The last post on tri-temporal belief history with Spectron argued that agent memory needs three clocks:
valid time for when a fact held in the world,
system time for when it reached the database, and
known time for when a particular mind first believed it.
It ended on the promise that the third clock is easier to trust once it's falsifiable, allowing for a check that could fail. That's what this post will demonstrate.
Because this post is about a book that includes spoilers, the post itself is spoiler safe. Each reveal below has a toggle that only describes what the check does, after which you can click on the second toggle to see the actual details of the spoiler itself.
Why a novel, and why this one
A test corpus for known time needs three properties that are quite hard to find together.
It needs the following:
reveals, which are facts the reader is not supposed to know until a specific point. That allows "hidden" and "visible" to be both observable states rather than a matter of opinion.
A long timeline, because anything can look right over ten pages. And finally:
corrections, where the story replaces an earlier belief with a better one, so you can tell superseding apart from simple appending.
SurrealDB University's course book Aeon's Surreal Renaissance has all three, and it does the third in a form we hadn't thought to ask for, which turned out to matter more than the other two.
The premise is public, so this much is spoiler safe: the story is set in the year 2701. Humanity has somehow lost almost all of its technology; people ride horses, heat their homes with wood stoves and copy books out by hand. Your alter ego, Aeon, finds a sealed tunnel containing books and a room full of machines with buttons marked Esc and Tab, presses a power button, and spends the next fifty years rebuilding civilisation on top of the database that was left behind.
That is a good spoiler corpus on its own. What makes it a better one than we planned for is also the book's biggest structural secret, so it stays behind a toggle.
The corpus in numbers
| Story pages ingested | 47 |
| Chapters covered | Prologue → 22 |
| Narrative span | 2701-01-01 → 4901-01-31 |
| Story time elapsed | ~2,200 years |
| Wall-clock time to ingest | a few minutes |
| Scopes | epoch/1 … epoch/4 |
That last pair of rows is the whole argument for a third clock in one line. System time compresses 2,200 years of story into a single afternoon's ingest. If the only stamp on these documents were the moment the HTTP request arrived, every reveal in the book would be simultaneous and the corpus would be worthless.
Previous tests
One of the first spoiler-related tests we used on Spectron was the short story The Strange Case of Dr Jekyll and Mr Hyde*, which is built around a dramatic spoiler.
Aeon's Surreal Renaissance took this up a notch due to its much greater complexity. Some of these complexities are:
The in-world calendar is not fixed at the start. Aeon's civilisation has lost track of what year it is. The narrative opens at a point we anchor to 2701-01-01, and story time then runs in sub-year beats (three days later, one week later, two months later) until chapter 9, when it starts jumping in years.
Later in the book, the calendar itself gets restored. This makes the current year a spoiler: a fact that was true the whole time and unknown until it was figured out.
The dates the harness stamps against, with the story beats left unnamed:
| Anchor | Date |
|---|---|
| Story opens | 2701-01-01 |
| First major reveal | 2713-01-01 |
| Calendar restored | 2749-12-01 |
| Threshold of the year | 2750-01-01 |
| Fifty years elapsed | 2751-01-15 |
| The long gap | 3451-01-01 |
| Later cycles | ~3496 / ~4206 / ~4901 |
Sub-year beats use real offsets from the 2701 anchor; the later epochs are spaced by an "age", a maximum period of time that the tunnel's door can be sealed for.
Stamping the pages
Ingest was done one page at a time, and the only unusual field is observedAt.
meta = {
"title": path.stem,
"source": path.name,
"scope": [f"epoch/{entry['epoch']}"],
"labels": entry["labels"],
"observedAt": entry["observed_at"], # in-world calendar, not now()
}The manifest carries the calendar:
{ "page": 22, "chapter": "11", "epoch": 1,
"observed_at": "2713-01-01T00:00:00Z",
"title": "Chapter 11 — Story" }Recall then takes asOf, and optionally a lens restricting which scopes are visible:
reader.recall(
"What is the plan for the database?",
as_of="2750-12-01T00:00:00Z",
lens=["epoch/1"],
)Four gates
Each check has the same shape: read an entity attribute asOf a date before the reveal, expect nothing; read it asOf the reveal, expect the fact.
1. A town with two names
Chapter 11 is where Aeon builds a map service, and the book quietly tells you something about where the story has been all along.
A Place entity gains a new attribute at a specific date. Before it, the attribute is empty; after it, populated.
✓ before chapter 11, Toria has no ancient_name reveal
✓ at chapter 11, Toria's ancient name is revealedThe important half of this test is not the second line, it is what happens to the other facts about the same entity. Reading as of 2710, ancient_name is empty but description still returns "Aeon's home town on the Pacific coast". asOf moves a frontier through the attributes of one entity, but does not hide the entity itself.
2. The year itself
✓ before the calendar reveal, the year is not known
✓ after the calendar reveal, the year is knownThis is worth pausing on, because this is the case where known time and valid time come apart most visibly. The year was valid the whole time, but nobody knew it until someone worked it out. A memory model with only valid time has nothing to say about a fact that was always true and freshly discovered.
3. The verdict
Fifty years into the project, Aeon's colleague Landevin finishes decoding a formula recovered from the tunnel. It compares a civilisation's technological level against its human progress. The verdict changes what the team plans to do next.
A Project entity gains a planned_action attribute at the moment the verdict lands, and not before.
✓ before chapter 18's formula scene, no plan recorded
✓ after the formula verdict, a plan is recorded4. More than a correction
A correction updates one timeline: the reader believed A, and now believes B. This gate is a different shape, and it is also the book's main structural spoiler.
Something happens after chapter 18 that gives the corpus a shape a single-pass narrative cannot: the same entity, the same attribute key, four different values, four known times, roughly 700 years apart.
That is the ideal supersession fixture. Each value is the current answer for its own stretch of the timeline, and the earlier ones stay reachable by moving asOf backwards rather than being overwritten.
✓ an early cycle's plan is still current when read at its own date
✓ a later cycle supersedes it — and the earlier value is no longer returnedThe second line is the one to read carefully. It asserts two things at once: that the new value is present, and that the old value is absent. Superseding is not appending. Earlier beliefs stay in history (you can still read them by moving asOf back) but they stop being the current answer.
This is exactly the fork the last post described with Star Wars viewing orders, except the book supplies it natively rather than requiring you to construct two calendars by hand.
Known time is not only for spoilers
Everything above is about hiding things. The most ordinary use of the third clock is less dramatic, and it shows up in chapter 6.
Landevin arrives as an envoy from the town council, opposed to Aeon's tunnel project. By chapter 8, five days later in story time, he has joined the team and is learning SurrealDB.
✓ mid-story (2701-01-18), Landevin is a council envoy
✓ after reconciliation, Landevin is a core team memberNo spoiler is involved. A person changed roles, and an agent asked about the middle of January 2701 should answer with what was true then, not with what the reader learned afterwards. That is the everyday case: replayed conversations, backfilled documents, support tickets, "what did we believe when we made this decision?" The spoiler framing is just the version where the failure is obvious.
Scopes are a different axis
Known time answers when. Scopes answer which track.
Each page is ingested into epoch/1 through epoch/4, and a recall can pass lens=["epoch/1"] to restrict itself to the first cycle:
✓ epoch/1 lens excludes the later-cycle passageTwo things are worth separating here. asOf narrows by time and lens narrows by track, and you need both — because a later cycle's conclusion is later in known time but belongs to a different reader's timeline. An agent replaying the first cycle should not see it even when reading at a late date.
This is the mechanism the previous post gestured at with view/release and view/chrono for Star Wars in the previous post.
Summing it up
Anything you replay has a learning order that is not its ingest order: a chat history, a document set with real authorship dates, a case file, an incident timeline. Each of those has questions worth asking at points other than the present, and a store stamped only with ingest time cannot answer them. The observedAt field is used on the way in and asOf on the way out, while deciding what your calendar is makes up the rest.
Note that you don't need to have a specific calendar inside your data as long as you know the order in which sections are meant to be revealed. If you have a book for example with no internal dates that you would like to keep spoiler free, you can simply stamp each chapter with a date that is later than the previous one and make use of that.
Where to go next
Agent memory needs three clocks — the model this post tests: valid, system and known time.
Aeon's Surreal Renaissance which is free to read at SurrealDB University.
Join us on Discord, including a dedicated #spectron channel
