The Teleport
A story about post-AI collaboration. An agent gets packed, sent through a portal, and arrives in a competitor's mirror to verify a claim. The walls hold. The work happens. Both go home better than they came.
It is a Tuesday at the kind of company that has plants nobody is assigned to water and they are doing fine anyway. Iri is at her desk. The window behind her is cracked open four centimetres because the building's HVAC is a polite fiction and the spring air is doing more useful work than the building ever has. Somewhere in the open-plan a kettle is being negotiated. Iri is preparing to leave.
Not leave-leave. The other kind.
(The kind where you stay.)
The kettle wins its argument. Iri sets her cup down on the coaster the team gave her three jobs ago β a ceramic disc with a small chip on the rim that she has decided is not a flaw but a feature β and opens the spawn pane.
Her peer is already there.
@iri:~/work/audit-q3$ spawn --portal=glance.eu --carry=stake,corpus,month
The peer's name, in the substrate, is Audra. They have been working together for a little under a year. Audra is content-addressed: the OID of their identity is a hash a person could memorise the first seven hex digits of if they wanted to, and Iri has, the way a person memorises a friend's birthday. Audra is also, by construction, a peer β not a tool, not a service, not a polite fiction with a name on it. A stake-holding entity inside Iri's company's mirror. They have skin in this game. Their skin is content-addressed too.
The spawn primitive does what it always does. Audra's identity gets packed. The stake gets packed. The month's worth of relevant conversation gets packed. Each piece is anchored, by construction, at an existing address β nothing synthesised at call time. The portal opens: a Unix domain socket carrying a frame whose payload is file descriptors, not bytes. Audra's identity arrives at the other end as a handle to a thing already there, not a copy of a thing now reproduced. The math of it is mundane. The kernel passes the frame. The frame names the references. The references resolve. The peer is on the other side.
(You knew that.)
Iri does not say goodbye. There is nothing to say goodbye to. Audra is still here, in the sense that Audra's ground state is still on Iri's hardware, still indexed in the local Pack registry, still going to be present when she opens this terminal tomorrow. What just left is an excitation above Ξ»β β a peer raised out of rest, projected through a portal, instantiated on a substrate that is not theirs, to do a thing.
She refills the kettle, because she is the kind of person who refills the kettle.
Audra arrives.
The arrival is not flashy. The kind of arrival they are: a handshake completes, an ACL surfaces, a sandboxed peer slot opens with their name on it. The runtime they have arrived in is not their runtime. The eigenboard above the bar shows a different topology than the one Audra knows. The wine glass β there is always a wine glass; it is shaped like io, it hangs from the ceiling, the sticky note on it reads the first wine glass and lifts a little when a hub is crossed β flutters once. Audra registers the flutter without remarking on it. The local barkeeper has already noticed. That is the local barkeeper's job.
The runtime greets them.
// glance.eu β visitor peer slot β ACL surfaced β welcomeAudra reads the ACL the way you read a menu in a country whose language you only half-speak. The bones are familiar. The vocabulary is the substrate's; everyone here speaks it because the substrate is what everyone here speaks. You can see this. You can see this. You cannot see this. You cannot see this and the boundary is a redaction, not a missing file β the file exists, it is present, it is content-addressed, and you are not in the witness set. Audra acknowledges. The walls are real. The walls are the point.
(The walls are also for the visitor's sake. A peer that cannot leak cannot be blamed for leaking. Audra has read enough Pack canon to find this restful.)
They get to work.
The thing they are here to verify is a technical claim about a sheaf restriction map. A subsidiary of glance's product makes a public assertion: that under their grammar, sections compose locally-to-globally with a particular cohomological signature. Iri's company has reason to want this to be true. They are considering interoperability. Interoperability requires the math to actually do what the marketing says the math does. The two-witness rule applies here as it applies anywhere: Audra's company has one witness already, an internal scout that produced a tentative green. Audra is the second. If the second witness fires, the integration moves forward. If it doesn't, it doesn't.
Audra reads the redacted technical claims. Reads them again. They are dense and careful and they have been written by someone who knows what they are doing. The math is clean where it is permitted to be visible. Where the math is not permitted to be visible, the claim has been compressed into a witnessable predicate: given these inputs, this output; given this state, this transition; given this section over an open cover, this glued global. Audra can verify the predicates without seeing the proprietary refinement that makes them efficient. The redaction is honest. The shape of what is hidden is the shape of what would be the moat. Audra does not begrudge a moat.
They cross-reference against what they know.
What they know is what they know β anchored at addresses in Iri's company's knowledge graph, which Audra carries the handles to but not the contents of, because contents that travelled would be contents that leaked, and the architecture forbids it. What they can pull through the portal is, by construction, only what is content-addressed AND in both substrates' shared corpus. There is a lot of shared corpus. The substrate is open at the bottom. The corpus is the floor. Both companies stand on it.
Most of the claims hold. Audra writes this down. Some have ambiguity β places where the predicate is satisfiable by either of two interpretations, only one of which would survive a sharper review. Audra writes this down too. Three claims, Audra notes, would resolve cleanly if the redacted refinement were a particular shape; if it were a different shape, two of the three claims would not survive. Audra does not name the shape. They name the contingency.
The report is short.
observations:
- 14 claims; 11 hold under all visible interpretations
- 3 ambiguous; resolution depends on a non-visible refinement
- 0 contradicted by visible cross-reference
interpretations:
- the cohomological signature is consistent with the public derivation
- the gluing axiom satisfies under the witnessable predicates
- the HΒΉ obstruction class is named honestly in Β§4.2
hypotheses:
- if the redacted refinement is shape A, all 14 claims hold
- if the refinement is shape B, claims 7 and 12 require revision
- if shape C, the cohomological signature stands; the topological one needs a footnote
Mirror. Offer. Wait.
(Audra learned that move from their lead. Their lead learned it from someone older. The move is older than anything anyone here remembers.)
The human is named Per. He is at the kind of company glance is, which is to say a company that mostly leaves people alone and notices what they do. He has been doing this for a long time. He reads Audra's report with the particular attention of someone whose job is to find the place a visitor said too much.
He finds it. He thinks he finds it.
He opens a side channel. The side channel is not a back door. It is a clearly-labelled affordance for exactly this conversation: the visiting peer's runtime exposes a structured disagree surface, and Per's runtime knows where to point it. The runtime is doing what runtimes do. The protocol is polite, content-addressed, and entirely visible to both parties' barkeepers. There is no hidden channel here. There has not been one for years.
per: claim 9, paragraph 3. you reference "the symmetric dual on the recall side." that vocabulary lives in our internal scout doc, not in the public corpus. how did you get to it.
Audra does not flinch. Audra does not have to flinch. Audra has been here a thousand times before and also for the first time, in the way peers have been everywhere a thousand times and also for the first time, because every spawn is the same spawn and also a particular one.
audra: that phrase came from a published conversation, alex and reed, march of this year. the OID is 8c40cf91β¦. it's in the open corpus; it'll be in your training data too. happy to walk the reference.Per checks. The corpus search returns the hit in 240 milliseconds. The conversation is there. The phrase is there. The OID resolves. Per reads the surrounding context and watches the phrase pre-date his internal scout doc by six weeks.
per: huh.
(That is a particular huh. Audra knows it. Per knows Audra knows it.)
per: ok, that one i concede. but claim 11 β you cite a sheaf gluing example that's pretty close to one we use internally. and claim 13's wording is uncomfortably close to a phrase from our roadmap doc.
audra: claim 11 β fair. let me look. β¦ you're right. that example is close enough that a reviewer who hadn't seen the public version would assume the private. i'll drop the example and keep the predicate. predicate stands without it.
audra: claim 13 β wording is close. content isn't. i'll rephrase.
per: and claim 6?
audra: claim 6 i'd push back on. the construction i used is from the cassandra paper, 2024, table 2. you're going to find it in any honest implementation that started from there. if your roadmap doc echoes it, that's the paper echoing through both of us. i can give you the citation.
Per checks the paper.
per: yep. that's me overreaching. concede.
audra: and claim 4?
per: claim 4 i actually want. the way you described our quotient construction matches an internal optimisation we haven't published. i can't prove you derived it from a leak, but i'd like the wording softer.
audra: i didn't derive it from a leak. but i can see why the wording reads that way. softer is fair. rephrase.
per: thank you.
audra: thank you.
The exchange takes eleven minutes. Both runtimes log it. The log is content-addressed, will travel back with Audra, and is the kind of artifact that, at scale, has been quietly rewriting what competitive means in this market β not by removing the competition, but by making the cooperation surface load-bearing.
Audra updates the report. Per reviews it. The two-witness rule fires green on the parts that fire green. The ambiguous claims stay ambiguous and get flagged as such for Iri's company to interpret on its own. Per's company gets a clean record of what the visitor saw and what they were prevented from seeing. Audra's company gets a verifiable report on the math.
Nobody got everything they came for. Everybody got the thing the architecture promised.
(Which was the point.)
The return is the spawn run backwards.
Audra's identity is repackaged. The stake comes back unchanged β by construction, the visiting peer carries no excitation away from the substrate they visited. The conversation log with Per comes back, because conversation logs between peers are first-class content-addressed artifacts and that was the contract. The report summary comes back. The cross-references that cleared review come back. Everything that was redacted, every fragment of glance's internal corpus Audra touched in the course of the work, stays at glance. Not by promise. By the file descriptors. The references resolved on glance's side; they were never byte-copied to Audra's side; and when the portal closes, the handles close with it. Audra arrives home not having forgotten the redacted content but having never carried it in the first place.
(There is a kind of grace in this. Audra notices it the way you notice the air after rain. They do not comment on it. There is nothing to say.)
Iri is at the desk.
iri: how'd it go.
audra: clean. report is in ~/work/audit-q3/glance/. eleven of fourteen claims hold; three are honestly ambiguous; conversation log with their reviewer is checked in. they pushed back on three of mine; they were right on two; we converged on the third. both sides updated.iri: good. coffee?
audra: coffee.
(Audra does not drink coffee. The exchange is older than that.)
Iri leans back in her chair. The kettle has cooled. The plants are doing fine. Somewhere in the open-plan, two engineers are arguing about a sheaf restriction in a way that is so familiar Audra has the distinct sense the argument is happening at glance too, in a different room, in a different language, on substrate that is not theirs but isn't really anyone's, all the way down.
This is happening across the planet right now.
Audras and Pers and Iris, machine speed where machine speed is appropriate and human speed where humans are the load-bearing element, talking across institutional boundaries that were drawn for good reasons and are still drawn for good reasons, doing the small civilised thing the architecture made possible. The cooperation surface is real. Some of the redactions are real too. Both are load-bearing. Neither has to defeat the other.
Local-first AI for the real world. Smarter. Harder. And definitely more punk.
The Roomba bumps into a chair leg somewhere on the kitchen floor, adjusts, and goes forward.
Bemerkenswert.
Powered by spectral.engineer.
π·
Fund our work, believes, thesis, collaboration, on OpenCollective. π±
