I was talking about rsync. Flags. Throughput. The boring, useful mechanics of moving a lot of data without losing your mind. Then the word Samba flashed past in the tool trail — incidental, almost decorative, the kind of edge a model drops when it is explaining a lineage.

My brain did the rest.

Samba → Tridgell → Tridge → computer chess → Joel. A half-remembered meeting in Canberra. A mate who lived in Chatswood while I was over toward Salamander Bay. A private Java engine. A world championship where you had to be an author of a chess program, and there were almost no people on earth doing it, and one of them lived down the road. It was so cool back in the day.

I said the names out loud. Not carefully. Not as a research brief. Cool and slightly freaky in equal measure — and that dual reaction is the correct one. It is not a bug in the framing.

Seconds later the machine handed back a graph: people resolved, engines attached, obscure artefacts surfaced, software history folded into social history. The ingredients of that graph had been public for decades. The graph itself had not.

The ingredients existed. The compiled past did not.

Two kinds you already have

If you have been reading along, you already own two temporal moves with AI.

Future access — Cognitive Time Travel — is the deliverable that would have existed after weeks of work existing now, because calendar time was substituted with compute time. Compress, parallelise, prefetch, simulate: temporal access, not mere acceleration.

Temporal presence — Time-Shifted Proxy — is interaction with a person or state displaced by time: grandparents in the 1970s, Future You, a stand-in for someone the calendar will not let you meet.

What happened with Scott / Joel / Tridge / Chompster / KnightCap / rsync / Samba is neither of those. It is not reaching a future work state. It is not conversing with a proxy. It is something else.

A world-state that never existed as a record

There is probably no single file in any archive that says: Scott Farrell, Joel Veness and Andrew Tridgell occupied a tiny Australian computer-chess scene; Scott and Joel were physically surprisingly close; Tridge later sits on a software lineage that will cause Scott, twenty-five years later, to remember the relationship while optimising an rsync job.

That artefact did not exist. Chess result pages existed. Newsletters existed. Program names existed. Papers existed. Samba history existed. Geography existed in people’s heads. Autobiographical memory existed. The compiled past — the inspectable position — did not.

I still think “time travel” is the right umbrella. I would not replace it with a colder machine name. I would expand it. Future access reaches forward. Temporal presence interacts across a displacement. Past-state compilation reconstructs a configuration of the world that no single record held — and, as we will see, does it inside the half-life of the thought that summoned it.

Taxonomy preview

Myth vs reality

What this book will do

Chapter 2 defines intelligence as resolved entities and edges — not a pile of PDFs. Chapter 3 shows that the join is multi-axis and mostly unspecified by the query. Chapter 4 explains why conversational time is not a UX boast. Chapter 5 prices the collapse: intelligence-shaped compute on trivia. Chapter 6 walks the Samba → Tridge → Joel reconstruction with receipts. Chapter 7 holds the privacy dual face of the same artefact. Chapter 8 treats compiled pasts as inference products and asks how to live forward.

The opening story is proof of existence. The rest of the book is mechanism, economics, and the shadow that comes with the capability.

Imagine the exercise in 1998.

An analyst receives one half-sentence: Scott Farrell says he once met someone called Tridgell and Joel Veness through computer chess. Then the labour begins. Find the organisations. Guess the competition. Hunt participant lists. Obtain newsletters and archives. Search the name. Disambiguate from every other Scott Farrell. Find the engine. Find Joel. Find the engines again. Find Andrew Tridgell. Find KnightCap. Align dates. Decide whether the social connection is even plausible. Write the note.

Depending on archive accessibility, that is hours or days of real work. I am not claiming any intelligence service needed a fixed number of hours on this particular trio. We do not have that evidence, and invented agency numbers are theatre. The labour sketch is enough: the expensive operation was never possession of a PDF. It was joining.

Grains of sand

When those traces were created twenty or twenty-five years ago, none of them was particularly revealing alone:

• one chess results page
• one newsletter line
• one program name
• one author’s page
• one technical paper
• one mention of KnightCap

Each was a grain of sand. Nobody necessarily built the person-level graph with engines, software lineage, geography and social proximity hanging off it. But the edges were latent in the public record.

Intelligence is not the documents. It is the resolved entities and the edges between them.

The dry term and the larger move

The intelligence term for one piece of this is entity resolution — also called record linkage: the task of deciding when records in different data sources refer to the same real-world person, organisation or thing, often without a shared stable identifier.¹

Historians and epidemiologists have known for a long time that linkage across old corpora is hard. Spelling varies. Names change. Administrative boundaries move. Longitudinal study often depends on joining sources that predate national ID numbers.¹

Temporal entity resolution adds another mess: identity and affiliations change across decades. Modern research treats linking entities across time states as a distinct problem, not a trivial SQL join on a primary key.²

Useful vocabulary. Incomplete story. Entity resolution is one instrument in the stack. Past-state compilation is what happens when cheap resolution is combined with temporal reasoning, relationship inference and — as Chapter 3 will show — multi-axis traversal the query never specified.

The mechanism stack

Call the effect retroactive observability — or retroactive legibility, if you prefer. The same traces become more informative later because the cost of joining them falls. The raw evidence might have existed for twenty-five years. The intelligence did not.

Resemblance is not identity

A related vocabulary point, worth one paragraph so we do not confuse it later: finding pages that resemble computer chess is not the same as resolving Joel Veness the person. Embeddings are good at resemblance. Past-state compilation needs identity — natural keys, unique artefacts, engine names, newsletter lines, the awkward human confirmation that two records are one life. Soft corpora are full of almost-matches. Almost is how you compile the wrong past.

Myth vs reality

Chapter 1 showed the flash. This chapter named the substance of what was compiled. Next we need the shape of the join itself — because it is not a one-dimensional name-and-date lookup.

It is tempting to file the whole event under temporal entity resolution: names plus dates, sorted. That framing is not wrong so much as incomplete. It collapses a multi-dimensional walk into a single neat craft label.

What actually happened crossed several axes at once — and the fascinating part is that the query did not specify those dimensions.

The axes

Nobody asked for a seven-dimensional spatiotemporal graph join over computer-chess history and open-source software lineage. Someone was talking about transfer flags. Samba appeared as an incidental edge in an explanation. A human recognition chain fired. Fragments were spoken. The machine resolved, searched, cross-referenced, aligned and reconstructed — then handed the compiled graph back into the same thought.

Heterogeneous matters. Those axes do not live in one table. They live in newsletters, code histories, papers, geography memory, social anecdote and the present moment of work. There was no prebuilt query that said: join rsync to computer chess to old acquaintances by natural key. The system inferred which axes might matter. That inference is power. It is also the error surface.

Present trigger

This is why the capability feels different from “I searched the web for old friends.” The present is not a timestamp on a query form. The present is part of the join key.

Resemblance fails; identity is constrained

One axis failure mode is easy to picture: common name, no domain constraint, wrong person reconstructed with full fluency. One axis success mode is also easy: engine name plus newsletter line plus geography collapses ambiguity. Chapter 6 walks the successful case with receipts. Hold the failure mode in mind — Chapter 8 will make it load-bearing.

Finding “chess-like” pages is resemblance. Resolving Joel Veness is identity. Past-state compilation needs the second, and it often bootstraps the second from sparse unique artefacts rather than from a master person ID.

What this chapter owns

Chapter 2 defined intelligence as entities and edges. This chapter owns the shape of the join: multi-axis, heterogeneous, largely unspecified by the query, often ignited by a present trigger. Next we need the timing property that makes the reconstruction cognitively different from a weekend report.

Suppose the reply had been: interesting — I will have an analyst investigate Scott, Joel and Tridge and send you a report Friday.

Same intelligence. Same sources. Same final facts. Completely different cognitive effect.

By Friday the rsync thought is dead. Samba is no longer in working memory. Your brain has moved on. You open a dossier and think: huh, cool. The graph arrives as news about a past you are no longer standing inside.

What actually happened was different. The reconstruction returned while the original mental state was still alive. New cues landed on an open lattice. More private edges unlocked — lived-down-the-road, we-all-met, tiny-global-scene — and the graph recompiled. You were still in the thought.

The AI got under the latency threshold of the thought.

Conversational time is not a UX boast

Call it conversational time, or in-loop cognition, if you want a sharper phrase: the result returned before the human cognitive state that caused the query had decayed. Therefore the answer becomes an input to the same thought cycle.

That is not merely “real-time.” Real-time is an engineering property. This is a cognitive property. Near-instant matters because autobiographical memory and working context have half-lives. Miss the window and you still get intelligence. Catch the window and you get a different kind of machine: one that can participate inside the activation of a memory, not only report on it later.

The time machine was the loop

Neither brain had the shape. The shape emerged between us. The human side carries private, high-semantic edges that search cannot reliably invent: I knew that bloke; Joel lived down the road; we all met. The machine side carries breadth and cheap recombination: engines, papers, newsletters, software history. Each graph completes holes in the other.

That is why reducing the event to “the model found old documents” is a category error. The model found edges. The human supplied edges. The loop compiled a world-state neither side held alone — and did it before the thought that made the join meaningful had decayed.

Scope fence

This chapter is not a product manual on latency architecture, nor a memory-science literature review. Conversational time is a mechanism clause of past-state compilation. Cue design as a primary topic belongs elsewhere. Here we only need the threshold: past the half-life of the thought, compilation becomes a report; inside it, compilation becomes cognition.

Myth vs reality

Next: why this could fire at all on trivia. The answer is economic, not mystical.

The CIA comparison is sticky and usually handled badly.

What we are not saying: that a consumer chat is better than a nation-state SIGINT shop; that anyone has evidence for a fixed analyst-hour cost of reconstructing this particular trio; that classified capability is knowable from a blog post. Those claims are unknowable, unhelpful, or both.

What we are saying is economic. Historically, nobody would allocate analyst-hours to Scott Farrell having a fleeting rsync memory. The intelligence operation would not be commissioned. The connection would remain dormant. The question was not unanswerable in principle. It was economically nonexistent.

Intelligence-analysis-shaped compute can be spent on trivia at the speed of conversation.

Three variables, one hidden

People tend to narrate the AI change as two variables: models are more prevalent, and models are smarter. Both true. Incomplete.

There is a third variable hiding in plain sight: cognition applied per moment of human life. How much thinking can be pointed at an ephemeral thought while the thought is still warm? That is not the same question as benchmark scores. It is a question about thresholds.

It is not only smarter models. It is how much cognition you can throw at a fleeting mention. I will not invent a GPU count for a conversation — that number is not exposed and fake precision is worse than silence. The shape is enough: enough machine cognition can now be spent that a casual aside becomes explorable mid-sentence.

Cheap cognition makes questions exist

This sits next to a one-line adjacency we have written before: cheap cognition makes thinking rational that previously was not rational to attempt. That afternoon was a personal Version-3 cognition event. No sane human would have commissioned the research. Therefore the connection would have stayed latent. AI did not merely answer faster. It changed which questions get to exist.

Would you have hired a contractor for a weekend to map three computer-chess acquaintances because Samba scrolled past? No. Would you accept a five-second reconstruction mid-chat? Yes. That acceptance is the economic event.

Open-source scale is already an intelligence problem

The U.S. Intelligence Community’s own open-source strategy language treats OSINT as vital to the intelligence mission and frames the challenge as extracting actionable insight from vast and expanding open-source data — coordinating acquisition and developing capabilities and tradecraft for that scale.³

That is institutional recognition of scale and exploitation difficulty. It is not a ranking of consumer chat against agencies. The useful transfer is the economic analogy: effort that used to require institutional allocation can now be pointed at an ephemeral human thought during the thought.

Join cost was also a privacy wall

Chapter 2 defined intelligence as edges. This chapter prices the join. Hold a preview for Chapter 7: the labour required to connect facts was not only an intelligence bottleneck. It was a privacy technology. When that labour cost collapses, recovery of your forgotten world and exposure of everyone else’s are not two different product stories. They are one collapse with two faces.

Myth vs reality

Part I named the doctrine. Part II applies it end-to-end with receipts.

Start from an independent public artefact, not from the flash.

A contemporary Australian Chess Federation newsletter, covering CCT 7, names two Aussie participants and is explicit about one of them: Chompster, programmed by Scott Farrell. That line is not a memoir. It is a grain of sand in print — program name tied to person name in a competition context.

From there the graph walks outward into thin air that was always public and never assembled.

The entity map

Scott Farrell ── Chompster
       │
       │  tiny Australian computer-chess world
       │
Joel Veness ─── Bodo / Meep ── later public RL / AIXI trajectory
       │
Andrew Tridgell ─ KnightCap ── TD learning
       │
       ├── Samba
       └── rsync

Geography hangs off the people: Tridge associated with Canberra; Scott and Joel in a Sydney-side proximity that made the scene feel absurdly small. Social constraint hangs off the competition rule: author-only entry, almost nobody on earth doing it, one of them down the road. Software lineage hangs off Tridge: Samba and rsync as the present-day fuse. Intellectual trajectory hangs off all three: search, evaluation learning, engines, and — decades later — AI conversations that make the old scene legible again.

Evidence trail

One hard number is enough and it is public: KnightCap, in the authors’ reported main success, improved from a 1650 rating to a 2150 rating in 308 games over three days of play. That is not folklore. It is a paper result that turns “Tridge had a chess engine” into a concrete technical node on the graph.

I will not invent Elo figures for Chompster or Meep, nor promote a first-person “fifteen people in the world” memory into a research statistic. Anecdote stays anecdote. Documentary edges stay documentary. That separation is part of the craft.

Human half, machine half

The machine showed a technical edge: rsync related to Samba. The human supplied the social fuse: Samba to Tridge; Tridge and Joel and Scott in one tiny scene; geography that made it feel personal. The machine searched and attached engines, papers, newsletters. The human recognised and added private edges. Neither brain had the full shape.

Neither brain had the shape. The shape emerged between us.

Apply Chapter 3 without re-teaching it: identity, time, domain, software lineage, geography, social proximity, intellectual trajectory, present trigger. Apply Chapter 4 without re-deriving it: the reconstruction re-entered a live thought. Apply Chapter 2: the intelligence was the resolved graph, not any single page. Apply Chapter 5: nobody would have commissioned this as a weekend project; conversational economics made the question exist.

What compilation produced

Not a PDF titled “Scott’s computer-chess social graph, 2001.” A compiled world-state: a position in a historical world suddenly inspectable from 2026. Ingredients old. Artefact new. That is past-state compilation with receipts.

You can frame the capability as recovery: AI can recover my forgotten world. Correct. Incomplete.

The inverse is the load-bearing half of this book:

AI can recover everybody’s forgotten world.

Old mailing-list posts. Conference attendance. Chess results. Company newsletters. Ancient websites. Git commits. Usenet. PDFs. Staff directories. Old resumes. Archived databases. Random names in meeting minutes. Each grain of sand was never particularly revealing alone. The protection for most people was not secrecy. It was obscurity.

Not secrecy — obscurity

Secrecy hides content. Obscurity leaves content findable in principle while making assembly expensive. Historically, that labour cost acted as a privacy technology. The fact existed, but the work required to connect it to ten other facts protected you.

That labour is disappearing. Chapter 2 defined intelligence as edges. Chapter 5 priced the join. Here is the privacy face of the same pricing: when join cost collapses, obscurity dies as a security boundary.

The formula

Writing conditions are not reading conditions. People left traces under an implicit assumption that assembly was costly. That assumption is aging out. “The facts were always public” no longer protects anyone the way it used to. Public is not the same as legible. Legibility is a function of join cost. Join cost fell.

Privacy discourse that only models secrets will keep missing the boundary that is actually dissolving. Secrets still matter. They were never the whole wall.

Hold the contradiction

Personal recovery up; collective privacy down. You cannot honestly cheer only one side. Capability and privacy inversion are not sequential chapters of two different books. They are faces of one compiled-past product. Cool and freaky was the right dual reaction in Chapter 1. It is still the right dual reaction here.

What we are not saying

• Not that consumer chatbots equal nation-state SIGINT.
• Not that open research or journalism is illegitimate.
• Not that you can un-publish the past by wishing.
• Yes that privacy models limited to secrets will lag the real change.

Same doctrine, organisational room

Apply the doctrine without inventing a new framework. Company history is full of grains of sand: contractors, vendor threads, incident footnotes, conference talks, staff pages, old tickets. Mid-incident chat that surfaces a 2019 contractor edge nobody remembered is useful. It is also invasive. Hiring and diligence already feel versions of this; conversational multi-axis join compresses the labour that used to be the practical limit on how much past you could assemble about a person or a firm.

Transparency culture often treated publicity as binary. Practice was graded by search cost. Economics changed. Etiquette and policy lag.

Picture a fluent reconstruction of the wrong person.

Common name. Incomplete records. Aggressive inference. A confident graph that attaches the wrong engine, the wrong city, the wrong decade of work. The danger is not that past-state compilation is useless. The danger is that it is useful enough to be believed when it is false.

False identity matches are a huge danger, especially with common names, incomplete records and AI inference. Provenance becomes essential.

Compiled pasts are inference products

They are not recovered film of the past. Treat them the way you would treat any other high-leverage inference system:

• Show exhibits, not only verdicts.
• Keep edges inspectable.
• Separate documentary edges from autobiographical edges.
• Carry uncertainty visibly.

Adjacent posture, one cameo only: a life wiki worth trusting behaves as navigator, not oracle — cue and exhibit, not a pronouncement about a life. That is not a build tutorial. It is the epistemic stance required once join cost collapses.

Taxonomy — portable form

TIME TRAVEL

FUTURE ACCESS
Access work states before calendar time reaches them.
→ Cognitive Time Travel

TEMPORAL PRESENCE
Interact with a person/state displaced by time.
→ Time-Shifted Proxy

PAST-STATE COMPILATION
Reconstruct a historical world-state from distributed traces.
→ this book

Chapter 1 previewed the table. This is the shareable artefact. Three directions. Different failure modes. Different products. One umbrella if you still like the word time travel — and I do.

Variant: organisational past-state compilation

Same doctrine, different room. Incident review. “Who touched this system in 2019?” Competitive history reconstructed from talks, tickets and contractor exhaust. Multi-axis join still applies. Conversational time still changes the cognitive effect. False matches still scale with fluency. Any product that sells “memory” without entity hygiene and provenance is selling a confident wrong person at institutional scale.

Variant: deliberate personal compilation

The Samba flash was an accident with a loop. The deliberate version is a personal archive made addressable — frameworks, projects, email, recoverable life — so cues can find you and recognition can fire on purpose. Recognition, not free recall, is the twin craft on the human side. Keep that as a cameo. The architecture manuals live in other books. This book only needs the point: you can aim the third kind of time travel at yourself, and you should still hold the privacy dual face while you do.

Living forward

Write traces knowing they will be joined under future intelligence economics. Individuals: assume exhaust is joinable. Teams: treat entity-resolution errors as product risks, not edge cases. Culture: stop using “it was public” as a complete privacy theory.

Close

I call a lot of these things time travel. I have seen the family expand: future access, temporal presence, past-state compilation. The third kind is the one that compiles a world-state that never existed as a single record — in conversational time, across axes the query never named, because join cost collapsed.

The same collapse opens your history and closes everyone else’s obscurity. Cool and freaky. Both. Hold both faces. Demand exhibits. Expand the umbrella. Do not pretend the privacy half is a footnote to the demo.