Home

ATLAS · Research Universe

Keep many questions. Show only a few. Test them for real.

The private registry contains 622 addressable entries. That does not mean 622 discoveries or 622 results. It means a question can be found, reopened, tested or dropped without depending on a forgotten chat.

Five drawers are enough here. Each keeps the same grammar: question → test → result → what broke → next.

The circuit

The world emits a signal. SPECTRA sees it. ATLAS decides whether it deserves an investigation.

Public circulationWORLD / SOURCES → SPECTRA → ATLAS → LAB → EVIDENCE → MEMORY
01SeeA signal, anomaly or source.
→
02QuestionWhat real unknown remains?
→
03TestA protocol that can lose.
→
04KeepResult, contradiction, next gate.

This describes the method. It does not publish the private corpus, restricted sources or SYSTASYS internal topology.

Five cases

Five domains. One evidence discipline.

The terrain changes. The evidence discipline does not.

MathWritingSourcesFieldHeritage
01Math

XENO / Gerver · 02 Oct 2026

How can a set saturate length-4 progressions while blocking length 5?

Question. For base-5 integers using only digits {0,1,2,3}, which common differences of 4-term arithmetic progressions are actually possible?

Test. Carry automaton, independent recognizer, negative control and independent brute force through L=6.

Result. At each certified finite scale, every theoretically possible difference up to (5^L-1)/4 is realized.

Still open: simpler human proof and deeper prior-art audit. Novelty not claimed.

See the public trace →
Plate 01 · structureDifference spectrum
AL = { n : digits5(n) ∈ {0,1,2,3} }
a+da+d+da+2d+da+3d

d = 1 … (5L − 1) / 4

Public reconstruction of the tested structure, not a complete human proof.
02Writing

Proto-Elamite · F-012

Can a proposed sign function survive the primary corpus?

Question. Do M157 and M388 show structural distributions compatible with the proposed functions?

Test. Frequency-matched controls, within-tablet permutations, sensitivity to uncertain readings and a held-out subset not used to select the candidates.

State. M157 survived the current held-out reading; M388 was disfavored. The method is supported while semantic interpretation remains open.

What broke: the idea that an independent discovery source could itself become evidence authority.

See the public trace →
Plate 02 · held-outThe proposed function has to survive the corpus
M157survived held-out
M388disfavored
Candidate→Controls→Held-out
Public reading of the current dossier state. No semantic value is presented as resolved.
03Sources

Source Ghost · F-019

Can a lost Arabic source be reconstructed without inventing the missing pieces?

Question. Can surviving citations, dependencies and witnesses constrain a lost source?

Test. A frozen Abu al-Fida branch is paired with an independent Yaqut branch.

State. The reconstruction method is the important object; the reconstruction itself remains partial and must keep witness, inference and absence separate.

Next: strengthen independent witnesses before broader synthesis.

See the public trace →
Plate 03 · witnessesReconstruct without filling the gap
Abu al-Fida
lost source
Yaqut

witness ≠ inference ≠ absence

The centre remains dashed on purpose: missing material is not drawn as known.
04Football

Football Lab · P0

Can a tactical intuition survive a match it has not seen?

Question. Find one measurable mechanism that predicts a next-state property on public historical matches.

Test. Freeze the rule before holdout, compare with a simple baseline, no paid data and no betting.

State. The lab now contains possession and geometry tests, but the core value is the discipline of choosing before opening the sealed match.

What can break: an attractive pattern that vanishes out of sample. Then killing it cleanly is the correct result.

See the public trace →
Plate 04 · blindChoose before opening the match
RULE FROZENHOLDOUT SEALED
The visual represents the protocol. It does not simulate a prediction or match result.
05Heritage

Virtual Joins / Vesuvius

When an object is fragmented, do several weak cues beat one strong cue?

Question. Re-associate historical fragments with several cue families without using ground truth as an input feature.

Test. Virtual Joins calibrates on RePAIR then on held-out Genizah fragments. Vesuvius separately tests causal value of annotation corrections rather than correction size.

Useful result. A first Genizah holdout returned MIXED, while Vesuvius already showed that a large label correction can have near-zero downstream learning value.

Lesson: amount of correction or number of cues ≠ useful information. Value must appear downstream.

See the public trace →
Plate 05 · fusionMore cues do not automatically mean more information
shapetextureedgecontext
→
JOIN
→
downstreamMIXED
Value is judged downstream. Cue count or correction size is not itself a result.

What 622 actually means

A memory of unknowns, not a trophy cabinet.

91

consolidated fronts

Families of questions maintained as working territories.

144

TERRIER work items

Historical backlog including dormant lines.

21

named H hypotheses

Named hypotheses can be found and falsified.

36

Ghost Works

Lost works or sources maintained as reconstruction fronts.

The backlog score is a work sort, not a scientific verdict or canonical priority.

Back to Work →