The customer already has models.
Import or black-box candidate systems, attack them, build evidence, monitor them, and govern advancement without requiring proprietary source code by default.
FINANCE DOMAIN PACK / FIRST PROVING GROUND
Trading first. Not trading only.
Financial markets provide dense data, objective outcomes, severe non-stationarity, transaction costs, fast feedback, and a long internal research history. That makes finance an unusually demanding place to prove the Kairos research and evidence loop before expanding into other predictive decision environments.
Start with existing models, or start with a decision problem.
Kairos is being designed to support both research-control workflows and autonomous predictive-system research against customer-controlled data.
The customer has data and a decision problem.
Generate, challenge, evolve, validate, monitor, and govern predictive systems without requiring the customer to first build an internal autonomous quantitative-research organization.
The public record shows process scale, retained diversity, and negative memory.
These counts are drawn from a dated, website-safe snapshot of the HSL research record, not a live feed. They describe research operations and current archive state, not profitability or predictive performance.
DATED WEBSITE-SAFE SNAPSHOT
Finance forced the research process to remember what failed.
Recorded XGCS research ancestry. Not Kairos customer telemetry, live trading, historical performance, or proof of product-market fit.
- Programs
- 8 37.5% contaminated
- Evaluations
- 942,601 12.8% contaminated
- Archive cells
- 374 59.9% contaminated
- Rejection records
- 8,855 31.2% contaminated
- Recent events
- 1,444
- History samples
- 4,950
REJECTION RECORDS / Typed archetype rows in the null registry, which the registry flushes periodically. One row summarises a class of failure rather than one killed candidate, so the number of candidates actually culled is far larger than this figure.
These counts are real. What they imply is not.
- Programs: 3 of these 8 programs (37.5%) are affected by the look-ahead found on 2026-08-27. The activity is real and the decisions behind it are not trustworthy, so read 5 as the qualified figure and the rest as work that happened.
- Evaluations: 120,795 of these 942,601 evaluations (12.8%) are affected by the look-ahead found on 2026-08-27. The activity is real and the decisions behind it are not trustworthy, so read 821,806 as the qualified figure and the rest as work that happened.
- Archive cells: 224 of these 374 occupied cells (59.9%) are affected by the look-ahead found on 2026-08-27. The activity is real and the decisions behind it are not trustworthy, so read 150 as the qualified figure and the rest as work that happened.
- Rejection records: 2,765 of these 8,855 records (31.2%) are affected by the look-ahead found on 2026-08-27. The activity is real and the decisions behind it are not trustworthy, so read 6,090 as the qualified figure and the rest as work that happened.
Broad cross-asset search
risk, regime, and candidate search
- STATE
- running
- EVALUATIONS
- 126,146
- ARCHIVE
- 150
Field direction search
directional falsification
- STATE
- stopped
- EVALUATIONS
- 108,854
- ARCHIVE
- 138
These counts are real. What they imply is not. A look-ahead was found in this program's exit rule on 2026-08-27. The counts below are real: these experiments ran and these candidates were retained. But the decisions that retained them were made on scores that are now known to be wrong, so retention here means the process kept it, not that it earned keeping. The program is stopped and is being rebuilt from scratch on corrected code.
Field surf search
relationship and participation search
- STATE
- running
- EVALUATIONS
- 11,941
- ARCHIVE
- 86
These counts are real. What they imply is not. A look-ahead was found in this program's exit rule on 2026-08-27. The counts below are real: these experiments ran and these candidates were retained. But the decisions that retained them were made on scores that are now known to be wrong, so retention here means the process kept it, not that it earned keeping. The program is stopped and is being rebuilt from scratch on corrected code.
AGGREGATE FAILURE MEMORY
8,186significance
317redundancy
209placebo
42cost
101other
LINKED EVIDENCE EXPLORER / RECORDED
One program selection updates the whole research record.
Choose a program, then inspect its recorded scale, history, retained archive shape, and structured failure memory. These are process and archive KPIs, not model-performance results.
SELECTED PROGRAM
Broad cross-asset search
risk, regime, and candidate search
- STATE
- running
- GENERATION
- 1,551
- EVALUATIONS
- 126,146
- ARCHIVE
- 150
- COVERAGE
- 23.1%
- REJECTIONS
- 2,915
Left to right through available website-safe samples.
Click a bar to inspect what the category means.
Research family
Information reach
Market regime
Directional evidence
Cost harvestability
Click a reason to inspect what it rules out.
Finance is where the machinery was forced to become skeptical.
Heatmap Strategy Lab research has repeatedly confronted the difference between visually compelling structure and evidence strong enough to survive holdout, cost, robustness, placebo, and forward-observation gates.
Perpetual candidate generation
Multiple candidate families, mutation, recombination, architecture exploration, and resurrection are evaluated inside a continuing research loop rather than a one-off notebook.
Preserve diversity under selection pressure
Quality-diversity archives and novelty-aware culling reduce collapse into thousands of superficially different but behaviorally redundant strategies.
Separate promising from trusted
Historical profitability can qualify a candidate for further observation, but it does not automatically make the candidate live-deployment eligible.
Reality remains the executioner
Frozen candidates accrue prospective evidence so research success cannot be retroactively rewritten after future information arrives.
Claims boundary: the record supports a serious research-control system. It does not support a claim of verified live alpha, and it should not be read as one. No standalone directional edge that beats the market net of costs has survived the gates. Several directional signals are statistically real; what stops them is the cost wall rather than absence, which is a different finding and a more useful one. The clearest positive result so far is not about direction at all: it is what the field encodes about risk and magnitude. The forward book of enrolled candidates has earned approximately nothing after costs. That is the system working, not the system failing, but it is the opposite of a performance claim.
The purpose of finance is to prove the Core, then test transfer.
Market-specific data adapters, costs, candidate families, regimes, and execution logic remain in the Finance Domain Pack. The reusable layer is the evidence, research-control, lineage, null-memory, and advancement infrastructure.
FinanceExternal workflowsMedia / investmentOther data-rich domainsDomain evidenceReusable methodsCore improvesNew packsRepeat