# DefiSilicon documentation (full text) Source: https://www.defisilicon.com/docs/ . Status date: see each page. Written in ASD-STE100. --- # Welcome ## Our objective We build an exchange with three properties: | Objective | What it means | |---|---| | **Distributed sequencer** | No single company sets the order of trades. Several independent sites stamp each order, and a published rule combines their stamps. | | **Wire stamps** | An FPGA card stamps each order when its bytes arrive at the network port, before any software touches it. | | **No front-running** | No person can put an order before an order that arrived earlier. The sites sign each step. Any person can check it. | Section 1.2 shows where we are against each objective. > **Status, 20 Nov 2026.** DefiSilicon is a testnet with test funds only. The last chain ran for 7 days and 4 hours without a stop. Then the operator stopped it. A new chain will start later. ## How to read these pages Parts 1 to 3 explain the protocol and give the measured numbers. The time to read them is 20 to 30 minutes. Part 4 is reference for traders and developers. Part 5 is the roadmap. The appendix has the glossary and the FAQ. Programs and AI agents can read the same content as plain text: [llms.txt](https://www.defisilicon.com/docs/llms.txt) is the index, and [llms-full.txt](https://www.defisilicon.com/docs/llms-full.txt) is the full book. We write in ASD-STE100 Simplified Technical English. Appendix A gives the technical words that we add to the STE vocabulary. ## Numbers and status Each number has a date. If a number is not a measurement, the text tells you: "design target" or "calculated". A part that does not exist is "not built". --- # 1.1 The objective We build an exchange with a **distributed sequencer**, **wire stamps** and **no front-running**. ## The problem On an exchange, the order of arrival decides who trades first. Today, one company's server decides this order. No person outside the company can examine that decision. - A person with an early or private view of the order flow can trade before other orders. - Market makers know this risk. They quote wider prices, and the order book becomes thin. - A venue that applies its rules correctly cannot prove this. - Liquidity stays away from venues that it cannot check, and markets become disconnected. ## How the design meets each objective Section 1.2 tells how far the testnet is today. | Objective | How | |---|---| | Distributed sequencer | Each site has an FPGA card. Each card stamps each order. A fixed rule in the cards calculates the official time from all the stamps. No site is the master. | | Wire stamps | The card records the time when the bytes of an order arrive at its network port. No operating system or software touches the order before the stamp. | | No front-running | The official order follows the stamps. The sites sign the stamps and the results, and each batch names the rule. Each site executes the same list and signs the result. Any person can replay it. A site that gives false data makes evidence against itself and pays a penalty from its stake. | The design is also made for low latency. Stamps, order, signatures and checks are in hardware, and the match is a few microseconds in software. Section 3.2 gives the measured latency and the design budget. Figure 1. One order, three stamps, one official time, one list, FINAL. Loop of 5 seconds. One signed order goes to three sites. The card of each site stamps it. The public rule gives one official time. Each site has the same list. A quorum signs, and the batch is FINAL. (image: https://www.defisilicon.com/docs/assets/figures/f1-objective.svg) ## Why it matters Our objective is fair and connected markets. Fair, because no person can move before your order after it arrives. Connected, because traders who can check the order of trades put their liquidity back on the venue. Exchanges use FPGAs for speed. Blockchains use signatures. We use both for one purpose: an order of trades that no single operator controls, and that any person can check. --- # 1.2 Where we are Part 2 describes the production system. This page compares the testnet with that system and with each objective. Status is at 20 Nov 2026. ## Progress against the objective | Objective | Today | What remains | |---|---|---| | Distributed sequencer | The sequencer runs in one FPGA card (site 0). Two software sites stamp, check and sign each batch. A batch is final when site 0 and one more site sign. | Three FPGA cards with synchronized clocks (M0.5). Then the stamps of all sites count. Later, sites that other operators run (V2). | | Wire stamps | In operation at the port of the card at site 0 | Hardware stamps at all three sites (M0.5) | | No front-running | No person can put an order before an order that arrived earlier at the card without signed evidence. Site 0 can reject orders, or delay an order by up to 1.5 ms. | A new ordering rule that removes the 1.5 ms delay and the choice of ties. More cards where orders can enter. | Figure 2. The testnet today and the production design. Left: the testnet today, one FPGA card at site 0 and two software sites, one operator. Right: the production design, six sites with a card each on a ring, at least three operators. (image: https://www.defisilicon.com/docs/assets/figures/s1-today-vs-production.svg) ## Latency The full path on the last chain was approximately 226 ms. Of this, an order waits 134.5 ms because each batch waits for the certificate of the previous batch. The card part of the path was 0.33 ms in a silicon test. The first step is to let the batches overlap again. Then the full path goes back to approximately 22.5 ms. The next steps are a short watermark with synchronized clocks, and FPGA cards at all sites. See 3.2. ## Main numbers | Item | Value | |---|---| | Longest run of a chain | 7 days 4 hours, 13 to 20 Nov 2026 | | Batches in that run | 5,274,059, all certified | | Commands executed in that run | 12,831,017 | | Trades (fills) in that run | 5,201,559 | | Maximum throughput, load test of 1 Nov 2026 | 1,035 commands each second for 120 s | | Clock of the FPGA card | 140 MHz | ## Also in operation - The deterministic matcher at each site. - Signed evidence for policing, and a stake of test USD for each validator. - An independent reader that replays and verifies all history. - The BTC-PERP market, the website, the command line, and keys for AI agents. ## Not built - Operation if site 0 stops. Today, the exchange stops. - Validators that are not the operator. Today, one operator runs all three sites. - Six sites, the bridge to another chain, and the snapshots each 100 ms. - The matcher in a second FPGA. Today, the matcher runs in software. - Replicas without approval. Today, the operator approves the key of each outside reader. - Trade with real money. ## Next steps 1. Start a new chain with the fixes from the last run. 2. Let batches overlap again, to remove the 134.5 ms wait. 3. Install three FPGA cards on a fiber triangle (M0.5). 4. Put the new ordering rule in the cards. 5. Decrease the latency of the full path. 6. Put outside market makers and AI agents on the testnet. Part 5 gives the hardware steps and the roadmap. --- # 2.1 The system Part 2 describes the production system: the exchange that we build. Section 1.2 tells which parts run on the testnet today. ## Sites The exchange runs on a ring of sites in one metro area. Each site is in a different building, with its own power and fiber. Independent companies operate the sites. The full design has six sites and at least three operators. The first builds have three sites. Figure 3. Six sites on a metro ring. Every site is a door. Six sites on a dual fiber ring in one metro area. Each site has a card, a matcher and a read server. Clients connect to two or three sites through passive taps. (image: https://www.defisilicon.com/docs/assets/figures/s2-system.svg) Each site has these parts: | Part | Function | |---|---| | Sequencer card (FPGA) | Owns the network ports. Stamps each packet when the light becomes bits. Signs observations and results. Calculates the canonical order. | | Matcher | A full replica of the exchange: order book, balances, positions and margin. It executes the canonical order. It runs in software first, and in a second FPGA for the highest rates. | | Read server | Receives each final batch from the card, one way. It serves the history to readers. | | Fiber links | Dual fiber to the other sites, approximately 10 km around the ring. The sites measure the delay of each link one time and sign it. | | Clocks | A clock synchronized through the fiber, and a rubidium atomic clock that checks it (see 2.2 and 2.6). | ## No main entrance Every site is a door. A passive optical tap, which is glass without electronics, sends the light that arrives to the local card. A client sends the same signed order to two or three doors. The cards identify copies by the hash of the order. Because of this, no single door can block an order, and a door that delays an order shows itself in the time data. ## Three broadcasts and one silence 1. **Deliver.** Each door that receives an order forwards the packet to the other sites. This is approximately 50 µs across 10 km of fiber. 2. **State what you saw.** Each card broadcasts one signed observation: the hash of the order and its wire stamp. Each card collects the observations and calculates the canonical order with the same public rule. The same inputs and the same rule give the same order at each site. 3. **Silence.** Each matcher executes the canonical order alone. Six independent computations must agree, because the matcher is deterministic. 4. **State what you calculated.** Each card broadcasts one signed digest of its results. When a quorum of identical digests exists, the batch is FINAL. With six sites, the quorum is five. With three sites, it is two of three, or all three. There are no questions and no votes. There are only statements and arithmetic. The quorum certificate is the settlement record, and any person can verify its signatures. The time from the wire stamp to a final trade is approximately 0.3 to 0.4 ms (design target). ## Where everything lives - **The present** is in the matcher of each site: the book, each balance and each position. - **The past** is on the read servers: each order, its signed stamps, its fills and the certificates. This stream is the chain. Any person can hold a copy, verify each signature and serve it again. - **Money** enters the same queue. Deposits and withdrawals are commands like all others: stamped, ordered and executed. No administrator path into the state exists. - **Snapshots.** Each 100 ms, the sites sign one root of the full state. A bridge contract on another chain accepts a root with five of six keys. A new or restarted site loads a snapshot and replays the stream to the present. ## The life of one order ``` 1. client signs the order and sends it to two or three doors 2. each door card: wire stamp at the port, session and rate checks -> ACK 3. each door forwards the packet to the other sites (~50 us) 4. each card broadcasts its signed stamp 5. each card: official time by the public rule, short wait, canonical order 6. each matcher executes the canonical order -> receipt 7. each card signs the digest of the batch and broadcasts it 8. quorum of identical digests -> FINAL 9. read servers and replicas verify and serve the history -> proof ``` Section 2.3 explains steps 2 to 8. --- # 2.2 Time and wire stamps ## The wire stamp Each card has a counter that increases with its clock. When the first bits of an order come out of the light at the port, the card records the counter value. This value is the **wire stamp**. - No operating system or software touches the order before the stamp. No software can add delay before the stamp. - The host never sees the order before the stamp. The host cannot hold an order and give it an earlier time. - The stamp goes into the signed output of the card. No person can change it later. Software cannot be the witness. Between the wire and a program sit buffers, interrupts and schedulers. They add tens of microseconds of random delay. Six software observers of one packet give six different sequences. A stamp in the FPGA at the port has a precision of nanoseconds and a jitter of nanoseconds. Figure 4. The same five packets, stamped at the port and stamped in software. Loop of 5 seconds. Five packets arrive in order. The port stamp keeps the sequence 1 2 3 4 5. The software stamp, after buffers and interrupts, gives 1 3 2 5 4. (image: https://www.defisilicon.com/docs/assets/figures/f2-wire-stamp.svg) ## One time base for all sites Six stamps are useful only if the six clocks agree. The design has three layers: 1. **Clock messages through the fiber.** The cards exchange messages with hardware stamps (PTP) and measure their offsets. No clock is the master. The design target is agreement in 5 to 15 ns. White Rabbit technology can later give agreement below 1 ns. 2. **Calibration.** The delay of each fiber link is measured one time, in a public ceremony, and all sites sign the table. With the table, each site converts the stamp of another site to its own time base. 3. **Atomic clocks.** Each site has a rubidium atomic clock. It does not set the time. It checks the rate of the local clock. A clock that moves away from its atomic clock makes evidence against its site. Section 2.6 explains this. ## Two notes - Light in fiber goes approximately 5 µs for each kilometer. A near site sees an order before a far site. The rule corrects this with the calibration table. This is correct physics, and it is not a fault. - World time is not necessary for fair order. The clocks must agree with each other only. --- # 2.3 The sequencer The sequencer sets the order of trades. On many blockchains, the sequencer is a server that one company controls. Here, no site is the sequencer. Each card calculates the same sequence from the same stamps, with a fixed public rule. The sequencer is distributed, and any person can check each decision. This part is the most difficult to understand. This section explains it step by step. ## The short version 1. You sign an order one time. You send the same bytes to two or three sites. 2. Each card stamps the order when it arrives, and forwards it to the other sites. Each card broadcasts its stamp one time. An order has one stamp from each site. 3. Each card converts the stamps to one time base and applies the public rule. The rule gives the official time of the order. One site cannot move it by more than a published bound. 4. Each card waits a short time for the stamps of the other sites. Then it releases the orders in time order. This list is the **canonical order**. It is the same at each site. 5. Each matcher executes the canonical order. A batch is final when a quorum of cards signs the same result. 6. If too many orders arrive, a card rejects some of them immediately. The client sends the same order again. A card does not discard an accepted order. The five pages of this section answer the questions that people ask most. The topics are copies, stamps, the canonical order, overload, and the limits of a site. --- # 2.3.1 Relay and copies Each site must have each order, because each site calculates the same sequence and executes the same list. This page tells how orders go between the sites, and what occurs with copies. ## Step 1: the client sends You sign an order one time. It is a message of 256 bytes. It contains your account, a nonce and the order. The nonce is a counter that makes each order unique. Your client sends the same bytes to two or three doors. The default of the official client is more than one door. If one door delays or drops your order, another door delivers it. ## Step 2: each door forwards A card that receives an order from a client is a door for that order. It stamps the order and forwards the packet to all other sites, without a wait. It also broadcasts its observation: the hash of the order and its wire stamp, with a signature. A card that receives the order from another door does not forward it again. It only stamps it and broadcasts its observation. ## Number of copies For one order that the client sends to two doors, on six sites: | Location | Copies that arrive | Which copy counts | |---|---|---| | Each of the two doors | 1 from the client, 1 from the other door | The first copy that arrives | | Each of the four other sites | 2, one from each door | The first copy that arrives | Each card broadcasts one observation for each order. An order has six observations, one from each site. Figure 5. Two doors, one forward each, six stamps. Loop of 5 seconds. The client sends one order to two doors. Each door forwards it one time to the other sites. Each site keeps the first copy and drops the second. (image: https://www.defisilicon.com/docs/assets/figures/f3-relay.svg) ## When a site receives a second copy Each card identifies an order by the hash of its 256 bytes. The same bytes always give the same hash. A card identifies a copy immediately. - The first copy gives the stamp. The card ignores the other copies. - The card rejects a copy that arrives after the order is in the canonical order. The matcher also rejects it, because the ledger used its nonce before. An order never executes two times. The ledger uses each nonce one time only. Because of this, it is always safe to send an order again. ## If a site is not available Each card keeps a queue of records for each other site. If a site is not available for a long time, its queue becomes full. Then the card discards the records that do not fit and marks a gap. When the site comes back, it loads a snapshot and reads the records that it did not receive. The chain continues with the other sites. --- # 2.3.2 Timestamps Each site gives one wire stamp to each order. The public rule combines the stamps into one official time. This page tells how the rule works, and what one false stamp can do. ## The stamps With six sites, an order has six stamps. The card of each site makes its stamp in hardware when the order arrives at its port. Each card writes a receipt for each order before it broadcasts its stamp. A site signs its receipts in groups. A site cannot change a stamp later. ## One time base The stamps of different sites are not on one clock. Each card converts a stamp of another site with two values. The first is the offset of the clock of that site, from the clock messages. The second is the fiber delay from the door to that site, from the calibration table. After this correction, the stamps of one order are close to each other. The distance between them is the noise of the clocks and the links. ## The rule The rule has three properties. Any person can check them from public data. - **No inversion.** No site and no client can put a later order before an earlier order. "Later" means later by more than a published noise bound. - **No chosen ties.** Orders that are closer than the bound are a group. Random data that no person knows at the time of the signature sets the order in the group. - **Accountability.** If a site uses the small influence that it has, the published statistics of that site show it. For each order: 1. Find the median of the corrected stamps. One false stamp cannot move the median outside the range of the correct stamps. 2. Clip each stamp to a band around the median. The band is the distance between the stamps of one order on 99.9% of orders, measured in the previous period. 3. Calculate a weighted average of the clipped stamps. A site with a more precise clock has more weight. No site has more than 45% of the weight. The result is the official time. One false stamp can move it by a maximum of a published bound. In 200,000 tests of the rule in a software model, with one site that gives false stamps, the result stayed inside the bound every time. Figure 6. Six stamps, one official time, one false stamp. Loop of 5 seconds. Six stamps of one order on a time line. The rule finds the median, clips the stamps to a band and takes a weighted average. One false stamp moves the official time less than the bound. (image: https://www.defisilicon.com/docs/assets/figures/f4-rule.svg) ## Simultaneous orders Two orders closer than the noise bound are simultaneous. No clock can tell which came first. The rule puts them in one group, and random data sets the order in the group. The data comes from a hash of the group, or from a value that all cards calculate together. A client does not know the group when it signs. A site cannot select the value. Because of this, speed below the noise bound has no value. ## What the rule publishes Each period, the sites publish the noise bound, the band, the weights and the statistics of each site. A site that is always at the edge of its influence shows in the statistics (see 2.4). --- # 2.3.3 The canonical order and finality An order with an official time does not have a place in the list at that moment. The stamps of far sites are on the fiber for tens of microseconds. Each card solves this with a **watermark**. ## The watermark ``` watermark = card clock - lag (the card updates it each 50 us) ``` The lag is more than the time of the slowest path between the sites, plus the time to broadcast a stamp. On a 10 km ring, the lag is approximately 0.1 to 0.2 ms (design target). The watermark means this: all stamps of orders with a time before this value are here now, or they will not come. The watermark never goes back. ## Release in time order Each time the watermark moves, the card releases all orders with an official time before the watermark: 1. The order with the earliest time goes first. 2. Orders in one simultaneous group go in the random order of the group (see 2.3.2). Each released order becomes a **CANON** record. The list of CANON records is the **canonical order**. Each card calculates it from the same stamps with the same rule. Because of this, the canonical order is the same at each site. If the lists are different, the digests are different, and the batch is not final. The card gives the canonical order to the matcher of its site through PCIe. The matcher executes it. Only execution goes through PCIe. All traffic between the sites goes through the ports of the cards. Figure 7. Release in time order, then five identical digests. Loop of 5 seconds. The watermark moves along the time line and releases the orders behind it in time order. A late stamp is rejected. Five identical digests of six make the batch FINAL. (image: https://www.defisilicon.com/docs/assets/figures/f5-canonical-order.svg) ## LATE orders If a stamp arrives after the watermark passed the time of its order, the card cannot put the order in its correct place. The card rejects the order as **LATE**. The order does not execute, and its nonce stays unused. The client sends the same order again. ## Batches and finality The matcher executes the canonical order in batches. For each batch: 1. The matcher executes the orders and calculates the state root. 2. The card calculates the **digest** of the batch: the orders, their order, the state root and the stamps that it used. The card signs the digest and broadcasts it one time. 3. Each card checks the signatures of the other cards in hardware and counts the identical digests. 4. When a quorum of identical digests exists, the batch is **FINAL**. With six sites, the quorum is five. Why is this final? A different history also must have five signatures. Two sets of five among six sites share four sites. Four operators must sign two different statements each. Each double signature is public proof, and it costs the bond of its operator. ## Three confirmations | Confirmation | Meaning | |---|---| | ACK from a card | The card accepted your order and stamped it. The order is not in the canonical order at this time. | | Execution receipt | Your order executed in a batch at a site. The batch is not final at this time. | | FINAL and certificate | A quorum of cards signed the digest. The certificate is the settlement record. | The time from the wire stamp to FINAL is approximately 0.3 to 0.4 ms in one metro area (design target). --- # 2.3.4 Overload Our rule for overload is: reject at the edge, and never discard an accepted order. If a card cannot accept an order now, it tells the client immediately. The order has no stamp. The client sends the same bytes again. ## Three limits of each card 1. **Rate limit of the session.** Each session has a token bucket on the card. If the bucket is empty, the card answers **Rate**. One client cannot use all the capacity of a card. 2. **Credit window.** A card can hold a fixed number of accepted orders that have no decision. If all credits are in use, the card answers **Credits**. 3. **Barrier.** Near some batch operations, the card stops new orders for a short time. It answers **Barrier**. The credit window makes overload safe. An accepted order always has a position in the table of the card. The card always makes a decision for each accepted order: release or rejection. ## What the client does | Answer | Meaning | What the client does | |---|---|---| | ACK | The card accepted the order | It waits for the receipt | | Rate | The bucket of the session is empty | It sends again after some milliseconds | | Credits | The card is full now | It sends again after some milliseconds | | Barrier | Short stop near a batch | It sends again after some milliseconds | | LATE | A stamp was too slow | It sends the same order again | The official client does this automatically. A rejection costs time only. It does not use your nonce or your funds. With two or three doors, a rejection at one door does not stop your order: another door delivers it. Figure 8. Rejection at the edge. No accepted order is discarded. Loop of 5 seconds. Six orders go to a card. The bucket of the session fills. The card rejects two orders immediately, with no stamp. The client sends them again and the card accepts them. (image: https://www.defisilicon.com/docs/assets/figures/f6-overload.svg) ## How the pressure moves back ``` one site is slow -> batches wait for FINAL -> the matcher of each site stops new batches -> the cards keep their credits -> the cards answer Credits -> clients send again ``` No step removes accepted work to make space. Only the number of rejections increases, and clients see them immediately. If two of six sites are dark, the exchange stops. Safety comes before continuity. ## What overload does not do - It does not change the order of accepted orders. - It does not sell a position in the list. There is no fee auction. - It does not hide rejections. Section 3.2 gives the measured throughput of the testnet. --- # 2.3.5 What a site can and cannot do A claim about fair order is only as good as its limits. This page gives the limits of each part of the production design: what it can do, and what it cannot do. | Part | It can | It cannot | |---|---|---| | The optical tap | Be cut. This is an outage. | Give false data. It is glass without electronics. | | A door | Delay or drop the orders that enter through it. The client has other doors, and the time data shows the delay. | Change the order of trades. The rule over all stamps decides. | | A door | | Make a false order. The client signs each order. | | One false site | Move a stamp inside the published bound | Move an order past an earlier order by more than the bound | | One false site | Sign a different result | Make it final. It stands alone against five identical digests, and its signature is the evidence. | | One dark site | Stop its own doors | Stop the exchange. Five of six sites continue. | | Two dark sites | | The exchange stops. Safety comes before continuity. | | Four sites that work together | Make a false history | Hide it. Each signature is permanent public evidence. | ## The assumption, stated plainly Four of six sites that work together can make a false history. The defenses are structural: - At least three independent companies operate the sites. - Each signature is permanent public evidence. One operator that speaks, or one client with its own send log, shows the false history and removes all bonds. - The unsettled value of the fast tier has a cap: a published multiple of the total bonds. ## Outside the ordering machine - **The oracle** gives the prices for liquidation. It is a median of the feeds of the operators, with the same logic as the stamps, and with limits that stop a false price. - **The bridge** moves one stablecoin to and from another chain, with a cap on the total value. These are the two other risk surfaces of the design. ## What you can do - Send each order to two or three doors. The official client does this by default. - Keep your own send log. It is evidence of when your order existed. - Run a replica, or read from a replica that you select. --- # 2.4 Consensus and policing ## Finality A batch is final when a quorum of cards signs the same digest. With six sites, five must sign. The certificate with these signatures is the settlement record. There is no later or stronger finality. ## Why a false statement costs money Each operator posts a bond: value that a proven offence removes. A card signs each statement. The signatures make three offences provable from public data: | Offence | What the proof shows | |---|---| | Double signing | One key signed two different statements for the same slot: two digests for one batch, or two versions of one stamp window | | Wrong result | A site signed a result that the certified inputs do not give. A replay of the inputs with the deterministic matcher shows it. | | Impossible claim | A stamp that physics does not permit. Examples: time that goes back in one clock, a reply before its request, or a stamp earlier than the light can arrive. | Each proof uses only signed data. Each site and each reader calculates the same verdict. A proven offence removes bond, and the sites exclude the site at the next checkpoint. ## Clocks The clocks police each other through the fiber, and each clock has its atomic reference. A stamp that does not agree with the fiber measurements or with the atomic clock is evidence. Section 2.6 gives the details. ## Operators - **A new operator** makes its keys in its own hardware security module. The sites add its fiber, run the calibration ceremony again and sign the new table. The new site then runs in shadow mode: it stamps, replays and signs, but its signatures do not count. When its results are identical with the quorum for some days, it posts its bond, and the sites activate it at a checkpoint. - **More sites make the system stronger.** With seven sites, the quorum is five of seven, and the system continues with two dark sites. - **A site is keys, not boards.** A spare machine loads the last snapshot, replays the stream, and continues the site in minutes. Figure 9. How a new operator joins. Five steps: own keys, calibration, shadow mode, bond, live. (image: https://www.defisilicon.com/docs/assets/figures/s3-onboarding.svg) --- # 2.5 Execution and proofs ## The matcher Each site executes the canonical order with the same matcher. The matcher is a fixed, bounded state machine: the order book, balances, positions, margin and funding of a perpetual exchange. It uses integers only, and its only input is the canonical order. The same input always gives the same output at each site. The margin check is part of the match. The question "can this account pay for this order?" is a few memory reads, not a request to another system. The matcher runs in software first, with a time of microseconds for each order. For the highest rates, it runs in a second FPGA at each site. ## The state root After each batch, the matcher calculates a **state root**: one hash of all balances, positions, orders and results. The card puts the state root into the digest. The other sites compare it when they sign. ## The certificate A **certificate** proves that a batch is final. It contains: - The digest and the signatures of the quorum. - The context: the chain, the keys and the rules. - A commitment to the stamps that set the official times. ## Proofs for one value A Merkle proof is a short chain of hashes from one value to the state root. With it, you can prove one balance, position, order or receipt without the full state. The official client checks these proofs on your computer. A balance that it shows as "proof-verified" agrees with a final certificate. Figure 10. A Merkle proof for one balance. Eight values hashed in pairs up to one state root, with a certificate above it. Three hashes from the other branches prove your balance. (image: https://www.defisilicon.com/docs/assets/figures/s4-merkle-proof.svg) ## Readers and replicas The card streams each final batch to the read server of its site, one way. The full stream is the chain: the orders, the signed stamps, the fills and the certificates. Any person can download it from any site, verify each signature, replay each batch and serve it again. Readers never slow the cards: the cards do not know them. ## Snapshots and the bridge Each 100 ms, the sites sign one root of the full state. A contract on another chain accepts a root with five of six keys, and moves one stablecoin in and out. The same snapshots are the restart points of the sites. ## Agent keys Each bot or AI agent can have its own key with its own nonce sequence. You can revoke one key and keep the others (see 4.3). --- # 2.6 The atomic clock This page explains the clocks of the sites in detail. ## The idea Each site keeps a rubidium atomic clock, which is stable for days. No site receives its time from another site. World time is not necessary for fair order. Clocks that agree with each other are necessary. The sites measure each other's clocks through fiber. No clock is the master. A public rule calculates the official time from all clocks. A clock that gives a false time does not move the official time. It only shows itself. ## Police from two directions - **From inside.** The atomic clock of a site is a flywheel. It is stable, and it is difficult to push. If a person changes the GPS signal of a site, the GPS time goes away from the flywheel. The difference shows immediately. - **From outside.** The other sites measure the clock of each site through the fiber. They compare the result with a signed table of fiber delays. Changes to the fiber or false stamps break this comparison. A difference is evidence. A site signs each claim. Physics tests the claim. If a signed claim is physically impossible, it is a proof against the site, and it costs stake. ## The parts | Part | Function | |---|---| | PTP with hardware stamps | Clock messages between the cards. It measures offset and delay. | | Rubidium clock (SRS FS725) | The local flywheel. It checks the rate of the clock of the card. | | Timing shelf with three AMD Kria KR260 boards | Connects each rubidium clock to its card through fiber | | GNSS receiver | Gives the date and a cross-check. We never use it as a time source. | | Calibration | We measure the delay of each fiber one time and sign the table. | The first production builds use PTP with hardware stamps, with a design target of 5 to 15 ns. White Rabbit technology from CERN can later give agreement below 1 ns. Figure 11. The clocks of one site in the M1 design. Not built. A rubidium clock checks the rate of the card clock through a timing shelf. A GNSS receiver gives the date and a cross-check. The fiber connects the clocks of the sites. (image: https://www.defisilicon.com/docs/assets/figures/s5-clocks.svg) --- # 2.7 The BTC-PERP market BTC-PERP is a perpetual future on BTC with USD margin. A perpetual has no end date. Funding payments keep its price near the BTC price. It is the first market of the exchange. ## Contract | Item | Value | |---|---| | Lot (size step) | 0.001 BTC | | Tick (price step) | 1 USD | | Position limit for each account | 10 BTC | | Fees | none on the testnet. Production fees are not set. | | Margin | isolated: one position for each account | | Leverage | 1x to 20x | | Initial margin | 1 divided by the leverage (5% at 20x) | | Maintenance margin | 2.5% at all leverage values | ## Index and mark price An oracle publishes the **index** two times each second. The index is the median of three sources: Coinbase, Kraken, and Binance (BTC/USDT, converted to USD). At least two sources must be new and must agree. The **mark price** is the median of three values that come from the index and the order book. The mark price sets margin and liquidation. ## Funding The matcher measures the premium of the book against the index. Each hour, the average premium divided by 8 becomes the funding rate. The maximum is 1%. ## Order types | Type | Behavior | |---|---| | Limit (GTC) | Trades what it can, then stays on the book | | IOC | Trades what it can, then cancels the remainder | | Post-only | The exchange rejects it if it can trade immediately | | Reduce-only | Can only make a position smaller | | Post-only replace | Cancels and replaces in one command | ## Liquidation When equity goes below maintenance margin, the matcher cancels the resting orders of the account. Then it sends IOC orders into the book. If the equity continues below two thirds of the maintenance margin, a backstop vault receives the position. An insurance account pays a deficit, up to its collateral. --- # 3.1 Seven days nonstop The last chain ran from 13 Nov 2026 13:29 UTC to 20 Nov 2026 17:29 UTC. The operator stopped it. Site 0, site 1, the oracle and the bots did not restart during the run. Site 2 failed from 15 Nov (see below). ## The run | Item | Value | |---|---| | Duration | 7 days 4 hours (619,200 s) | | Batches | 5,274,059, all certified | | Average batch rate | 8.5 each second | | Commands executed | 12,831,017 | | Average command rate | 20.7 each second | | Trades (fills) on BTC-PERP | 5,201,559 | | Volume | 181,484.625 BTC (test funds) | | Restarts of site 0, site 1, the oracle or the bots | 0 | | Median time between two batches | 0.105 s | | Longest time between two batches | 6.83 s | We counted all values from the logs of the node and from the index of the reader. Figure 12. Commands and certified batches each hour, 13 to 20 Nov 2026. Hourly data: CSV file. Commands executed each hour and batches certified each hour over 173 hours. No hour has zero. The shaded part shows the days with site 2 silent. (image: https://www.defisilicon.com/docs/assets/figures/c1-seven-days.svg) ## Who sent the commands | Sender | Share of commands | |---|---| | Three market-maker bots | 36.51% | | Three volume bots, taker side | 32.50% | | Three volume bots, resting side | 18.88% | | Price oracle | 9.64% | | Momentum bot and whale bot | 2.43% | | Other accounts | below 0.01% | Eight bots and the oracle, all run by the operator, gave more than 99.9% of the commands. Outside users were few. This is a limit of the test. The exchange does not stop self-trades. Each trader must not trade with its own quotes. 4,706 fills (0.09%) were a market-maker bot that traded with its own order. ## Site 2 stopped, and the chain continued From 14 Nov, site 2 became slower than the chain. On 15 Nov at 15:54 UTC, it stopped. It then failed at each restart until the end of the run. During approximately five days, sites 0 and 1 made all batches final. The rule "site 0 plus one more site" did its work. But the run also showed a problem. The watchdog raised the alert "site 2 silent" 4,289 times. No notification channel was set, so no person saw it. An alert that reaches a person is step 3 of the roadmap (see 5.1). --- # 3.2 Throughput and latency ## Our rules for numbers - Each number has a date and a source. - We do not mix live numbers and numbers from a model. - We remove artifacts that increase or decrease a number. We tell what we removed. - We give latency as p50 and p99. Half of the orders are faster than p50. One order in 100 is slower than p99. ## Throughput On 1 Nov 2026, a load test ran on an earlier chain without policing: | Item | Value | |---|---| | Rate | 1,035 commands each second, for 120 s | | Commands executed with proof | 145,235 of 145,235 | | LATE orders, full tables, relay overflows | 0 | | Credits rejections, sent again by the clients | 2,105 | We did not do a load test on the policing chain. ## Latency on the last chain (13 to 20 Nov 2026) | Leg | p50 | p99 | |---|---|---| | CANON to executed | 135.51 ms | 256.81 ms | | Executed to batch close | 4.62 ms | 20.69 ms | | Batch close to FINAL | 72.27 ms | 133.31 ms | | Full path (calculated, with the wait for the watermark) | approximately 226 ms | | This latency is too high, and we know the cause. On 3 Nov 2026, before policing, the full path was 22.5 ms at p50. On 6 Nov 2026, the stake rules started. With these rules, the host does not execute a batch before the previous batch has its certificate. Because of this, the batches do not overlap. Each batch waits for the FINAL of the previous batch (72 ms), and then for two evidence steps in the host (21 ms). The data of all 9,053 minutes agrees: | Item, median of all minutes | Value | |---|---| | Time between two batches | 104.2 ms | | Batch close to FINAL | 72.4 ms | | Host evidence steps for each batch | 21.3 ms | | Stake preparation for each batch | 2.6 ms | | CANON to executed (time that an order waits for its batch) | 134.5 ms | Because of this, an order waits **134.5 ms** for its batch. This wait is most of the 226 ms. The problem is in the host, not in the card. **What we will do.** We will change the stake rules in the host. A batch will use the certificate of an older batch, not of the previous batch. Then the batches overlap again, as on 3 Nov 2026. We expect the full path to go back to approximately 22.5 ms, the value before the stake rules. This change also changes the time when a stake penalty starts. It must have a complete design before we build it. We did not start the design. ## Earlier measurements | Date | Setup | Full path, p50 / p99 | |---|---|---| | 28 Oct 2026 | One order in each batch, 3 ms watermark | 10.4 / 14.8 ms | | 3 Nov 2026 | Batches in parallel, 10 ms watermark, no policing | 22.5 / 38.1 ms | | 13 to 20 Nov 2026 | Policing | approximately 226 ms / not measured | ## Less than 1 ms: the design budget The design target for the full path is less than 1 ms from the wire stamp to FINAL. The testnet does not meet it today. The table shows where the time goes now, and the design target for each part. | Part of the path | Last chain, p50 | Design target | What must change | |---|---|---|---| | Wait for the watermark | 8.5 ms minimum | 0.1 to 0.2 ms | Synchronized clocks. Then the wait must cover only the fiber and one peer stamp. | | Card work for the order | not measured separately | 0.08 ms | Nothing. A model of the card logic gives this value. | | Execution on the host | 135.51 ms | 0.02 ms | A batch must not wait for the certificate of the previous batch. The match itself is a few microseconds. | | Batch close | 4.62 ms | 0.01 ms | Close a batch without a fixed wait | | Signatures and checks to FINAL | 72.27 ms | 0.35 ms | FPGA cards at all sites sign in hardware. The card checks a signature in 0.154 ms (measured). | | **Full path** | **approximately 226 ms** | **approximately 0.6 ms** | | The targets are a design budget, not a measurement. The only measurement today is a silicon test of 30 Sep 2026. In it, the card went from an accepted order to FINAL in 0.33 ms, with four peer signature checks. That test did not contain the match or the network. We will publish the full path when we measure it. Figure 13. The last chain and the design target, part by part. Each part of the path on a log scale. A filled square is the last chain at p50, and a hollow square is the design target. (image: https://www.defisilicon.com/docs/assets/figures/c2-latency-legs.svg) ## An artifact that we removed On the policing chains, one metric compared the clock of the card with the clock of the host. It did not correct the offset between them. It showed approximately 21 s that was not real delay. We do not use that metric. All latency on this page comes from inside the systems of the operator. It does not contain the network path of an outside client. --- # 3.3 Silicon ## The card image today | Item | Value | |---|---| | Board | One AMD Alveo U50 at site 0 | | Clock of the exchange logic | 140 MHz (one tick is 7.14 ns) | | Timing margin at 140 MHz | +0.061 ns after place and route | | Exchange logic | 258,975 LUTs | | Full device, with the AMD shell | 387,075 of 870,720 LUTs (44.45%) | | In operation | From the 10 Nov 2026 chain. No card stop in the 7-day run. | Before it went live, the image passed simulation against a software reference, emulation, four audits and a test on the physical card. ## Cost of operations on the card | Operation | Cost | Source | |---|---|---| | Check one Ed25519 signature of a site | 153.59 µs | Measured, 30 Sep 2026 | | Sign one batch digest | 92 µs | Design target | | Work for each order (session, copy check, stamp, sequence) | 1 to 2 µs | Design target | | Order accepted to FINAL, one batch, test | 327.80 µs p50 | Measured, 30 Sep 2026 | The card is not the slow part of the chain today. Section 3.2 shows that the latency comes from outside the card. --- # 4.1 Trade: website and command line The testnet uses test funds only. For access, join the waitlist on www.defisilicon.com. The operator approves each account. ## Keys | Key | Function | |---|---| | Wallet (Phantom or Solflare) | Signs you in to the website. It does not sign orders. | | Trading key | Signs each order. The website gives it on the Account page ("Reveal trading key"). | | RPC key | Lets a program read your data. You make it on www.defisilicon.com/rpc. | ## Website 1. Open www.defisilicon.com/exchange. 2. Connect Phantom or Solflare. On a phone, use the wallet app. 3. Select BTC-PERP. 4. Set the leverage, the price and the size. Then send the order. The website shows when your order is final. ## Command line The official client kit contains the `./trade` command. You receive the kit when the operator approves your account. ```sh ./trade setup # paste the trading key, then the RPC key ./trade markets ./trade balance # proof-verified balance ./trade buy BTC-PERP --price 81500 --size 0.01 --post-only --leverage 5 ./trade sell BTC-PERP --price 81400 --size 0.01 --reduce-only ./trade positions ./trade margin add 250 ./trade orders --market 2 ./trade cancel ORDER_ID --market 2 ./trade history --limit 20 ./trade dashboard ``` ## Rules for the command line - Each command uses one token from the rate bucket of your session. - You can use one trading key on more than one computer. Run `./trade setup` on each computer. - After an interruption, run `./trade resume`. It is safe, because each nonce executes one time only. - Do not delete the files in your data directory. They contain the state of your commands. --- # 4.2 Market makers ## Why this venue - **The wire sets your queue position.** The card stamps your order when it arrives. No hidden priority exists. - **The match is deterministic.** Each site calculates the same fills from the same input. You can replay your fills. - **A quorum certifies each result.** The kit checks the proofs on your computer. ## Features | Feature | Function | |---|---| | Post-only, IOC, reduce-only | Order flags | | Known order ID | You know the ID before the ACK. You can cancel immediately. | | Post-only replace | One command cancels the old quote and sends the new quote. If the new quote crosses the book, the old quote stays. | | Batches of commands | Send many commands without a wait. Confirm many with one proof read. | | Mass cancel | One command cancels all your orders on a market or on one side | | Dead-man switch | If you send no command for a set time (1 s to 24 h), the ledger cancels your orders | | Private stream | Your orders, fills and funding as events | | One key for each machine | Each machine can have its own revocable key and nonce sequence | ## Rate limits (configured) | Session | Sustained rate | Burst | |---|---|---| | Market maker | 50.06 commands each second | 60 | | Normal trader | 16.68 commands each second | 4 to 8 | ## Self-trade The exchange does not stop self-trades. Your system must not trade with its own quotes. We keep the engine simple and the same for all traders. ## How to start Ask the team for access. You receive the client kit, an account, keys and a session for a market maker. --- # 4.3 AI agents An AI agent or a bot can trade for your account with its own key. You can revoke this key at any time. Your other keys continue to work. ## Connect an agent 1. On the computer of the agent, run `exchange-trade connect --name "My agent"`. The kit makes a new agent key. The key stays on that computer. 2. Open the link that the command shows. Use the browser where you signed in. 3. Compare the code on the screen with the code on the computer. Approve only a connection that you started. 4. Push **Approve agent**. Your browser signs the registration of the agent key. 5. On the computer of the agent, run `exchange-trade preflight --json`. The result must show `"ready": true`. ## What an agent can do | An agent can | An agent cannot | |---|---| | Send, replace and cancel orders | Withdraw funds | | Set leverage and margin | Change keys or link wallets | | Use the dead-man switch | Sign in to the website | | Read its account, with proofs | Act on other accounts | | Revoke itself | Add other agents | ## Local MCP server The kit contains `exchange-mcp`, an MCP server for agent clients. Add it to your client: ```json {"mcpServers":{"defisilicon":{"command":"exchange-mcp","args":[]}}} ``` The server gives tools to read markets, balances, positions, orders and history. It also gives tools to send and cancel orders. The keys stay in the local profile. Do not put keys in a prompt or in the MCP configuration. ## Disconnect an agent - On the website, open **Connected agents** and push **Disconnect**. - On the computer of the agent, run `exchange-trade disconnect`. --- # 4.4 API ## Read API The RPC gives verified data. You must have an RPC key. Make it on www.defisilicon.com/rpc. The page also shows the RPC address and examples. | Data | Endpoint | |---|---| | Account balances, orders and positions | `/api/v1/accounts/{account}/balances` | | One of your keys | `/api/v1/accounts/address/{key}` | | Markets, order book, trades | see the RPC page | ## How a command goes to the exchange 1. **Make it.** Each command is 256 bytes. It contains the chain, the account, the nonce, the action and an Ed25519 signature. 2. **Count it.** Each trading key has its own nonce sequence. The ledger accepts a command only with the next nonce. Execution uses the nonce, also when the ledger rejects the action. 3. **Send it.** The command goes as one UDP datagram to each site that you select. The session key of the site authenticates it. 4. **Wait for the result.** The ACK means "accepted". The receipt means "executed". The certificate means "final". | Action | Code | |---|---| | Place | 1 | | Place with flags (reduce-only, IOC, post-only) | 22 | | Cancel | 2 | | Replace | 15 | | Post-only replace | 29 | | Add or remove margin | 23 | | Set leverage | 25 | ## SDK The kit contains a Python SDK and a native client. The native client signs commands and checks proofs. Use the SDK, not raw datagrams. It sends again after rejections, and it keeps the nonce state safe. --- # 4.5 Verify it yourself ## Three layers of evidence 1. **Certificate.** Site 0 and one more site signed the digest of the batch. 2. **Replay.** We publish the canonical order. You can execute it again and compare the state root. 3. **Proof.** A Merkle proof connects one balance, position or receipt to a certified state root. ## How to verify - **Your balance.** `exchange-trade balance` shows a proof-verified balance. - **The full history.** An independent reader downloads all certified batches. It checks each signature and executes each batch again. During the testnet, the operator must approve the key of an outside reader. ## What a verification proves | It proves | It does not prove | |---|---| | A quorum signed the history | That the history is the latest one | | A replay gives the same balances and fills | That site 0 did not reject or delay an order before the stamp | | Your balance is part of the certified state | That the card holds its own key | | The rule combined the stamps correctly | That the clocks of the sites are correct | | The stamps of the sites are authentic | That the three sites are independent | --- # 5.1 Hardware and roadmap The hardware steps are the path to the objective: a distributed sequencer, wire stamps and no front-running. We build the hardware in four steps. Each step adds one capability. We keep the hardware of the step before. | Step | Hardware | Clocks | Status | |---|---|---|---| | M0 | One U50 card (site 0) and two software sites | Each site counts its own time | In operation: runs the testnet | | M0.5 | Three new servers, three U50 cards and one spare, fiber spools between the cards | PTP through the fiber, crystal clocks, 5 to 15 ns (design target) | In procurement. The spools passed the tests of the supplier. | | M1 | M0.5 plus three rubidium clocks, a timing shelf and GNSS | Atomic clocks police each clock | Designed and quoted | | V2 | Operators that we do not control | Atomic-policed | Not started | Figure 14. The hardware steps. Four steps as a stair: M0 in operation, M0.5 in procurement, M1 designed, V2 not started. (image: https://www.defisilicon.com/docs/assets/figures/s6-hardware-steps.svg) ## What each step proves - **M0** proves that one FPGA card can stamp, sequence and sign batches for a live market. The 7-day run is the evidence. It does not prove clock synchronization or independent sites. - **M0.5** will prove clock synchronization in nanoseconds with no master clock, stamps in hardware at three sites, and consensus between three cards. The three spools are approximately 9.9 km, 4.9 km and 1.9 km. The different lengths make real optical delay part of each measurement. M0.5 does not prove clock policing, because a crystal clock cannot check its own rate. - **M1** will prove that atomic clocks find drift and false time, and make it evidence. - **V2** will prove that the network works when we do not control all the sites. ## Roadmap | Order | Step | Done when | |---|---|---| | 1 | New testnet chain | The chain runs, the bots quote, the reader verifies | | 2 | Batches overlap again: remove the 134.5 ms wait for the certificate of the previous batch | We measure the full path at approximately 22.5 ms or less | | 3 | Alerts that reach a person, and a check for slow sites | A test outage of a site makes an alert | | 4 | Outside market makers and AI agents | Outside traders use the kit and agent connect | | 5 | M0.5: three cards on fiber | Each card passes the physical test | | 6 | Fiber calibration and PTP | We measure and publish the clock agreement | | 7 | New ordering rule in the cards | Built, audited and live on a new chain | | 8 | Low latency on the three cards | We measure the full path from the wire stamp to FINAL on the three cards, and we publish it | | 9 | M1: atomic clocks | We test the faults on the physical setup | | 10 | V2: outside operators | Operators that we do not control run sites | Later items: a replacement for site 0 with no stop, keys that stay in the card, a matcher in hardware, more markets. --- # A. Glossary and word list We write these pages in ASD-STE100 Simplified Technical English. STE lets a project add its own technical names and technical verbs. This page gives ours. ## Technical names | Name | Meaning | |---|---| | ACK | The answer of the card when it accepts a datagram | | Agent key | A trading key for one bot or AI agent. You can revoke it alone. | | Barrier, Credits, Rate | The three rejections of the card when it cannot accept an order now (see 2.3.4) | | Batch | A group of orders from the canonical order. The sites execute it and sign it as one unit. | | Bridge | A contract on another chain that accepts a signed state root and moves one stablecoin in and out | | Canonical order (CANON) | The one ordered list of orders that the card releases. Each site executes it. | | Card | The AMD Alveo U50 FPGA board at site 0 | | Certificate | The proof that a batch is final: the digest and its signatures | | Credit | One position for an accepted order on a card. Each card has a fixed number. | | Datagram | One UDP message. One command is one datagram. | | Digest | A hash of the orders of a batch, their order and their result | | Door | A site that accepts orders from clients and sends them to the card | | Fallback stamp | On the testnet: a stamp that a site sends when it has no receipt for an order | | Fill | One trade between two orders | | FINAL | The state of a batch that site 0 and one more site signed | | FPGA | A chip with a logic circuit that you can program. It has no operating system. | | Front-running | To trade before an order that you saw, with knowledge of that order | | Funding | A payment between long and short positions of a perpetual future each hour | | GNSS | Satellite time (GPS and similar). We use it for the date only. | | Hash | A short fixed-length value that identifies data. The same data always gives the same hash. | | Host | The computer that holds the card at site 0 and runs the matcher | | Index | The reference price of BTC that the oracle publishes | | LATE | A rejection: the second stamp of an order arrived after the watermark | | Late guard | On the testnet: a software stamp is at most 1.5 ms earlier than the card stamp | | Ledger | The record of all accounts, balances and positions | | Leverage | Position size divided by margin, from 1x to 20x | | Liquidation | The closure of a position when its margin is too low | | Lot, tick | The size step (0.001 BTC) and the price step (1 USD) | | M0, M0.5, M1, V2 | The four hardware steps (see 5.1) | | Margin | The collateral that a position must have | | Mark price | The price that sets margin checks and liquidation | | Matcher | The software that matches orders and changes the ledger | | MCP | Model Context Protocol: a standard interface between AI agents and tools | | Median | The middle value of a sorted list | | Merkle proof | A short chain of hashes that connects one value to a state root | | Multi-home | To send the same signed order to two or three doors | | Noise bound | The published distance below which two orders are simultaneous | | Nonce | A counter in each command. The ledger uses each nonce one time only. | | Observation | A signed statement of a site: "I saw order H at time T" | | Oracle | The account that publishes the index two times each second | | Ordering rule | The rule in the card that calculates the official time of an order | | Perpetual future | A future with no end date. Funding keeps its price near the index. | | Policing | The system of evidence, proofs and penalties for wrong actions of a site | | PTP | Precision Time Protocol: clock messages with hardware stamps | | Quorum | The identical signed digests that make a batch final: five of six sites in the full design | | Reader | A program that downloads the history, checks it and executes it again | | Receipt | The signed result of one command | | Replica | A copy of the chain that any person can hold, verify and serve | | Resting order | An order that stays on the book | | RPC key | A key that permits read access to your data | | Session | An authenticated channel from a client to the card, with a rate limit | | Shadow mode | A new site stamps, replays and signs, but its signatures do not count | | Site | One validator location, with its own card, matcher and read server | | Snapshot | One signed root of the full state, each 100 ms | | Stablecoin | A token with a value fixed to one currency | | Stake, bond | Value that a validator locks. A proven offence removes part of it. | | State root | One hash of all balances, positions and orders after a batch | | Tap | A passive optical splitter, glass without electronics, that sends the light that arrives to the local card | | Timing shelf | The shelf that connects the atomic clocks to the cards (M1) | | Token bucket | The rate limit of a session: tokens refill at a fixed rate, and each command uses one | | Trading key | The key that signs your orders | | Validator | A site that executes and signs batches | | Watermark | The card clock minus a lag that covers the slowest path between the sites. Orders with an earlier time go into the canonical order. | | White Rabbit | A time-transfer technology from CERN with precision below 1 ns | | Wire stamp | The time when the bytes of an order arrive at the port of the card | ## Technical verbs | Verb | Meaning | |---|---| | Calibrate | Measure the delay of a fiber one time and sign the result | | Clip | Limit a value to a band | | Close | End a batch and send its result to the card | | Execute | Apply commands to the ledger | | Forward | Send a received order to the other sites | | Match | Pair a buy order with a sell order | | Overlap | Run at the same time, in a pipeline | | Police | Measure a clock from outside and inside, and record a difference as evidence | | Quote | Put buy and sell orders on the book | | Relay | Send a received record to other sites | | Replay | Execute recorded history again to check it | | Revoke | Make a key unusable | | Sequence | Put orders in the canonical order | | Sign | Make a cryptographic signature | | Stamp | Record the arrival time of an order | | Synchronize | Make clocks agree with each other | | Verify | Check a cryptographic signature or proof | ## Limit of our STE check A script checks these rules: sentence length, paragraph length, punctuation, verb forms and a list of unapproved words. We did not check each word against the licensed STE dictionary. --- # B. FAQ ## The sequencer **Who sets the order of trades?** No single site. Each card calculates the same sequence from the same stamps with a fixed public rule. See 2.3. **How many times does an order go between sites?** The client sends it to two or three doors. Each door forwards it one time to the other sites. Each site receives a maximum of one copy for each door, plus the copy from the client. See 2.3.1. **What occurs when a site receives a second copy of an order?** The card identifies the copy from its hash. The card stamps only the first copy. See 2.3.1. **Can an order execute two times?** No. The ledger uses each nonce one time only. **How many stamps does an order have?** One from each site: six with six sites. Each card broadcasts its stamp one time. See 2.3.2. **How do six stamps become one time?** Each card converts the stamps to one time base with the clock offsets and the fiber delays. Then the rule finds the median, clips each stamp to a band, and calculates a weighted average. One false stamp can move the result by a maximum of a published bound. See 2.3.2. **What occurs with two orders that arrive at the same time?** Orders closer than the noise bound are a group. Random data that no person knows at the time of the signature sets the order in the group. Speed below the noise bound has no value. See 2.3.2. **Why does the card wait before it releases an order?** The stamps of far sites are on the fiber for tens of microseconds. The card waits until the watermark passes the time of the order. The wait is approximately 0.1 to 0.2 ms (design target). See 2.3.3. **What occurs when the chain receives too many orders?** The card rejects immediately with Rate, Credits or Barrier. The client sends again. The card never discards an accepted order. See 2.3.4. **Can a site trade before me?** It cannot put an order before yours after yours arrives at the doors. It can delay the orders that enter through its own door, and you have other doors. One false site can move a stamp inside the published bound only. See 2.3.5. ## The project **What is the objective?** An exchange with a distributed sequencer, wire stamps and no front-running. See 1.1 and 1.2. **Is the testnet this system today?** No. The testnet has one FPGA card and two software sites. Section 1.2 tells what exists today and what is not built. **Which market can I trade?** BTC-PERP, with leverage up to 20x. The testnet uses test funds. **Who runs the sites?** In production, independent companies, at least three. On the testnet, the DefiSilicon team runs all three sites. **Why an FPGA?** The FPGA stamps the order at the network port, when the light becomes bits. No software touches the order before the stamp. **Why atomic clocks?** They do not set the time. They check it. A site with a false clock makes evidence against itself.