Skip to main content

Laboratories

Environments built to observe before they conclude

XGCS tests whether different ways of reading market structure remain useful beyond their first fit. Each laboratory defines the subject, instrument boundary, provenance, and evidence standard before it makes a claim.

LAB 01 / Public

Heatmap Strategy Lab

A public-safe record of multi-program candidate search, archive geometry, failure memory, and recorded research activity. Source, timestamp, coverage, and exclusions remain visible inside every instrument.

Subject
Candidate search, research throughput, archive diversity, prospective observation, and controlled failure memory.
Public mode
Verified website-safe snapshot with a bounded recorded replay. No continuous live connection is claimed.
Private boundary
Internal formulas, strategy logic, and sensitive research data remain private.
Claim boundary
The instrument does not promise investment performance or remove uncertainty.

The finding / What the search has settled

Direction is null. Risk is not.

Across every price and OHLCV-derived ordering the laboratory has tried, the field carries information about risk and magnitude, and direction comes back null. That is the clearest result the work has produced, and it took the larger part of the search to reach.

The precise form matters. No net-of-cost, market-beating standalone edge has survived. That is not the same as saying there is nothing there: several directional signals are statistically real and then die on transaction costs, which makes them harvest-walled rather than absent. Nor is it a claim that the space is exhausted or closed. It is a statement about what has been tested, on the data the laboratory currently produces.

It is published here because a laboratory that only reports the questions still open is not reporting. The exhibits below are the record this conclusion was drawn from, including the register of what failed and the retraction of the one performance figure this site published.

Public artifact / Field census

First see the machine as a whole.

The census shows which approved research programs were active, the scale of recorded evaluation, and a bounded sample of sanitized process events. It describes research activity, not trading performance.

XGCS / HSL FIELD CENSUS

INSTRUMENT
RECORDED RESEARCH ACTIVITY
ARTIFACT
hsl_20260826T221952Z_30f6d81ac86e
SOURCE
Verified website-safe HSL snapshot
AS OF
Aug 26, 2026, 10:19 PM UTC
RECORDED / BOUNDED REPLAY 8 PROGRAMS / 942,601 EVALUATIONS

Who / What / Why

A research machine that keeps both survivors and failure.

Heatmap Strategy Lab searches across multiple research programs. Candidate structures are generated, challenged, archived, rejected, revived, or moved into forward observation. The public view preserves the shape of that work while excluding formulas, candidate identities, returns, forecasts, and positions.

Programs
8
4 running, 3 paused, 1 stopped
Evaluations
942,601
research throughput, not performance
Archive cells
374
current occupants across three programs
Rejection records
8,855
typed records, each batching many rejected candidates

Read these totals with the contamination in mind. A look-ahead was found in three programs' exit rule on 2026-08-27. Those programs account for 224 of 374 archive occupants (60%), 120,795 of 942,601 evaluations (13%) and 2,765 of 8,855 rejection records (31%). Their work is counted here because it happened, not because it stands: the decisions that retained those candidates were made on scores now known to be wrong. The programs are stopped and are being rebuilt from scratch on corrected code.

When / How

Recorded activity across a bounded event sample.

EACH MARK = ONE EVENLY SAMPLED RECORDED EVENTREPLAY ORDER / EARLIER → LATER
Recombine 418Research record 131Advance 268Archive 46Reject 23Resurrect 46Forward observe 26Checkpoint 1

WHERE / Heatmap Strategy Lab, recorded in snapshot hsl_20260826T221952Z_30f6d81ac86e.

FRESHNESS / ARCHIVAL SNAPSHOT. Exported at Aug 26, 2026, 10:19 PM UTC. This page does not claim a continuous live connection.

EXCLUDED FROM THIS INSTRUMENT / Positions, orders, forecasts, returns, formulas, genomes, model weights, raw event text, and private topology. The forward track further down this page used to publish a return; it was withdrawn on 2026-08-27 after a look-ahead was found in the program behind it, and no return figure is published anywhere on this site today.

A public-safe census of 8 approved HSL programs and a representative replay drawn from 1,444 sanitized recent events. The event record is bounded, not complete history. Motion shows recorded process activity, not live computation or performance.

Public artifact / Model evidence

Now test whether an attractive result survives.

This approved aggregate follows one formulation from a positive selection-period result into a negative independent validation result. Click each mark and control to inspect why the candidate stopped before the locked forward holdout.

PUBLISHED SPECIMEN RECORD / REJECTED AT VALIDATION

A promising fit is allowed to fail.

This recorded case answers a narrow question: does the laboratory stop an initially attractive candidate when independent validation and controls disagree?

CLAIM TESTEDSelection-period evidence should survive independent validation.
RECORDED VERDICTREJECTED
MARKET SCOPEBroad US equity panel
NEVER MEASUREDthe 2026+ test was left unopened
Recorded stage results. Click a row to inspect why the candidate stopped.
Stage Rank correlation Overlap-safe t Dates Reading
+0.0249 1.123 38 Not distinguishable from zero
-0.0051 -0.202 26 Not distinguishable from zero
NOT CONSUMED Never opened

Neither measured stage clears its own significance test. This is not a relationship that existed and then reversed; it is a number that was never separable from noise, and a second period that confirmed it. The locked holdout was never opened.

CLICK A CONTROL TO INSPECT WHY THE CANDIDATE STOPPED
SAMPLE / 13,388 pooled validation observations
UNCERTAINTY / Overlap-safe t: selection 1.123 across 38 dates; validation -0.202 across 26 dates.
COST CONTROL / Rank correlation is non-economic; a separate 3 bps validation control also failed.
Read the publication limits
  • This is one rejected cross-sectional divergence formulation, not a conclusion about all field or market-structure research.
  • Selection and validation estimates are historical aggregate tests, not live trading results or future-performance claims.
  • The locked 2026+ holdout was not evaluated, so no forward estimate exists.
  • No implementation internals, raw series, or actionable outputs are included.

PROVENANCE / hsl-x2w-rejected-41bc25a5bce2658d · approved aggregate · data through Jan 1, 2026

Public artifact / Archive reconstruction

The laboratory shows its search shape.

This artifact uses the verified website-safe HSL snapshot. It can show current archive occupants and recorded update generations; it cannot support conclusions about quality, performance, or the archive's prior history.

XGCS / LIVING FIELD LAB

INSTRUMENT
QUALITY-DIVERSITY ARCHIVE
ARTIFACT
HSL-QD-ARCHIVE-01
SOURCE
HSL website-safe snapshot / Broad cross-asset search
AS OF
Aug 26, 2026, 10:19 PM UTC
RECORDED / VERIFIED CURRENT AT EXPORT 150 / 648 NICHES

Recorded field / Five dimensions / One public boundary

A search field shaped by what survives.

Six primary families hold the other four axes still: reach runs vertically by regime; direction runs horizontally by cost. The compass stays fixed while the evidence field changes around it.

What to take from this: the search has reached 150 of 648 possible behaviours, about 23% of the space. The lit cells are where it has found something worth keeping. The empty ones are mostly where it has not looked yet, not where it tried and failed.

ROWS reach × regime
COLUMNS direction × cost
COLOR + HATCH recorded occupant update generation
01

field

73 / 108
02

correlation

4 / 108
03

cross-sectional

18 / 108
04

rolling

32 / 108
05

time-series

12 / 108
06

other

11 / 108

One-pass reconstruction · pauses offscreen and in background tabs

SELECTED NICHENone; choose a lit cell
LAST CLAIMUnknown
MODERecorded / verified current at export
AS OFAug 26, 2026, 10:19 PM UTC

Inference boundary. Update generation is not discovery time, persistence, quality, return, or confidence. The reconstruction describes current occupants in one recorded snapshot only.

Current-occupant reconstruction from the verified website-safe HSL snapshot. Each lit cell is one occupied behavior niche; color and hatch indicate its recorded update generation. Earlier occupants, archive history, per-candidate fitness, returns, candidate identities, formulas, and verdicts are omitted from this snapshot. Empty means unoccupied in this snapshot; it does not mean failed.

Public artifact / Failure memory

The laboratory keeps what did not survive.

The archive alone would show only survivors. Failure Memory restores the other half of the research process by retaining aggregate rejection records without exposing candidate identities, formulas, or individual outcomes.

XGCS / NULL REGISTRY

INSTRUMENT
FAILURE MEMORY
ARTIFACT
HSL-NULL-REGISTRY-01
SOURCE
HSL website-safe snapshot / Eight approved research programs
AS OF
Aug 26, 2026, UTC
RECORDED / VERIFIED CURRENT AT EXPORT 8,855 AGGREGATE REJECTIONS

Null is a result / absence is not zero

What did not survive remains part of the instrument.

Most public research surfaces display only the surviving line. This register gives typed failures a durable place without publishing candidate identities, formulas, individual outcomes, or raw event text.

VIEW All aggregate rejection records

8,855 of 8,855 aggregate rejections visible.

Deliberately excluded: candidate names and identifiers, genomes, formulas, fitness, forward P&L, verdicts, individual OOS results, and event text.

One mark equals one aggregate rejection record in the verified website-safe snapshot. Marks are grouped by controlled failure class for legibility, not time. The record spans approved HSL programs; counts do not measure strategy quality, loss, return, or confidence.

RECORDED OVER 10 DAYS / Aug 17 TO Aug 26

How many candidates actually make it through?

Every other view of this record is a running total. This is the only one that moves: candidates entering, waiting, surviving, and being rejected, sampled continuously across 4,950 readings.

  • Waiting on more evidence
  • Enrolled for observation
  • Still surviving
  • Rejected by this program
Field surf search 546 readings, peak 953
Waiting on more evidence
553 → 72
Enrolled for observation
560 → 953
Still surviving
49 → 905
Rejected by this program
0 → 0unchanged
Broad cross-asset search 788 readings, peak 194
Waiting on more evidence
92 → 106
Enrolled for observation
102 → 114
Still surviving
7 → 7unchanged
Rejected by this program
194 → 194unchanged
Field direction search 763 readings, peak 12
Waiting on more evidence
0 → 0unchanged
Enrolled for observation
9 → 12
Still surviving
7 → 10
Rejected by this program
2 → 2unchanged
Market field world model 714 readings, peak 2
Waiting on more evidence
2 → 2unchanged
Enrolled for observation
1 → 1unchanged
Still surviving
0 → 0unchanged
Rejected by this program
0 → 0unchanged

Recorded XGCS research activity, sanitized for publication. Each line is scaled to its own program's peak, so heights are not comparable between programs. 4 of 8 programs recorded no enrolment activity in this window and are not drawn. A flat survivor line is a real result, not a missing one.

  • Recorded research activity is not live trading telemetry.
  • Archive occupancy and research throughput do not establish predictive performance.
  • No positions, orders, forecasts, returns, formulas, genomes, or model weights are published.
hsl_20260826T221952Z_30f6d81ac86e

Withdrawn 2026-08-27

This exhibit published a return. It should not have.

Until 2026-08-27 this page showed a forward track: 136 candidates from the field-direction-surf program, frozen on 2026-07-21, watched for 25 sessions without intervention, most of them positive. It was the only forward performance figure on this site. That program is named in the weekly research notes on this site, so this retraction can be checked against our own published record rather than taken on trust.

It was wrong, and the defect was in our own code. The exit rule for that program read its inputs at the close of a bar and was then scored against that same bar's return, so it closed losing positions using information that did not exist at the moment the decision would have been made. Shift the exit by the one bar it should always have been shifted by and the result does not shrink, it disappears.

Four candidates from the program, measured on holdout, n=581. Before is what we published. After is the same candidate with the one-bar alignment corrected.
Candidate Sharpe, gross Sharpe, after costs Information coefficient
Best enrolled 13.10.200 11.5-1.8 0.1090.004
Median enrolled 16.00.500 14.1-1.4 0.1060.003
Worst enrolled 4.0-0.300 2.9-1.3 0.039-0.001
Already culled 15.30.300 13.6-1.4 0.1050.001

Every one of them goes to approximately zero gross and to a loss after costs. The information coefficient, which is the part that would have to be real for any of this to mean anything, goes to within rounding distance of nothing. This was checked against the corrected code as deployed rather than an offline patch, and the original figures can no longer be reproduced.

The forward track carried the same defect, because it re-runs the same exit. It was never a clean out-of-sample number, and no amount of forward observation would have revealed that, because the leak travels with the rule rather than with the data.

We found it ourselves, by attacking our own result rather than defending it. Nobody reported it, no counterparty lost money on it, and it was caught before it was ever acted on. That is the machinery on the rest of this page doing the job it exists to do: a result is held against the tests that would kill it, and this one died.

It is the reason the exhibit is gone rather than footnoted. A qualified performance number would ask you to trust a judgement about how much of it survived. There is no such judgement available here, and the honest version of this page is one with a hole in it. A research record is only worth reading if the people keeping it will also publish the entries that cost them something.

Nothing replaces it today. The programs unaffected by this defect are honest and earn roughly nothing after costs, and putting those here instead would be choosing a number because it was available rather than because it meant something. The affected programs are stopped and are being rebuilt from scratch on corrected code. If a replacement figure ever appears in this space, it will have been through the same tests that removed this one.

Every exhibit above is a still. The training gym puts the same record on a time axis and lets you scrub it.