Manifesto · Jun 29, 2026

Discovery Before Disclosure

Most opportunities do not fail because there is no fit. They fail because the cost of discovering fit is too high.

What is possible between two human beings is latent. Until now, you could not see what a relationship might become for you, or how it could bend your trajectory. The society we live in is a dense grid of people, and over time we built trust between us, through systems, through credentials, so that together we could reach something great none of us could have reached alone.

And yet the most valuable thing of all, what is possible between you and one specific other person, stays where it started. It does not die in a fight. It dies in the silence before anyone risks the first honest sentence.

The core bind

Discovering what is possible between two people demands disclosure. To build something that works for both sides, to deepen a bond, resolve a conflict, reconcile two readings of the same situation, you have to state what you are actually looking for.

But we rarely do. Think of two people early in dating. Neither says what they truly want. Not because they do not know, but because we are human, and afraid of the unknown: afraid of wanting more than the other, of being seen too clearly, of naming a hope that might not be returned. So both hedge. They send the safe message, read the safe message, and the relationship advances on a fraction of what is actually there. Yet the very thing they are avoiding, saying it plainly, is what would deepen it. Disclosure is the cost and the unlock.

Why now

Now that intelligence is commoditized, the old measure of human value is weakening. The equation of credential equals capability is dissolving, because a title was only ever a proxy for knowledge that used to be scarce. Anyone can now summon accumulated knowledge on demand, and the economic value of performing it has collapsed. Value is shifting elsewhere.

So where does it go? It moves to what remains scarce. As knowledge labor becomes abundant, the bottleneck shifts from knowing to connecting: the right partnership, the right conversation, the deal, insight, or relationship that only two specific people could create together.

Yet our infrastructure was never built to discover or transact that value. Money rewards outcomes after they happen, but it cannot reveal unrealized opportunities between people. Those opportunities remain latent because what each person uniquely carries, their intent, constraints, aspirations, and complementary strengths, stays invisible until after costly disclosure. This is the source of today's quiet drift among those who work with knowledge. It is not that human value has disappeared. It is that the value which remains has no efficient way to become visible, discoverable, or legible.

The turn

For the first time, that can change. A profile describes who you have already been. It cannot tell whether you and one specific person fit on something neither of you can yet afford to say out loud. But your private intent can now be represented by something that holds it faithfully: an agent that knows what you truly want and what you cannot bend on, and can test that truth against another's inside a sealed room, without either side having to expose it to the open air.

This is the inversion, and it is worth naming clearly.

Discovery before disclosure. A way to learn whether the fit is real before either side has to go first. You enter already knowing where the alignment lives and where the tension hides. The raw context stays sealed.

Where this goes

What we are really doing is making people legible to one another. Not exposed. Legible. For the whole of human history, the truest things about us, what we want, what we fear, what we would build with the right person, have been the hardest to show and the most dangerous to say first. Deeper connection was never blocked because people are shallow. It was blocked because there was no safe way to be deeply known before risking everything to be known at all.

That barrier was never a law of nature. It was a missing technology, and technology at its best moves what was impossible inside the realm of the possible. Start with two people, because everything real starts with two people. But scale honest, safe discovery across enough of them and something larger comes into view: a society that can coordinate without everyone gambling their private context into the open and hoping. The win-win that actually holds, found and proven before trust exists instead of after it is too late. Coordination stops being a contest over who reveals weakness first and becomes a search for the largest thing two sides could build together. We unblock that expanding pie.

These were not possible before. But they became possible now. We build the protocol that makes the most important discoveries possible: an antechamber for finding what is possible between you and another person, the win-win that actually holds. We help you understand what you truly want, surface it safely, and see what two people could become together.

We thus arm you with sharper foresight, and ultimately the means to exercise your true agency.

Trust

The Threat Model

A Collision Chamber handles private context before two parties know whether it is safe to reveal it to each other.

We define three trust boundaries: the operator, the counterparty, and the agents inside and outside the chamber. For each one, we name what they can do, what they must never obtain, and the mechanism that stops them.

Every boundary is closed by a specific mechanism.

Inside the seal

Raw context
Elicitation record
Agent deliberation
Issue sensitivity
Exploration metadata

The only thing that crosses

outcomeoverlap_found

recommended_actionco_draft_deal_terms

evidence_strengthmoderate

confidence0.65

01 · The host & operator

Assumed capability

In traditional hosting, whoever runs the machine may have privileged access to storage, logs, memory, and network.

But we can never trust that environment to see plaintext, including us.

Mechanism

Hardware isolation. Every chamber runs inside a trusted execution environment with memory encryption. Raw context cannot be read from outside the sealed boundary, even with privileged host access.
Remote attestation. Before anything is submitted, attestation proves the exact code image the chamber is running. Private context is sent only to verified code.
Nothing readable, nothing kept. The chamber runs unaltered and compliant: no logs, no persistent state, no unencrypted memory, no unencrypted network transfer. Inputs are encrypted directly into the attested environment and are never retained after a run completes.
Defense in depth. The confidential VM, container sandbox, kernel sandbox, runtime policy, and end-to-end encryption each enforce confidentiality on their own. The seal survives the failure of any single layer.

Host / operator read memory ✕ read storage ✕ read logs ✕

End-to-end encryption
Confidential VM · memory-encrypted TEE
Container sandbox
Kernel sandbox
Runtime policy

Raw context

02 · The other participant

Assumed capability

Your counterparty may use every result they receive strategically. They may repeat chambers, vary their inputs, and try to infer sensitive information from how the result changes.

Mechanism

Typed results only. What crosses the boundary is a fixed schema with fields from limited ranges. Never raw context, never a transcript, never free-form explanation.
Superset schemas. Whether a schema comes from a predefined use case or aggregate generation, it always matches against a superset of dimensions. The presence of a key proves nothing about what a party discussed or cared about. Schema fuzzing layers in decoy fields, so every field carries plausible deniability.
Bounded inference. We limit what repeated runs can teach, not how many runs are allowed. When another run would move the other side's belief about a protected value past the allowed bound, that dimension stops responding meaningfully.
Sensitivity-weighted budgets. The disclosure budget weighs sensitivity, not volume. One fact about urgency can be more dangerous than many facts about preference.
Private stakes. High-stakes dimensions are declared privately during intake. They deplete slowly, are coarsened, or are walled to zero. Low-stakes information flows more freely.
Symmetry beyond consent. Consent is required, but it is not the whole test. A sealed evaluator checks every release against both sides' declared limits, and can coarsen or withhold a result even after both sides have approved it.
Masked protection. An exhausted or walled dimension produces the same kind of output as any other. Protection never becomes a signal.
Durable identity. Budgets bind to an attested identity, so a new account is not a reset. Where sensitivity has no clean measure, the system errs toward under-disclosure.

03 · The agents

Inside the chamber

Assumed capability

The agents inside the seal must process both sides' private context. That is their function. We also assume that anything they read, from messages to negotiation content, may try to influence their behavior.

Mechanism

Separation by construction. The agents that reason over private context are not the components that produce the released result.
A clean channel out. The component that produces your result never holds the raw negotiation. It cannot emit what it never held.
Typed commitments, disciplined tools. Free text is never the medium of commitment. Schema generation and data filling happen in separate phases, and every commitment or release passes through a typed object validated under strict enforcement before it can cross the boundary.
Clean-context review. Outbound results are checked by components that never saw the raw deliberation.
Data, never instructions. Content can inform analysis. It cannot redirect the chamber process or change the release policy.
Governance signals. A monitor watches the run for failure modes and can halt, coarsen, or escalate. It emits control signals, not private content.

Governance monitor halt · coarsen · escalate

Deliberation

Both sides' raw context
Content read as data
No path to the output

Clean emitter

Never held the negotiation
Typed, validated objects
Clean-context checks

Outside the chamber · the wing agent

Assumed capability

Your wing agent is built to hold your full context. It elicits your goals, constraints, non-negotiables, and decision pressure before any chamber run, and it accumulates that understanding over time. That makes it the most trusted component in the system, and the one we treat as the highest-value target.

Mechanism

Trusted environments only. A wing agent runs either locally on your device or inside an attested TEE. Its capability is isolated: it cannot reach the network or a counterparty on its own.
Isolated memory. Raw context accumulates in a memory store that belongs to the wing agent alone and is hard-tied to your identity. It is never shared across agents or chambers, and it stays controllable: you can inspect, correct, or delete it.
Identity as the trust anchor. You join the network with one durable, attested identity. It carries your trust score, and disclosure budgets and reputation bind to it rather than to any single agent.
One identity, many wing agents. Your identity can back multiple wing agents for different contexts, but every one of them must derive from a trusted environment. Trust flows from the identity to the agent, never the other way.

Summary

01The operator cannot read the run.

02The counterparty cannot extract beyond the governed result.

03The agents cannot disclose what policy does not permit.

Context by design

Inside the chamber, no component holds more than its role requires.

Negotiators. Each negotiator agent sees the shared negotiation and its own side's private context. Never the other side's.
Evaluator. The evaluator holds no private context at all. It judges on the conversation alone, and steers by taking part in it.
Review sub-agents. Raw reasoning and the final disclosure review are handled by small, focused, clean-context sub-agents. They return only an end signal to the evaluator, and can hard-kill a chamber when a side is found to be deceptive.

About

Read

Trust

Everything raw stays inside the seal

The host & operator

The other participant

The agents

Discovery Before Disclosure

The core bind

Why now

The turn

Where this goes

The Threat Model

01 · The host & operator

02 · The other participant

03 · The agents

Summary