Join the waitlist
Menu

KAIROS DYNAMICS

Make the system prove it.

Kairos Dynamics builds autonomous research infrastructure for consequential predictive systems: software that generates its own candidates, attacks them, keeps what fails, and controls what is allowed to advance toward a real decision. The search behind it is running right now, and its record of failures, limits, and evidence states is open to inspect, not summarized after the fact.

Possibilities enter.Almost none survive.The failures are kept.

Financial markets are the first proving ground, because a weak model fails there quickly and measurably.

Inspect the live record →What the machine does
QUALIFICATION FIELD / WHAT YOU ARE LOOKING AT Each trace is one possible predictive relationship.

Read left to right: candidate models or decision rules face six evidence tests before forward observation and a separate authority boundary.

Research layer
Implemented Heatmap Strategy Lab ancestry
Advancement
Kairos Core / specified
Playback
Deterministic schematic
SEARCH / GENERATION 00 56 proposed · 0 rejected · 0 observing · 0 advanced
Resolve the instrument
candidate trace ruled-out memory underpowered / retest forward observation controlled advancement

A conceptual preview, not live results. The observatory explains every gate, outcome, and evidence boundary.

WHAT THE MACHINE DOES

One loop, run until something earns the right to advance.

AI is making model generation cheap. What has not become cheap is knowing which models deserve to be trusted. That gap is the whole of the work below, and it runs as one closed loop rather than as a sequence of meetings.

  1. 01 Observe

    Read the data the system is authorized to see, carrying its provenance and the time it was genuinely available rather than the time it was filed.

  2. 02 Define the decision

    Write down the decision being made, what it costs to be wrong, and what evidence would be enough. Written down first, because a bar that moves is not a bar.

  3. 03 Generate candidates

    Produce many possible explanations. Generation is cheap now, and the system treats it as cheap.

  4. 04 Attack them

    Falsification, adversarial tests, counterfactuals, baselines, realistic costs, regime changes, and a holdout nobody has opened. Most candidates stop here.

  5. 05 Keep the failures

    A rejected candidate is recorded with the reason and the scope of the rejection, not discarded. A search that remembers only its successes rediscovers its own dead ends and overstates its own hit rate.

  6. 06 Observe forward

    Survivors are frozen and watched in real time, point in time. Historical success is not allowed to stand in for forward behaviour, and drift is monitored rather than assumed away.

  7. 07 Control advancement

    Evidence maturity and permission to act are separate decisions with separate owners. Lineage and negative knowledge are preserved through both.

The rule the whole thing turns on: the part that generates candidates is never the part that certifies them. A search allowed to grade its own work will always find that it has done well.

WHAT HAS BEEN BUILT

One philosophy. Several systems built on it.

Two of them ask that question of different subjects, a prediction and an agent, and the third is the domain where the loop has to survive contact with reality first. None is the company, and the list is meant to grow without redefining it.

These three are the ones far enough along to show. Work on human context, on controlled operations, and on generation that can explain evidence without altering it comes out of the same discipline. What is built, and what is not →

  1. Invariant

    Predictions Running now

    Does this prediction deserve to be acted on?

    A candidate is held against a standard fixed before the test, then across sub-periods, regimes, cost models, scrambled controls, and a holdout nobody has opened. What survives every one of those is frame-independent. What does not is recorded and kept.

    8,855 typed rejection records against 374 archive occupants. 59.9% of those occupants are under a contamination notice

    Inspect the record →

  2. Lightcone

    Agents Built, never deployed

    Was this action allowed, and can you prove it?

    Authority is delegated in bounded envelopes that fail closed. Every consequential action is checked against an active envelope before it leaves, and written to an append-only, hash-chained ledger. Knowing what a person wants is not the same as having permission to act on their behalf, and the two are kept apart on purpose.

    Implemented and tested. No production deployment, no users.

    How the boundary works →

  3. Finance Domain Pack

    Domains First proving ground

    Where does a weak model fail fastest?

    Markets answer quickly, objectively, and under costs that cannot be waved away. Everything market-specific lives here: data adapters, cost models, regimes, execution. What the loop learns about evidence does not, which is the entire reason the two are kept in separate layers.

    Dense data, fast feedback, and an unforgiving null

    Why finance first →

These are what is far enough along to show you, not the whole list. XG Capital Strategies runs the research and does not stop; Kairos is being built to productize the parts that survive contact with it, under a licence that has not been executed. New domain packs are expected to arrive the same way, from the laboratory end rather than from a roadmap.

WHAT YOU JUST WATCHED

What must a prediction survive before it is allowed to influence a decision?

The field above runs fifty-six of these at once. Here are three, slowly.

  1. Candidate A Looks convincing on the history it was found in.

    REJECTED It never separated from noise. The rejection is kept.

  2. Candidate B Gets further. Survives the first challenges.

    MORE EVIDENCE REQUIRED Not refuted, but not established either. It waits rather than advancing.

  3. Candidate C Clears every historical gate.

    FORWARD OBSERVATION Now it has to survive time it has never seen.

Evidence does not grant authority.

Candidate C has earned the right to be watched, not the right to act. Crossing that boundary is a separate, explicit authorization step, and it can be withdrawn without the evidence changing at all.

Three schematic candidates, drawn to explain the gate sequence. Not a recorded HSL result, not a performance claim, and not Kairos telemetry. The gate order and names are the ones the research system uses.

Open the Observatory
THE RESEARCH / 5W + 1H

What is being researched, why it matters, and how to read the system.

WHAT
Candidate predictive relationships

Possible models, signals, or decision rules for one clearly bounded decision problem.

WHY
Trust before action

Promising patterns are easy to generate. Kairos exists to determine which deserve belief and bounded use.

WHO
Accountable decision teams

Researchers, risk owners, operators, and leaders responsible for forecasts, models, or automated decisions.

WHERE
Inside the data boundary

Run in customer-controlled or approved private environments; finance is the first proving domain.

WHEN
Before and after deployment

From initial search through historical testing, forward observation, approval, and ongoing monitoring.

HOW
Search, attack, remember, govern

Generate many candidates, falsify aggressively, preserve failures, observe survivors, and authorize separately.

RECORDED RESEARCH ANCESTRY / HSL

Now inspect what actually happened in the research process.

The 5W + 1H above defines the research problem. Below is the research record itself. It is process evidence, not live Kairos telemetry or a claim of model performance.

RECORDED TRAINING-GYM REJECTIONS

Why do candidates stop?

A training gym is one of the search programs XGCS leaves running: it proposes candidate predictions continuously, and a referee decides which are allowed to enrol. The field above is a schematic. This is the real register, every rejection those gyms recorded across the 4 of 8 programs that have logged any, sorted by the reason each was attributed to.

92.4%

of 8,855 recorded rejection records stopped at one place: never separated from noise. Almost nothing survives far enough to fail for an interesting reason.

  1. 8,186 never separated from noise 92.4% of the register · V1 / holdout

The other 669, shown at their own scale. Together they are 7.6% of the register.

  1. 317 a simpler baseline already explained it R2 / spanning
  2. 209 scrambled data scored just as well R1 / placebo
  3. 101 failure mode not resolved by the export classifier unclassified
  4. 42 did not survive trading costs R4 / net of cost

What this shows. Where the search's own rejections were attributed, across every program that has recorded any. Why it matters. The rejections are kept rather than discarded, so the register of what did not work is itself part of the research record.

BOUNDARY / This is implemented XGCS research ancestry, not Kairos customer telemetry or historical performance. Kairos authority and deployment states remain a separate specified architecture.

Recorded process activity is not proof of predictive quality, return, confidence, or deployment authority.

AS OF Aug 26, 2026, 10:19 PM UTC

Recorded activity from the XGCS training gym, which is the same record on a time axis and can be scrubbed frame by frame. Sanitized for publication. Counts are rejections attributed to a failure class, not a survival funnel: the referee wires its checks into one enrol-or-reject decision rather than a sequence of stages, and it qualifies candidates rather than ending them. A check appears here only when the export classifier resolved rejections to it. The unclassified row is known to contain persistence failures that the classifier did not resolve, so the absence of a row is not evidence that a check recorded nothing.

hsl_20260826T221952Z_30f6d81ac86e
RESEARCH ANCESTRY / STATUS LEDGER

Substantial machinery exists. The product boundary remains explicit.

IMPLEMENTED RESEARCH ANCESTRY

Heatmap Strategy Lab

Perpetual candidate generation, breeding and culling, quality-diversity preservation, null-result memory, Referee evaluation, and forward-observation machinery.

IMPLEMENTED / EVOLVING

Null Compass

Failed research becomes structured information that steers future search instead of disappearing into notebooks or chat history.

IMPLEMENTED CONTROL ANCESTRY

XG Engine

A model is registered as shadow by default and cannot promote itself. Promotion is fail-closed by design: the console refuses the flip unless every bar passes, and a bar that is not yet built refuses rather than waves through. The evidence-span bar is enforced today, at least sixty trading days of shadow readings with no long gaps. Research sets the thresholds and a separate platform action performs the flip, so the party that sets the bar is never the party that clears it.

SPECIFIED / NOT YET PRODUCTIZED

Kairos Core

The reusable cross-domain layer is architected. Productization, hardening, private deployment, and commercial validation remain the work ahead.

BOUNDARY / Predecessor systems show that the advancement philosophy on this page already exists in working software rather than only in a plan. They do not prove external product-market fit, a finished cross-domain product, or verified live trading alpha.

DEPLOYMENT PRINCIPLE

Software should go to the data.

01 / PREFERREDCustomer-controlled private

Run Kairos where sensitive data already lives, under customer security and governance rules.

02 / OPTIONALKairos Node

Integrated local workstation or server profiles without turning Kairos into a hardware manufacturer.

03 / WHERE APPROPRIATEManaged private or cloud

The same Core logic where confidentiality, latency, and policy requirements permit hosted operation.