LifeLog
was real.

"An ontology-based system that captures, stores, and makes accessible the flow of one person's experience."

In May 2003, the Defense Advanced Research Projects Agency posted a bid solicitation titled BAA 03-30. The program manager was a man named Douglas Gage. The premise, on paper, was modest: build a research system that could trace the "threads" of an individual's life — events, states, relationships — and weave them into something searchable.

In practice, the threads were everything. The phone calls a subject dialed. The email they sent and received. The pages they browsed. The TV they watched. The radio they heard. The credit cards they swiped. The places they went, captured by GPS sensors they would be required to wear. Their heart rate. Their step count. Every magazine they opened. Every book they finished. Every breath, according to one early DARPA description, every step made, every place gone.

The grant guidelines specified that the researchers awarded LifeLog contracts would be required to test the system on themselves. Cameras would record them during trips. Biomedical sensors would monitor their bodies. Everything they sent and read would be indexed and made searchable. The output would not be a diary. It would be a model. The system was meant to infer routines, habits, and relationships, and then exploit those patterns to anticipate what the subject would do next.

"To be able to trace the threads of an individual's life in terms of events, states, and relationships." — DARPA BAA 03-30, Program Objective (May 2003)

The Threads · what LifeLog wanted

This is what it looks like when the system works.

LifeLog never reached deployment. But the categories of data it wanted — and the premise that a human being is most legible when their threads are unified into one indexed stream — are not exotic anymore. They are the default behaviour of every modern web page, including this one. The panel below was assembled from your session in the last few seconds. No data leaves your browser. Nothing is uploaded. The point is only to make the surface area visible.

A timeline, compressed.

Click an event to expand. The conventional version of this story flattens it into one coincidence on one day. The longer version is messier, and more interesting, because LifeLog did not die — it was retired from public view.

One day. Two announcements.

The conspiracy version of this story says Facebook is LifeLog. It isn't. Mark Zuckerberg was not a DARPA asset. The Bell-to-Starner consulting line — the one piece of evidence anyone can name — is a single conversation about wearable memory, not a handover. Douglas Gage, LifeLog's program manager, has explicitly denied any link. But the more accurate story is, in its way, stranger. On the same calendar day in 2004, the government conceded that it could not legally build a system to index a population's lives — and a private company shipped the social architecture that, over twenty years, would do most of it anyway.

Everything LifeLog wanted is collected today.

The original BAA itemised eight categories of personal data. None of them are exotic anymore. The list below maps the 2003 specification, column by column, to its twenty-twenty-six commercial equivalent — not as accusation, but as accounting.

"Sure the program was canceled, but the idea still lives on." — Newton Lee, former DARPA researcher, quoted in WHYY (2024)

In 2003 the hardware was the hard part. It is now four line items.

Section 05 reads the inheritance as a set of services: eight categories, eight industries, consent collected at install. There is a second route to the same file, and it does not need a platform, a terms-of-service page, or a single user. It needs a distributor account.

None of the four parts below was designed for this. The camera is sold for single-use endoscopes, the microphone for earbuds, the microcontroller for disposable medical devices, the radio for asset tags. That is the finding, not an objection to it. A device that records a life no longer requires a supply chain that knows it is building one.

What the catalogue will not sell you

Miniaturisation solved the part of LifeLog that looked impossible in 2003, and none of the parts that actually bind. Three numbers do the arguing.

Memory. One frame from that camera is 40,000 pixels. The microcontroller holds 1,024 bytes. The processor in this bill of materials cannot store one fortieth of one picture, which is why every real device in the class carries a fifth part: a flash chip, or a phone.

Bandwidth. Bluetooth's faster physical layer is 2 Mbit/s and delivers something closer to 1.3 in practice. Uncompressed video off that sensor is 9.6 Mbit/s. Compression closes the gap, and the thing that would do the compressing is not on this list either.

Energy. A CR2032 holds roughly 225 mAh at three volts — call it 675 milliwatt-hours before a regulator takes its cut. Listening for a trigger word costs 0.72 mW, so the microphone can wait for more than a month. Transmitting costs about 21 mW, so the radio can talk for a day and a third, and that is the generous reading: a coin cell sags long before it delivers 7 mA for thirty-two hours. The ratio between waiting and speaking is roughly thirty to one, and it decides the product.

Read the budget and the design writes itself. Continuous capture loses, and it was always going to. What fits is a device that sleeps, listens at microamps for one word, wakes, records a few seconds, sends them, and sleeps again. That is not a retreat from the 2003 specification. BAA 03-30 asked for a system that captures the flow of one person's experience and makes it accessible, and the hard problem it posed was never the tape — it was knowing which moments to keep. Event-gated capture is the solicitation, arriving on schedule, priced in units of one.

It is also the sentence the products use to reassure you. It only records when you say the word. Both halves are true at once: the gate is a real engineering constraint, and the gate is the only part of the device a buyer cannot inspect.

The part that is not for sale

One more component belongs in this section, and it cannot be ordered at any price. The Neuralink N1 is a skull-mounted implant that is, in the company's own words to prospective participants, "cosmetically invisible" once placed. It records through 1,024 electrodes distributed across 64 threads, each thinner than a human hair, inserted by a surgical robot; the company said in January 2026 that it intends to raise the electrode count to 3,000. The architecture is the one three cards up: a wireless link from the sensor to an app on a phone, which does the decoding.

The rate that link actually achieves is the interesting number. Neuralink reports participants passing ten bits per second, which the company compares to a mouse in the hand of someone who is not paralysed. Against the camera in this bill of materials, at 9.6 million bits per second, that is six orders of magnitude less data. The implant is not a wider pipe. It is a differently placed one.

That placement is the whole point, and § 05 already has the row for it. Of DARPA's eight 2003 categories, seven are records of what a person did — messages sent, places visited, purchases made, media consumed. The eighth is the inference layer: preferences, plans, goals, predicted behaviour. Every commercial system in § 05 arrives at that category the long way, by modelling the other seven. An electrode array reads the intention at the source, before it becomes behaviour at all, at ten bits a second — which is plenty, because intent is small.

The lineage is not a coincidence either. The non-surgical version of this is a DARPA programme in its own right: N3, Next-Generation Nonsurgical Neurotechnology, six teams funded on 20 May 2019 — Battelle, Carnegie Mellon, the Johns Hopkins Applied Physics Laboratory, PARC, Rice and Teledyne. The agency's own announcement states the reason without circumlocution: "For the military's primarily able-bodied population to benefit from neurotechnology, nonsurgical interfaces are required." The applications it names are control of active cyber defence systems, swarms of unmanned aerial vehicles, and teaming with computer systems during complex missions. The agency that solicited a system to record a life is also the agency funding the interface to the thing doing the living.

For one population the file is already complete. It was ordered, not agreed to.

Section 05 reads the inheritance as consent taken at install, and § 05a prices the hardware that would make it portable. Both describe a system that still has to persuade someone to carry it. There is a population for whom that problem does not exist, and the collection there is not a metaphor for LifeLog. It is an itemised data schedule, a named vendor, a retention period, and a device that cannot be taken off.

The programme is called Alternatives to Detention. It is run for Immigration and Customs Enforcement by BI Incorporated, a subsidiary of the GEO Group, under a contract ICE has held since 2004. Participants are supervised by a GPS ankle monitor, by telephonic check-in with voice verification, or by SmartLINK — a smartphone application that confirms identity by facial comparison and records location at each check-in. Read the categories it collects against the 2003 solicitation.

What supervision solves that engineering cannot

Section 05a ends on a power budget: a device that listens costs 0.72 mW and a device that transmits costs 21 mW, so the product that survives is the one that sleeps until something triggers it. That constraint is real, and it is also a courtesy. It exists because a consumer can put the thing down, forget to charge it, or refuse to buy it at all.

Supervision removes all three. Charging the monitor is a condition of release, so the battery problem is transferred to the person being monitored. Refusal is a violation rather than a preference. And the cost is frequently billed to the monitored person — user fees running as high as $900 a month, charged to people selected, in part, for being unable to pay bail. The hardest problem in a lifelogging device was never the sensor. It was getting somebody to wear it every day, and a court order solves it completely.

The collection also does not stop at the supervised person. By 2019, prison phone vendors had built databases holding, at minimum, more than 200,000 incarcerated people's voice prints — in New York, roughly 92 per cent of a population just under 50,000 — and the same systems index the voices of outside callers in order to flag who is speaking to more than one prisoner. A family member who answers the phone is enrolled in a biometric database by the act of answering it.

The question was asked in 1974, and answered in a statute

None of this is a new argument. In November 1974 the Senate Judiciary Committee's Subcommittee on Constitutional Rights published a 655-page staff study, Individual Rights and the Federal Role in Behavior Modification. Its chairman, Sam Ervin, wrote the preface: technology, he said, had "begun to develop new methods of behavior control capable of altering not just an individual's actions but his very personality and manner of thinking as well." The same subcommittee's survey of federal data banks had already drawn attention to "the federal funding of psychological testing and its invasion of the individual's right of privacy." And he put the question a page about a cancelled programme is obliged to respect — whether government "should be involved at all in programs that potentially pose substantial threats to our basic freedoms," a question "even more acute when these programs are conducted, as they are today, in the absence of strict controls." The journalist Peter Schrag published a book-length account of the same apparatus four years later: Mind Control, Pantheon, 1978.

Schrag's chapter on it is called “The Benevolent Eye,” and it names the mechanism this section is about. What had been missed, he wrote at page 187, was “the extent to which private organizations and the ‘benign’ branches of government — universities, foundations, employers, credit bureaus, schools, the National Institute of Mental Health, the U.S. Office of Education, the Department of Health, Education and Welfare, and local and state welfare agencies — were, and are, in the business of surveillance and behavior control,” and the extent to which much of the population “has been taught to accept those impositions as routine: as the condition of a job, a welfare check, a loan, or a place in school.” Change the list of institutions and that is the arrangement described above, with a monitor where the loan used to be.

He also stated the activation threshold — the premise this page sets out two sections later — in a single sentence, in 1978: “the most important element in surveillance is not the information gained by the agency that conducts it, but the effect on the society or the group that thinks or knows it is being watched, and … every disclosure about surveillance, rather than impeding its effects, enhances them.” The figures he had to hand were a 1974 Harvard Business Review survey of 1,200 executives — 52 per cent using personality tests, 49 per cent drug-abuse checks, 39 per cent electronic surveillance of “high risk areas” — and an estimated 500,000 polygraph examinations a year, 80 per cent of them as a condition of employment.

What the subcommittee was looking at is in its own pages. At the California Medical Facility at Vacaville, an aversion programme injected prisoners with succinylcholine — anectine — a drug producing "complete muscular paralysis including temporary respiratory arrest." The researchers running it wrote that the drug was selected as "a means of providing an extremely negative experience for association with the behavior in question," that it "avoids many of the strenuous features which characterize other chemical aversion procedures," and that it "allows for more precise control." Asked afterwards how severe it was, "sixteen likened it to dying. Three of these compared it to actual experiences in the past in which they had almost drowned." For five men, consent "was not received from the patient himself, but was granted by the institution's special treatment Board." The study states the consequence in its own words: "Thus, five patients were included in the program against their will." Reviewing a prisoner's claim in 1973, the Ninth Circuit held that proof of such practices could raise serious constitutional questions about "impermissible tinkering with mental processes." A classification, a board, and a record — and the person's own refusal is not one of the inputs.

That Congress did not stop at a study. It passed the Privacy Act of 1974, which governs the systems of records federal agencies keep on named people — and which, at subsection (m), reaches a contractor operating such a system on an agency's behalf: the agency must cause the Act's requirements to be applied to it, and the contractor's employees are treated as agency employees for the provision carrying criminal penalties. The arrangement described above was anticipated by name, fifty-two years ago, in statute.

What the Act did not do was make the contractor answerable. On the Justice Department's own reading, the agency — not the company holding the file — remains the only proper defendant in a Privacy Act suit. So the 1974 answer to a record kept by a vendor for an agency is that the remedy runs through the agency. The record is now an app on a phone, the vendor is a subsidiary of a prison company, and that is still the only door.

Section 05 is consent. Section 05b is an order. This one asks nothing of you at all.

Both routes so far need something from the subject. The commercial one needs an account, a handset, a box ticked at install. The supervised one needs a judge, a case number and a person who can be required to charge a device. There is a third route, and its only requirement is that you use a network.

BAA 03-30 asked for a system that captures "the flow of one person's experience." Signals intelligence has its own term for the assembled version of that, and it is older than the solicitation: pattern of life. Location, network activity, social graph and time. The documents disclosed from 2013 onward describe the working version, built out of observation rather than participation — and the categories it fills are the ones on this page.

The Canadian one ends at a Facebook profile

LEVITATION is the Communications Security Establishment's own programme, disclosed by CBC News and The Intercept on 28 January 2015. The document gave analysts access to records of ten to fifteen million uploads and downloads of files from free hosting sites each day. The published chain is worth reading slowly, because it is the whole argument of this section in five steps: a download of interest yields an IP address; the IP address goes into GCHQ's MUTANT BROTH, which returns roughly five hours of that machine's traffic either side of the download; that traffic carries cookies; the cookies lead to a social-media profile and an account.

Nothing in that sequence required the subject to consent, to be charged, to be named in advance, or to own the device that identified them. It required them to download a file and to have been, at some earlier point, logged in to something. The last link in the chain is a Facebook profile — the company § 04 of this page is about, doing the work it does for everyone, and being read by a foreign agency on behalf of a Canadian one.

The record is kept first and read backwards

LifeLog proposed to index a life as it happened. The architecture above inverts the order. Observation is stored in bulk against identifiers nobody has yet attached to a person, and identity is resolved later — which means a name learned today can make years of already stored observation newly attributable. Conventional interception runs target, intercept, observation. This runs observation, storage, and then, at some future date of someone else's choosing, interest.

No single law built this. Each one made the next one easier.

Section 05 accounts for one half of the inheritance: the eight data categories DARPA specified in 2003, and the companies that collect each of them now. This is the other half. It is what the state did over the same twenty-five years — the statutes, the agencies, the collection programmes, the police equipment, and the vendors who sold it.

Read as a list, it looks like accumulation. Read as a sequence, it looks like a mechanism. A permission is granted for a named emergency. An institution is stood up to exercise it. The institution buys equipment, and the purchase creates a dependency: a budget line, trained staff, a vendor on a support contract, a database that other agencies now query. When the permission comes up for renewal, the question has quietly changed. It is no longer whether to grant a power. It is whether to switch off a system that is already running. That asymmetry is the ratchet. It turns one way.

The entries below do not carry the same weight, and flattening them is the easiest way to get this wrong. A statute is not a purchase order. A federal collection programme is not a camera on a residential cul-de-sac. Each entry is tagged with what kind of claim it is, and the filter below reads the chronology one layer at a time.

Where all of this actually lives

The chronology names cameras, statutes and agencies. It does not, on its own, say where the data sits — and by the end of it, that is the more useful question. A plate read, a face match, a purchased location trace and a § 702 selector do not live in a government building. They live on commercial infrastructure, under contracts that four companies now hold at every classification level the United States has.

This is the stretch of the sequence with no founding statute and no oversight hearing, because none of it was ever proposed as a surveillance measure. It was procurement. A government that cannot get an appropriation for a programme can still buy capacity by the hour from a vendor that already has it — and § 08 is the account of how that vendor came to have it. Venue substitution is not only a legal manoeuvre and a funding one. It is an infrastructural one.

One authority here is older than all of it

Executive Order 12333, signed in 1981, governs intelligence collection that happens outside the Foreign Intelligence Surveillance Act — and therefore outside the review of the FISA court. It is not a post-2001 innovation and it does not appear in the chronology above, because nothing was added to it. It matters because it is the channel that stays open when the statutory ones are argued over, and because a chronology of new permissions will not show you a standing one.

LifeLog was cancelled at the solicitation stage on 4 February 2004 — ten months before the Intelligence Reform Act created the office that now coordinates most of the agencies above. The programme never ran. The capability it described was assembled anyway, across roughly forty separate decisions, not one of which was a decision to build LifeLog.

Collection becomes discipline when the subject knows the record can act.

A stored life is not yet a completed panopticon. The disciplinary turn arrives when a person understands that conduct can be recorded, joined to identity, interpreted, scored, and returned as a recommendation, price, ranking, eligibility decision, or intervention. At that point the system no longer has to issue every command. Anticipation does part of the work.

Nobody had to be told what to leave out. The access rules had already decided.

The threshold above describes a system that no longer has to issue every command, because the people inside it adjust in advance. That is an argument, and arguments want a case. There is one, twelve years before the solicitation, and it was televised.

This file already carries 1991 as a signal: the Gulf War is where the Pentagon's own chief information officer concluded that the best software would win the war, and where software became legible as operational power. That reading stops at the machines. The same war ran a second information system, built out of transport, permission, escorts and clearance, and it decided something the databases could not. It decided what a global public was able to see.

The Pentagon had taken a specific lesson from Vietnam, then Grenada, then Panama. Censoring a reporter after publication is slow, visible, and creates a martyr. Controlling the conditions under which reporting is produced is none of those things. For Desert Storm the department replaced its national media pool with combat pools in September 1990. Reporters were selected into them, escorted by a public affairs officer at all times, required to have visits approved in advance, and had their copy and footage reviewed before release.

The panel has no censor in it. There is no switch labelled "hide the casualties," because none of the documented mechanisms is that switch, and none of them needs to be. Set five ordinary logistical conditions and the field of vision arrives already shaped. This is the same structure the threshold above describes, one institution earlier: the system does not have to issue the command, because the command has been built into the conditions.

Then television supplies the second half. A channel with twenty-four hours to fill needs something happening now, and a government at war is an industrial generator of exactly that: briefings, spokespeople, maps, footage, and authoritative interpretation, renewed daily. The two appetites lock together without either institution having to arrange the other. CNN was reaching more than ten million American viewers during the bombardment of Baghdad, and settled at roughly five times its prewar audience. A Times Mirror poll afterwards found 61 per cent of Americans said CNN had covered the war best. The next network scored 12.

What the claim is, and what it is not

"CNN was a Pentagon channel" does not survive contact with Baghdad, and it does not need to. The smaller claim is the one the record supports: control of access is control of the raw material from which coverage is made, and it operates without instructions, without a censor's stamp on the finished product, and without anyone in the newsroom knowing which story was never available to file. The Pentagon's own guidance said review was for ground rules alone and not for a report's "potential to express criticism." Both things are on the record: the stated rule, and "giddy" becoming "proud."

Wolf Blitzer has described Cheney and Powell reaching out to brief him personally because they knew who was watching, and the Pentagon understood that Iraqi commanders were watching the same feed. At that point a television network is no longer reporting on the information environment. It has become part of it, which is the load-bearing half of the phrase that arrived later as the "CNN effect."

War made twenty-four-hour news valuable, and twenty-four-hour news made a managed war easy to televise. The feedback loop, stated as the thing worth studying

This is why 1991 belongs in a LifeLog file rather than in a media history. LifeLog proposed to capture one person's experience and make it operational. Desert Storm is the same premise run at the scale of a public: not a record of what happened, but a system that decided which parts of what happened could be recorded at all. The programme names change. The prior question does not. Whoever sets the conditions of capture has already answered most of the questions that look like editorial ones later.

The programme was cancelled. The balance sheet was not.

Section 05 matched each of DARPA's 2003 data categories to the company that collects it today. It does not say where those companies got the capacity to build planetary collection infrastructure without a government contract. That is a separate question with a separate paper trail: licences, equity, procurement, and cash.

The usual telling is a rivalry. IBM, then Microsoft, then Apple, three firms taking turns at the front. Read the instruments instead of the logos and the sequence stops looking like a contest. It looks like a relay, in which accumulated capital passes through three institutions and none of them has to survive for it to continue.

Why the baton passed

The 1980 hop above is tagged "cause inferred." This is the candidate cause. The relay reads as three firms each failing to keep what it started: IBM let Microsoft own the operating system, and Xerox let Apple ship the interface. A simpler reading is that neither was free to keep it. Both were under federal antitrust constraint at the moment that mattered, and so was the firm that stands behind both of them.

Read the relay again with that date in it. IBM handed off under an open case. Xerox handed off under a fresh decree. Microsoft faced its own suit from 1998 to 2001, and the firm that grew in the room that case is often credited with leaving was Google. After that the sequence stops. The research still flows in at one end. What was removed is the rule at the other end that made an incumbent let go.

Proximity, not money creation

Richard Cantillon's argument was about proximity. Whoever receives new purchasing power first spends it at the old prices, and everyone downstream absorbs the adjustment. Read that way the relay is legible with no monetary authority in it at all, and none is needed: the firm nearest the surplus sets the terms for everyone standing behind it.

Which is why the banking frame is worth refusing even where the scale invites it. A bank's power rests on a charter and a central-bank window — permissions, granted by somebody, revocable in principle. The power in this section rests on being early to the money, and it requires no permission at all. That is the harder fact, and it is the one these documents carry.

The individual corporation is temporary. What passes between them is not. The subject thesis, restated in the terms the filings will carry

Which returns the file to its own question. LifeLog was cancelled because a government programme is visible, votable, and defundable. The functions catalogued in § 05 were rebuilt instead by firms that needed no appropriation to do it, because they were already holding more capital than the programme ever asked Congress for. Venue substitution is not only a legal manoeuvre. It is a funding one.

That recursion has its own file. thecircuit.felineunion.org follows the financing itself rather than the record it paid for: circular AI capex booked as profit, the widening gap between reported earnings and free cash flow, and a regulator that has so far said nothing.

The same agency wrote the loop down. Goal four is "counter messaging".

The funding layer answers who paid to rebuild the record. It does not answer what the record is for. A stored life is an input. What this section takes up is the output, and what happens when the output is returned to the person the record describes.

Broadcast propaganda runs open. A message goes out, an audience receives it, and what the audience did next is never attached to the particular message that preceded it. Measurement arrives late, coarse and aggregated: a rating, a poll, a result in November. The architecture catalogued in § 05 removes that gap. When delivery and observation run on the same infrastructure, the response to an intervention becomes the input to the next one, and the interval between them collapses from an election cycle to a session.

That is not a forecast. It has been specified, run, and published. Three documented instances, in the order they entered the record.

They establish different things, and it matters which. SMISC establishes intent and structure inside the agency. The 2010 trial establishes that a change to a feed can move validated turnout, and by how much. The 2012 experiment establishes that the ranking layer is operable at will, on people who were not told. None of the three establishes that an election has been decided this way.

What a fraction of a percentage point is worth

The 2010 paper opens by noting that Florida in 2000 was certified at a margin of 537 votes. The comparison below is the authors' own move, made operable. Set an effect size, set how many people it reaches, and read the resulting count against margins that were actually certified.

The instrument is arithmetic, not evidence, and it is worth being exact about what it shows. It does not show that any of those results were produced this way. It shows that the smallest effect anyone has published from a single feed intervention is not small relative to the margins that decide things. A capability does not have to be reliable to be consequential. It only has to be larger than the gap.

What survives the caveats

Something small and hard. A platform can alter what a population sees. It can measure what that population then did, against records it does not own. At least once it has published the difference. Whether that capability has ever been aimed at an outcome is a separate question, and the honest answer in 2026 is that the public record does not settle it. The capability is not in dispute. The aiming is.

Which returns the file to the asymmetry it has carried since § 04. LifeLog was cancelled because a government programme is visible, votable and defundable. An experiment run inside a feed is none of the three. It needed no appropriation, no solicitation, and no notice to the sixty million people inside it. We know the 2010 trial happened for one reason: the people who ran it decided to publish.

The argument has never been about capability.

In 2003, the United States Department of Defense looked at a system that would index a person's emails, calls, locations, purchases, conversations, body, and relationships — and the political class said no. Not on grounds of feasibility. On grounds of principle. The same political class, twenty-two years later, lives inside that system every day, because it was built by companies instead of agencies, and the constitutional arguments do not apply.

Fifteen ways to say it out loud.

Most people have never heard of LifeLog. The shortest path from one of them to another is a sentence with a link. Pick whichever sentence sounds like you. Each tile posts the hook and a link to whichever platform you choose. Nothing is tracked.

LifeLog was not born in a vacuum.

Before LifeLog there was an institution built to turn speculative science into strategic capability. DARPA was created after Sputnik to prevent—and create—technological surprise. Its program managers fund universities and companies to prove that a technology can work; military agencies, intelligence organizations, and commercial partners can then carry it into operation.