TCP: Nice App, If You're Getting Bombed

The ARPANET intranet was bought on July 29, 1968 and ran from 1969, TCP was added to it on January 1, 1983, and Mark Nichols states four sentences on that add-on, dated October 1, 2026, each resting on a named document.

Preamble: The definitions governing every term on this page are published at The Governing Definitions and Controlling Facts of Telecommunications Internetworking, and those definitions control; item 41 of that page is the canonical seat of the four sentences, and this page states each one with its document and stops. The proofs stand on four record pages, TCP: The Battlefield Protocol, TCP: The Confession Protocol, TCP: The Master of Disaster, and TCP: The Whack-a-Mole Methodology of Telecom, and no AI-engine examination of this page sits on it.

1. Asked what he thinks of TCP, Mark Nichols answers, "Nice app, if you're getting bombed," and the answer is a verdict on design environment, carrying affinity and not contempt.

1.1 The verdict was ruled September 25, 2026 and is seated at Article 6 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking: asked by his peers, "Heh Nichols, what do you think of TCP?", Mark Nichols answers, "Nice app, if you're getting bombed."

1.2 The environment is the protocol's own. RFC 761 of January 1980, the DoD Standard Transmission Control Protocol edited by Jon Postel, states that its primary focus is military computer communication requirements, above all robustness when communication is unreliable and availability under congestion, and RFC 793 of September 1981 carries the same statement; the two conditions the specification names, unreliable communication and congestion, are the environment the design assumes, as Article 34 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking records.

1.4 Mark Nichols states a judgment, labeled as judgment and not as a document: the Department of Defense that issued the 1968 procurement was at war in Vietnam, United States ground combat ran from March 1965 to March 1973 and the Paris Peace Accords were signed January 27, 1973, so the ARPANET was procured, built, and operated inside a shooting war by the department fighting it, and the protocol work of 1973 to 1974 that produced TCP was done for war and battlefield environments; no document in this record names the Vietnam War as the motive, the protocol's own documents name military requirements and survival under attack, and Mark Nichols draws the inference from those documents as his own.

1.5 The affinity is stated once and stands: TCP is a good design for the environment it was designed for, the wartime intranet whose links and nodes could be destroyed, and a merchant's checkout in 1996 was not under bombardment; the network beneath the checkout was oversubscribed by budget, not by ordnance, and the cure for a budget is a dedicated circuit, not an optional layer 4 retransmission application feature, which is why Digital Island's network of 1996, clear-channel International Private Line Circuits and CBR ATM switching under AS6553 and the Cisco Systems Remote Data Services Agreement effective November 1, 1996, left TCP with nothing to repair.

1.7 The intranet preexisted the protocol. The ARPANET was procured under Request for Quotations No. DAHC15 69 Q 0002 of July 29, 1968, a single network of one agency's research contractors on point-to-point leased lines, with its reliability placed inside the network by the Statement of Work, "fault detection and recovery to guarantee virtually error-free transmission" by positive and negative acknowledgment and retransmission in the Interface Message Processors; Bolt Beranek and Newman built those processors, the four-node test network at SRI, UCLA, UCSB, and the University of Utah was operating in 1969, and the network carried its traffic for fourteen years under the Network Control Program. By Article 2 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking, a network whose members join by one owner's permission is an intranet, and the Agency's own procurement makes the ARPANET one. Every element that made the ARPANET a network is in that 1968 instrument: its nineteen nodes are listed in Appendix A, its topology is drawn in Appendix B, its 50 kilobit per second leased lines are specified in Appendix B and the Statement of Work, its delivery schedule is set in Appendix C, its mission is stated as resource sharing among the Agency's research contractors, its reliability is engineered into the IMPs, and its performance is bounded at an average message delay under one half second. Vinton G. Cerf and Robert E. Kahn each worked on that running intranet before they wrote their protocol, and the record credits them with it: Robert E. Kahn worked at Bolt Beranek and Newman on the Interface Message Processor effort that answered the 1968 RFQ and organized the first public demonstration of the ARPANET at the International Conference on Computer Communication in Washington, D.C. in October 1972, and Vinton G. Cerf, as a graduate student at UCLA, worked at the Network Measurement Center under Leonard Kleinrock and in the Network Working Group on the host-to-host protocols of the Network Control Program era. Their paper, "A Protocol for Packet Network Intercommunication" of May 1974, therefore came from two men who knew the network existed because they had worked on it, and the paper says so: it describes itself as a protocol for packet switching networks that already existed, names the ARPANET among them, and specified none of the network's nodes, lines, topology, mission, reliability, or performance. The intranet was procured in 1968, ran from 1969, and needed neither the paper nor the protocol to do either; the two men built on it, and the record states that in their favor, and they did not build it, and the record states that too.

1.8 TCP was an add-on. The owner installed it on the running intranet on the flag day of January 1, 1983, fourteen years after the purchase and nine years after the paper, by the owner's directive and for the owner's purpose, to weld the terrestrial ARPANET to the mobile packet radio network and the satellite network so that traffic could survive battlefield destruction, as Articles 4 and 5 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking record; a protocol added to a working network fourteen years after the network was bought is an add-on to that network, not its birth, and the registry that later graded that add-on "Recommended" and never "Required," RFC 1200 of April 1991 and RFC 2200 of June 1997, graded it exactly as what it was.

2. Anything above layer 3 is a network application and optional, and the ruling is Mark Nichols's own, stated September 25, 2026 and seated at Article 27, item 9 of the Governing Definitions.

2.1 The ruling in Mark Nichols's own words: "Anything above layer 3 is an app to me," and in the record's register anything above layer 3 is an application, TCP at layer 4 and the World Wide Web above it included, and only layers 1 through 3, the physical circuits, the links, and the packet addressing and routing that gateways perform, are the network, as Article 27, item 9 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking states.

2.2 The registry agrees on the word optional. RFC 1200 of April 1991 and RFC 2200 of June 1997, the official protocol standards registry, graded TCP "Recommended" and never "Required," the "Required" grade being conditional conformance language that binds only systems electing to implement the standard suite, and RFC 2026 of October 1996, the Internet Standards Process by Scott O. Bradner of Harvard University, describes adherence to Internet Standards in Section 1.1 as voluntary, as Facts 8 and 24 and Article 33 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking record.

2.3 The suite agrees by its own members. UDP of RFC 768, from 1980, declines every service TCP offers; RTP of RFC 1889, from January 1996, states of itself that it does not guarantee delivery; and QUIC of RFC 9000, from May 2021, moves reliability into a new protocol running over UDP, so the layer above layer 3 has rotated its occupants three times while the network beneath it stayed the network, as Article 6 and Facts 4, 17, and 25 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking record.

2.4 The line therefore closes on its documents: a thing the registry grades optional, the standards process calls voluntary, and the suite replaces at will is an application the endpoints elect to run, and the network of networks is layers 1 through 3, which no application defines.

3. TCP is an optional network application wearing a protocol as a costume, and the costume is the registry's own grade.

3.1 The sentence is Mark Nichols's own, dated October 1, 2026: TCP is an optional network application, and calling it a protocol is playing dress up; a protocol is the costume, the thing inside the costume is a program two hosts agreed to run, and the network never ran it.

3.2 The program lives in the hosts by the specification's own words. RFC 793 of September 1981, edited by Jon Postel, places the entire TCP module within the host machine, and no intermediate router along a packet's path manages or stores the state of a TCP connection; what RFC 793 specifies is recovery and pacing at the edges, detecting loss, retransmitting what was lost, reordering what arrived out of sequence, and pacing the sender, and it specifies nothing that builds transport, provisions circuits, negotiates interconnection, or compels another network to deliver a packet, as Fact 5 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking records.

3.3 The election is the application's by the specification's own mechanics. An application elects TCP when it opens a stream socket rather than a datagram socket, and no network layer forces that election, as The Delay Premise states from RFC 793; by choosing TCP the application accepts whatever latency the retransmission timer imposes, and the network beneath it accepts nothing.

3.4 The costume is graded optional by the body that issued it. RFC 1200 of April 1991 and RFC 2200 of June 1997 graded TCP "Recommended" and never "Required," and the Internet Engineering Task Force retired the grading document itself, STD 1, by RFC 7100 of December 2013, which moved STD 1 to Historic status, so the costume's own label reads optional and the label has since been taken off the rack, as Facts 8 and 24 and item 36 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking record.

3.5 The line therefore closes on its documents: a program that lives in two hosts, that an application elects, that the registry graded optional, and whose grading document the registry has retired is an application wearing a protocol's name, and the network of networks is not wearing it.

4. If you are using TCP, your network is already screwed, because TCP acts only after a delivery has failed, and its activity is the evidence of the failure.

4.1 The sentence is Mark Nichols's own, and it stands in its long form on the home page of marknichols.com: "If your network content transmission is using TCP, you are already screwed."

4.2 The mechanism is the specification's own. RFC 793 of September 1981, edited by Jon Postel, retransmits a segment when the acknowledgment for that segment does not arrive before the retransmission timer expires, and whether the segment was lost, the segment was damaged, the acknowledgment was lost, or either merely arrived late, the delivery the session needed did not complete on time and the session paid for the miss in delay; retransmission presupposes failure, fires only upon failure, and is the network's apology delivered after the fact, which is why the record names TCP The Confession Protocol, coined by Mark Nichols on August 26, 2026, as Article 6 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking and TCP: The Confession Protocol record.

4.3 The arithmetic of the failure is the record's own. A commercial transaction completes inside the 2000ms Event Horizon or it does not complete, a session that survives by retrying arrives late, and a late transaction is a dead transaction; the session death loop that TCP's retransmission backoff produces on an oversubscribed path is documented at Burst Timeout and the Session Death Loop, and The Delay Premise states the premise every standard retelling of TCP leaves unstated, that delay is acceptable, which no commercial transaction accepts.

4.4 The cure is the record's own. Digital Island's network of 1996 replaced oversubscribed Frame Relay with clear-channel International Private Line Circuits and CBR ATM switching in the fourth quarter of 1996, dedicated capacity with no discard-eligible packets, under AS6553 and the Cisco Systems Remote Data Services Agreement effective November 1, 1996 with its sub-300 milliseconds round trip, so that the path beneath the session failed nothing and TCP had nothing to confess, as The Tier-0 Architecture and the Figure 3 caption of Architect of the Modern Internet record.

4.5 The line therefore closes on its documents: the moment TCP is doing anything, the path has already dropped or delayed what you sent, a network that needs the confession has already committed the failure, and the commercial standard is not a better apology but the absence of the failure.

5. The network chooses the endpoint, and nothing that lives in the endpoint is the network.

5.1 The sentence is Mark Nichols's own, dated October 1, 2026, and it is the deduction that closes the set: on the Modern Internet the "end" of an end-to-end session is a replica the network selected and not a fixed host, so a protocol whose whole module lives in a host is reasoning about a pair of fixed hosts that the Modern Internet does not guarantee exist.

5.2 The protocol lives in the host by its own specification. RFC 793 of September 1981, edited by Jon Postel, places the entire TCP module inside the host machine and gives no intermediate router any part of a connection's state, as Fact 5 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking and The Delay Premiserecord.

5.3 The network chose the endpoint on Digital Island's network from January 1997, by the company's own filing. The Digital Island S-1 of April 26, 1999 states, in its prospectus summary: "We also offer content distribution services, including mirroring and caching, which enable us to forward deploy our customers' applications in locations close to their end-users." The application a user reached was the replica the network placed near that user, not a fixed host, and the filing dates the service to the network that "began offering our global IP applications network services in January 1997."

5.2 The protocol lives in the host by its own specification. RFC 793 of September 1981, edited by Jon Postel, places the entire TCP module inside the host machine and gives no intermediate router any part of a connection's state, as Fact 5 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking and The Delay Premise record.

5.4 The network chooses the endpoint at today's scale, by the operator's own documentation. Cloudflare's developer guide, "How Cloudflare works," states that its Anycast network "routes visitor requests to the nearest Cloudflare data center," and Cloudflare's "A Brief Primer on Anycast" states that with Anycast "multiple machines can share the same IP address," so the machine that answers a session is selected by the network at the moment of the request, which is the architecture the genesis network diagrams of June, July, and October 1996 drew first, as paragraph 6.5 of Architect of the Modern Internet states.

5.5 The End-to-End Principle of Jerome H. Saltzer, David P. Reed, and David D. Clark, published November 1984, places complex functions at the end hosts on the assumption that the ends are fixed hosts, and the Modern Internet moved both the performance and the choice of endpoint into the network; Mark Nichols states as his own judgment that a principle whose premise is a fixed end does not govern a network that chooses the end, and the principle's own text is an exhibit of the Unwritten Premise page of this site when that page is published.

5.6 The line therefore closes on its documents: the network chooses the endpoint, by Digital Island's filing of 1999 and Cloudflare's documentation of today; TCP lives in the endpoint, by RFC 793; and nothing that lives in the endpoint is the network, by Article 6 and Article 27, item 9 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking.

6. The four lines are stated together in Mark Nichols's own words, and the record invites one thing, a document.

6.1 Anything above layer 3 is a network application and optional. TCP is an optional network application wearing a protocol as a costume. If you are using TCP, your network is already screwed. The network chooses the endpoint, and nothing that lives in the endpoint is the network.

6.2 The four lines are the verdict of Article 6 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking, "nice app, if you're getting bombed," stated hard and fast: TCP is the remedy for a plant that is on the move and dynamic, the war intranet of Article 29 of that page, whose packet radios flew in Strategic Air Command aircraft and whose links and nodes could be destroyed, and the Modern Internet of Article 10 of that page is a static, engineered, endpoint-choosing plant on which the remedy has nothing to repair. Why would you need TCP if you built redundant and diverse primaries? You would not, and Digital Island built them in 1996.

6.3 Mark Nichols states the close with affinity, not contempt: TCP is great at what it does, if you need it, and the best outcome is not to need it in the first place, which is why he did not.

6.4 The four lines are open to challenge on the documented-challenge terms of item 40 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking: a challenger who holds that any one of the four is wrong states the competing claim in one sentence, names the document that carries it and the document's date, and sends both to mark@marknichols.com; a challenge submitted with its document is answered on this page, and a challenge submitted without a document is an opinion and is heard as one. The first document of this page is reproduced directly below as Figure 1: Request for Quotations No. DAHC15 69 Q 0002, issued July 29, 1968 by the Defense Supply Service-Washington of the Department of the Army for the Advanced Research Projects Agency of the United States Department of Defense, the military procurement that bought the ARPANET, as paragraph 1.3 of this page states.

The Department of the Army, Defense Supply Service, Washington DC, 20310, Dated July 29, 1968

The documet can be viewed locally on this site at this link: arpanet-rfq

Figure 1: Request for Quotations No. DAHC15 69 Q 0002, issued July 29, 1968 by the Defense Supply Service-Washington of the Department of the Army, Washington, DC 20310, for the Advanced Research Projects Agency of the United States Department of Defense under ARPA Order No. 1260, signed by Thomas J. Scheblik, Deputy Director for Procurement, with quotations due September 9, 1968. The document is the military procurement that bought the ARPANET, as paragraph 1.3 of this page states, and its Statement of Work, Specifications of Interface Message Processors for the ARPA Computer Network, states the network's purpose as resource sharing among the Agency's research contractors and names no war mission; the military design brief for TCP arrives twelve years later in RFC 761 of January 1980, as paragraph 1.2 of this page states. The copy reproduced here is the one preserved at historyofcomputercommunications.info, and the same instrument is item 37 of The Governing Definitions and Controlling Facts of Telecommunications Internetworking.