: POSIT uses order-size priority to break ties when satisfying minimum-quantity instructions [9].
• Price / size / time composite. Virtu's 2019 POSIT rebuild moved to price → size → time with price improvement to the liquidity provider [9].
• Random / uniform midpoint allocation. When many orders sit at the same midpoint with no economic differentiator, some venues allocate randomly / uniformly to remove any deterministic edge. This is the priority rule most directly portable to a setting where arrival timestamps cannot be trusted.
• Broker preferencing / anti-internalization. Cboe Periodic Auctions: if a port opts in, allocation is Broker × Price × Volume × Time — a broker's own two-sided flow matches itself first, then general allocation runs; all orders still contribute equally to price formation [8]. The mirror-image control is anti-internalization: ATSs let a subscriber prevent its own buy and sell orders from crossing each other (avoids wash-trade / self-match). Both are counterparty-routing overlays on top of the base priority rule.

C. Order types

• Midpoint peg. Order priced to the reference midpoint; it re-prices as the reference moves so neither side pays spread. The workhorse dark order type. IEX adds M-Peg (pegged to midpoint) and D-Peg (pegged but discretionary — steps off the midpoint, and is withheld when IEX's "crumbling quote" signal fires, i.e. when the quote is about to move) [3].
• Conditional orders + firm-up handshake. A conditional order is non-firm: it lets a participant advertise the same interest in many venues without over-committing. When the venue finds a likely contra, it sends a firm-up request; the participant's system commits real shares and returns a firm order; only then does the match execute. Liquidnet's H2O ATS does exactly this — LPs rest "actionable order flow" conditionally; before execution Liquidnet sends a firm-up request, the LP returns remaining unexecuted shares, and the firm-up protects the LP against over-execution across venues [1]. The risk is the firm-up itself leaks information and can be declined ("fade"), so venues track firm-up reliability.
• IOIs (Indications of Interest). Non-binding advertisements of interest. An actionable IOI is, per the SEC's 2009 concept release, one conveying symbol + side + a price at or better than NBBO + size ≥ one round lot — the SEC proposed treating these as firm quotes precisely because they function as hidden orders [7]. Targeted IOIs are sent to selected counterparties to source contra liquidity; they are a recognised information-leakage vector (false-liquidity signalling, front-running risk) [7]. Cboe LIS uses curated IOIs to invite block negotiation.
• Minimum quantity (MEQ) / minimum acceptable quantity (MAQ). A floor on fill size: the order will not trade unless at least the floor can be filled. Primary defence against information leakage from small "pinging" fills. POSIT MEQ logic is sophisticated — with multiple MEQ orders it perturbs the pro-rata formula with an 80/20 split (80% of an order's pro-rata share allocated normally, 20% reserved to satisfy other orders' minimums), and uses size priority as the tie-breaker when minimums can't all be met [9]. ASX Centre Point Block orders carry a user-defined MAQ and execute only when aggregate executable volume ≥ MAQ [11].
• MELO (Minimum Execution Life Order) and per-order speed bumps. Nasdaq's M-ELO requires an order to rest a holding period before it can match, and M-ELO orders match only other M-ELO orders — an asymmetric, self-segregating speed bump: by accepting a minimum life, an order joins a pool free of latency-arbitrage algos [2]. IEX's 350-microsecond "shoebox" delay is the venue-wide symmetric variant — it delays both liquidity-adding and liquidity-removing orders to neutralise speed [2].

D. Anti-gaming / anti-leakage controls

• Speed bumps. Symmetric (IEX, 350µs, all orders) vs asymmetric (only liquidity-removers delayed, protecting resting quotes). Purpose: stop latency-arbitrage picking off stale midpoints.
• Minimum execution life. MELO as above — time-based, makes an order useless to a sub-millisecond strategy.
• Pinging detection. "Pinging" = sending tiny IOC orders to detect hidden size, then trading ahead of it. Defences: minimum-quantity floors (a ping below the floor never fills, so it learns nothing), and surveillance that flags subscribers whose flow is dominated by tiny immediate-or-cancel orders.
• Toxicity-based counterparty segmentation. The most important leakage control. Venues score subscribers by markout — short-horizon price drift after their fills — and bucket them. Virtu's POSIT segments subscribers into Neutral / Move-Towards / Move-Away tiers via a long-horizon markout methodology, and any subscriber can opt out of interacting with the Move-Towards or Move-Away segment on a per-order basis [9]. Several ATSs let clients restrict interaction to "low-toxicity" tiers to cut adverse selection [12]. Form ATS-N was introduced specifically so subscribers can see how a pool segments flow and how affiliated/principal desks interact [12].
• Fill-size floors. MEQ/MAQ again — a leakage control as well as an order type: it both blocks pinging and ensures block interest only meets block interest.
• Post-trade deferral. EU LIS waivers permit delayed publication of large trades so the market cannot reverse-engineer a working block [13] (a leakage control on the print, not the order).

E. Periodic-auction mechanics (detail)

Cboe Europe Periodic Auctions Book is the reference implementation. An auction match is price determination, then execution allocation [8]:

• Price determination. Not constrained to bid/mid/offer — the clearing price can be any half-tick point in between. The chosen price is the one maximising executed volume within a price collar; ties broken by smallest imbalance, then by proximity to a volume-weighted reference (≈ midpoint) [8]. All orders — including broker-preferenced ones — contribute equally to price formation [8].
• Allocation. Price × Volume × Time, with optional broker preferencing prepended (Broker × Price × Volume × Time) [8].
• Auction trigger & duration. An auction starts when matchable interest exists; duration is short and randomised so the exact uncross time cannot be gamed. Randomised end-time is the key anti-gaming feature — it defeats last-millisecond order placement.
• Regulatory context. EU FBA venues grew because periodic-auction volume is not counted against the MiFID II double-volume-cap on dark trading. ESMA investigated whether "non-price-forming" FBAs were just DVC-circumvention dark pools in disguise and issued an opinion clarifying that genuine FBAs must be price-forming (limit orders, equilibrium price) to qualify as lit [6][4]. The DVC itself caps two waivers — the reference-price waiver (price imported from another venue) and the negotiated-transaction waiver — at 4% per venue / 8% market-wide per instrument [5].

F. How the cross price is set (noted; WS5 owns this)

Every venue here is a price taker: the cross price is imported, not discovered.

• US ATS: NBBO midpoint (consolidated SIP feed). Midpoint-peg = exactly half the NBBO spread.
• EU dark / SI: EBBO or primary-venue BBO midpoint, under the reference-price waiver — by rule the price must derive from a "reliable" reference market [5].
• Periodic auctions: semi price-forming — the clearing price is computed from the auction's own order book but collared to the lit BBO so it cannot stray [8].
• Crypto OTC/RFQ: no canonical consolidated tape — the desk quotes a price built from its own inventory cost, hedge cost on lit venues, and a risk/liquidity premium; RFQ networks (Paradigm) surface the best of several dealer quotes [14][15].

The single most transferable observation for WS5: all of TradFi dark crossing depends on a trusted, manipulation-resistant external reference, and the venue does not produce it. A decentralized cross must supply that reference itself — there is no NBBO for cross-chain assets.

Region subsections

US equities ATS — live & proven

Form ATS-N (mandatory since 2018) makes US ATSs the most transparently-documented dark venues anywhere — matching logic, order types, segmentation and affiliate-interaction rules are public [12]. Representative venues:

• POSIT (Virtu) — continuous midpoint crossing; price/size/time priority; pro-rata with 80/20 MEQ reallocation; Neutral/Move-Towards/Move-Away toxicity segmentation [9].
• Liquidnet (H2O ATS) — buy-side block pool; conditional orders + firm-up handshake; large minimum sizes [1].
• IEX — 350µs symmetric speed bump; M-Peg / D-Peg with crumbling-quote withholding [3].
• Nasdaq M-ELO — minimum-execution-life, self-segregating order type [2].
• UBS ATS, Instinet, IBKR ATS, Luminex — variants on continuous midpoint crossing with per-subscriber counterparty controls.

EU/UK MiFID II — live & proven

Dark trading is structured around waivers and is volume-capped:

• Reference-price waiver — pure dark midpoint crossing; capped by the DVC (4%/8%) [5].
• Negotiated-transaction waiver — bilaterally agreed trades formalised on a venue; also DVC-capped [5].
• Large-in-scale (LIS) waiver — uncapped; the buy-side's main route for genuine blocks. Cboe LIS (with BIDS), Turquoise Plato Block Discovery, Liquidnet sit here, using conditional orders + IOI-driven block invitations [13].
• Systematic Internalisers (SI) — investment firms dealing on own account, bilaterally, outside a multilateral venue; quote-driven, not a crossing network in the multilateral sense [2-SI source].
• Periodic auctions — DVC-exempt, semi-price-forming, randomised-duration (section E).

Asia-Pacific — live but more constrained

Hong Kong, Australia and Japan hold most APAC dark liquidity, but rules are tighter and venue-operated dark pools are more restricted:

• Australia — ASX Centre Point. Exchange-operated dark book; midpoint matching; Centre Point Block orders are midpoint-only single-fill with a user-defined MAQ; Sweep orders dual-post to lit + dark; order preferencing to lift crossing rates [11]. ASIC imposes a meaningful-price-improvement rule — sub-block dark trades must improve on the lit spread.
• Hong Kong — HKEx ALP regime. Broker "Alternative Liquidity Pools" are permitted but tightly governed: institutional-only, broker priority must be behind client orders, and — critically — all trades must be reported and cleared through HKEx, so an ALP can earn matching fees but not clearing/settlement revenue [APAC source].
• Japan — PTS + ToSTNeT. Off-exchange crossing runs through Proprietary Trading Systems (separate PTS licence) and the TSE's ToSTNeT/JNet facility for negotiated and basket/block trades; bank/broker dark pools (Credit Suisse, Instinet, ITG, Chi-X Japan, SBI Japannext) route here.

Maturity: live & proven but with smaller share and stricter price-improvement / reporting obligations than the US.

Crypto OTC desks / RFQ block trading — live, the de-facto crypto dark venue

Crypto has no Reg ATS / MiFID equivalent and no consolidated tape; large trades move OTC:

• Principal desks — Cumberland (DRW), B2C2 (SBI), Wintermute, Galaxy. Core flow is RFQ: client requests a quote for a stated size and side, desk returns a firm price (inventory + hedge cost + risk premium), client accepts or not. Settlement is bilateral/off-chain into custodial wallets, which removes on-chain MEV/front-running but replaces it with full counterparty trust in the desk [14][15]. This is bilateral negotiation, not multilateral crossing — closest TradFi analogue is a MiFID Systematic Internaliser.
• Multi-dealer RFQ networks — Paradigm. A liquidity network (270+ counterparties) where a client requests two-way quotes from several dealers simultaneously, disclosed or anonymous, without revealing direction, and executes on the best bid/offer; block trades clear on the connected exchange (Deribit etc.) [16]. This is the closest crypto analogue to a dark crossing/IOI venue: pre-trade-dark to the wider market, anonymous, but the RFQ hub still sees everything.

Maturity: live & proven commercially, but as a trust architecture it is the weakest — a single desk or RFQ hub is a fully trusted, fully sighted operator, exactly the adversary 00-framing.md wants designed out.

Comparison table

Scored on the 00-framing.md rubric (✅ good / ⚠️ partial / ❌ poor / n/a). "Privacy" here = pre-trade opacity to other participants — in every TradFi case the operator is fully sighted, so trust assumptions are uniformly operator-trust.

Venue / mechanism	Privacy guarantee	Trust assumptions	Liveness / censorship	Throughput / latency	Cross-chain	Maturity	MEV / gaming exposure
POSIT (Virtu) — continuous midpoint cross	⚠️ dark to peers; operator fully sighted	❌ trusted operator sees all flow	⚠️ operator can halt/exclude	✅ continuous, sub-ms	❌ single-market, single-asset	live & proven (since 1987)	⚠️ latency arб on stale midpoint; mitigated by markout segmentation + MEQ
Liquidnet H2O — conditional + firm-up	✅ strong; block-only, conditional orders never firm until match	❌ trusted operator; firm-up leaks to LP	⚠️ operator-run	⚠️ slower — firm-up handshake adds a round-trip	❌	live & proven	✅ low — block floors + conditionals defeat pinging; ⚠️ firm-up fade
IEX speed bump + D-Peg	⚠️ dark order types; operator sighted	❌ trusted operator	⚠️ operator-run	⚠️ deliberate 350µs delay	❌	live & proven	✅ speed bump + crumbling-quote withholding neutralise latency arб
Nasdaq M-ELO	⚠️ dark; operator sighted	❌ trusted operator	⚠️ operator-run	⚠️ minimum holding period by design	❌	live & proven	✅ minimum execution life self-segregates out HFT
Cboe EU Periodic Auction (FBA)	⚠️ short dark window then uncross; operator sighted	❌ trusted operator	⚠️ operator-run	✅ near-continuous; randomised sub-second windows	❌	live & proven (EU, ~2017)	✅✅ batch + randomised duration kill the speed race (Budish et al.)
Crypto OTC / RFQ (Cumberland, Paradigm)	⚠️ dark to wider market; desk/hub fully sighted	❌❌ single trusted desk or RFQ hub; counterparty credit risk	❌ desk can refuse/fade quote	✅ fast quote turnaround	⚠️ multi-asset in practice, but settlement is bilateral not atomic-cross-chain	live & proven commercially	⚠️ no on-chain MEV (off-chain settle) but last-look / quote-fade by the desk

Implications for our engine

What transfers to a decentralized setting

• Periodic / frequent batch auction as the core matching regime. The single strongest transfer. Discrete-time batching is also the natural fit for a decentralized engine: a block / batch interval already exists, sealed-bid + threshold-decrypt maps cleanly onto a batch, and Budish–Cramton–Shim's result — speed advantage within a window is worthless — directly neutralises MEV ordering games. A continuous crossing book would re-import the latency race on top of an adversarial mempool.
• Randomised auction duration. Cheap, proven anti-gaming primitive; a decentralized engine can source the randomness from a VRF / beacon and defeat last-moment placement without trusting an operator.
• Uniform-price clearing + random/uniform allocation at the clearing price. Random midpoint allocation is the only priority rule that survives the loss of trusted arrival timestamps — a decentralized batch cannot trust "time", so allocation is pro-rata or uniformly-random among same-priced orders.
• Minimum quantity / MAQ floors. Directly portable and arguably more important — they defeat pinging, which is the cheapest leakage attack against any hidden book, decentralized or not. Enforce as an in-order parameter the matching circuit checks.
• Conditional orders + firm-up. Maps onto intent + commit–reveal: an intent is a non-firm conditional; the "firm-up" is the reveal/commit step. The firm-up fade problem maps to intent non-execution — handle with collateral/bonding so a revealed intent that fails to settle is penalised.
• Toxicity-based counterparty segmentation. Transferable in principle — markout scoring needs only post-trade price data — but it requires identity or persistent pseudonymous reputation, which is in tension with a permissionless decentralized venue. Likely a reputation/staking layer rather than a KYC tier list.

What does NOT transfer

• Operator-as-trusted-matcher. Every TradFi venue's privacy is "dark to peers, transparent to the operator." The framing explicitly designs the operator out — so all TradFi privacy guarantees must be rebuilt cryptographically (ZK / MPC / TEE / encrypted-mempool — WS2's job). The matching logic transfers; the trust model does not.
• Imported reference price (NBBO / EBBO). TradFi crossing is parasitic on a lit market's consolidated quote. Cross-chain multi-asset pairs have no NBBO and no consolidated tape — WS5 must construct a manipulation-resistant reference from scratch. This is the hardest non-transfer.
• Targeted / actionable IOIs. IOIs assume a trusted broker selectively routing advertisements. In a decentralized venue a "targeted IOI" is just a selective information leak with no trusted router — drop it in favour of uniform sealed-bid submission.
• Broker preferencing / anti-internalization by identity. Depends on knowing broker identity and on a trusted operator applying the priority overlay; doesn't map onto a permissionless, identity-light setting.
• Continuous-time price-time priority. Relies on trusted, ordered timestamps the operator assigns. In an adversarial-mempool / multi-proposer setting there is no trustworthy "time" — this is the priority rule to abandon, reinforcing the batch-auction choice.
• Post-trade deferral as a leakage control. Assumes a trusted reporting facility that can withhold a print. On a public chain settlement is visible; pre-trade privacy must do all the work — there is no post-trade deferral.

Open questions

1. Batch interval length. EU FBAs run sub-second; a cross-chain decentralized engine is bounded by block times and threshold-decryption latency. What interval balances immediacy against giving sealed-bid privacy time to work? (Feeds Q1.)
2. Allocation without trusted time. If random/uniform allocation replaces time priority, what randomness source is both unbiased and unmanipulable by a matcher — and how is fairness proven to participants? (Feeds Q1.)
3. Decentralized toxicity segmentation. Can markout-based counterparty scoring work against pseudonymous, possibly Sybil identities? Is a staking/bonding reputation layer enough, or does adverse selection force some identity? (Feeds Q1 / fee model Q4.)
4. Firm-up / conditional fade penalties. A revealed intent that fails to settle is the decentralized analogue of firm-up fade — what bond size deters it without excluding honest participants who genuinely lost the contra? (Feeds Q3.)
5. MEQ enforcement under encryption. Minimum-quantity floors must be checked inside the matching computation while orders are still hidden — does this constrain the privacy mechanism choice (ZK circuit vs MPC vs TEE)? (Feeds Q2.)
6. Reference price for cross-chain pairs. There is no NBBO — every TradFi mechanism here assumes one. Deferred to WS5 but flagged as the hardest non-transfer.

Sources

1. Liquidnet, Inc. — Form ATS-N/MA (H2O ATS), conditional orders & firm-up: https://www.sec.gov/Archives/edgar/data/0001110644/000111064422000009/H2OExh_3_23feb2022.pdf
2. IEX speed bump & Nasdaq M-ELO as asymmetric speed bump (Traders Magazine / Coalition Greenwich): https://www.greenwich.com/blog/case-asymmetric-speed-bumps-us-equities
3. IEX — Breaking Down M-Peg and D-Peg (crumbling-quote withholding): https://www.iex.io/article/breaking-down-m-peg-and-d-peg
4. ESMA Opinion — pre-trade transparency & price determination in frequent batch auctions: https://www.esma.europa.eu/press-news/esma-news/esma-opinion-clarifies-application-pre-trade-transparency-and-price
5. Double Volume Cap (DVC) transparency regime under MiFID II — reference-price & negotiated-transaction waivers: https://www.emissions-euets.com/internal-electricity-market-glossary/2021-double-volume-cap-transparency-regime-under-mifid-ii
6. ESMA — call for evidence on Frequent Batch Auctions and DVC circumvention: https://www.esma.europa.eu/press-news/esma-news/esma-address-regulatory-concerns-over-frequent-batch-auctions
7. SEC 2009 concept release on actionable IOIs / dark pools (Cadwalader summary): https://www.cadwalader.com/resources/clients-friends-memos/securities-and-exchange-commission-proposes-regulation-of-indications-of-interest-and-dark-pools
8. Cboe Europe Equities — Periodic Auctions Book Guidance Note (price determination, broker preferencing): https://cdn.cboe.com/resources/participant_resources/BCE-GuidanceNote_Periodic_Auctions_Final.pdf
9. Virtu — new matching rules & segmentation logic for US POSIT ATS (price/size/time, 80/20 MEQ pro-rata, Neutral/Move-Towards/Move-Away): https://www.tradersmagazine.com/departments/brokerage/virtu-changes-matching-rules-for-us-posit-ats/
10. Budish, Cramton & Shim — The High-Frequency Trading Arms Race: Frequent Batch Auctions as a Market Design Response (QJE 2015): https://academic.oup.com/qje/article/130/4/1547/1916146
11. ASX Centre Point — Block orders, MAQ, Sweep order types, preferencing: https://www.thetradenews.com/guide/asx-centre-point/
12. Are Dark Pools All the Same? Form ATS-N Says "No" — segmentation & affiliate interaction (Traders Magazine): https://www.tradersmagazine.com/news/are-dark-pools-all-the-same-form-ats-n-says-no/
13. MiFID II's Trading Hereafter: Systematic Internalizers & Block Venues — LIS waiver, Cboe LIS/BIDS, Turquoise Plato (FlexTrade): https://flextrade.com/resources/mifid-iis-trading-hereafter-systematic-internalizers-block-venues/
14. What Is OTC Trading in Crypto — RFQ mechanics, off-chain settlement, principal desks (DEXTools): https://www.dextools.io/tutorials/what-is-otc-trading-in-crypto-guide-2026
15. Best Crypto OTC Desks for Large Block Trades — Cumberland, B2C2, Wintermute, Galaxy principal model: https://blog.tokenmetrics.com/p/best-crypto-otc-desks-for-large-block-trades-2026
16. Paradigm — Multi-Dealer & Anonymous RFQ (anonymous two-way quotes, best-bid/offer execution): https://www.paradigm.co/blog/multi-dealer-rfq-launched