Separate signal from story
Controls, nulls, resource ceilings, and provenance travel with every result.
FIELD NOTES / EVIDENCE-BOUND RESEARCH
CHELATEDAI
A public lab notebook for asking better questions of retrieval, memory, and computation—then keeping the answer honest when the evidence says not yet.
PUBLIC NOTE // This archive distinguishes a reproducible code or fixture result from a scientific claim. A promising lead is still a lead.
CALIVE
LEDGER
LEDGER
MODEOPEN / AUDIT
CLAIMSBOUNDED
01 / ORIENTATION
The map before the myth.
Research is most useful when the boundary around a result is visible.
CHELATED AI explores whether retrieval and agent systems can treat evidence as a changing, structured environment rather than a flat list of nearest vectors or disposable summaries. The work moves between adaptive correction, graph memory, controlled perturbation, governed variation, and computation that may live closer to storage.
The archive is a translation layer between working sessions and durable evidence. Each entry names the question, records what was actually run, links to the source surface, and states what remains unproven.
Working rule: a validated harness proves that a test executed correctly. It does not, by itself, prove that the proposed mechanism improves a real model.
Keep the useful unknown
Discovery claims need an off-span or held-out test—not just a better score on familiar data.
Publish the dead ends
Closed, blocked, and negative lanes are part of the map and prevent accidental rediscovery.
Prefer reversible state
Interventions should be resettable, lineage-aware, and measurable at query time.
01B / CURRENT FIELD NOTE
Rotate the route. Preserve the manifest.
Resonant holographic path compression / active mechanism hypothesis
ACTIVE HYPOTHESIS
CONSTRUCTED FIXTURE ONLY
What if frequently used expert routes were represented as rotations of shared low-dimensional slices instead of duplicated as full paths? Rare structure could remain available as sparse residual “holograms” that stack only when a route needs them.
The route itself becomes an ordered specifier chain. A fixed-width holographic signature can help retrieve or compare that chain, but it never replaces the lossless manifest that records station order, permutations, residuals, and provenance.
Recurrent unbind–clean–rebind steps seek a stable relative phase. Convergence is the lock signal; persistent cycles are treated as standing refractions and rejected, rather than blurred away by smoothing.
01C / DEEP RESEARCH SURFACES
Follow the equation into the evidence.
Four ways into the same research record
LABS / 15
Interactive explainer atlas
Manipulate the equations, assumptions, controls, and falsifiers behind every full paper and active topline theory—including RHPC.
Explore every mechanism ↗ PAPERS / 10Full preprint library
Canonical web editions, typeset PDFs, BibTeX, formal methods, primary references, limitations, and falsifiable promotion rules.
Browse research papers ↗ CLAIMS / LEDGERFindings matrix
Search what survived, what failed, what remains blocked, and the exact boundary between a fixture and a scientific claim.
Inspect findings ↗ TIME / TRACEResearch log
Walk the durable chronology—including nulls and no-promotion decisions—without reading raw machine transcripts.
Trace the journey ↗02 / RESEARCH LANES
One repo. Several pressure tests.
Choose a lane to see its present evidence boundary.
HOW TO READ THE FIGURES
Observed or implemented
Proposed or open
Boundary, unresolved gate, or negative result
Explanatory schematics—not measured plots. Every caption states the evidence limit.
03 / MILESTONES
A trail of tested ideas.
Dates mark the archive, not a claim of finality.
04 / STATUS BOARD
What the evidence says now.
Status labels are deliberately conservative.
Bounded evidence confirms a fixture, implementation surface, or analytic result.Unconfirmed is a hypothesis or lead awaiting independent utility evidence.Blocked names a dependency or missing control; it is not a negative result.Closed / negative means the tested version was ruled out or narrowed.
05 / RESEARCH PAPERS
Full manuscripts. Explicit evidence boundaries.
Citable web editions and typeset PDFs / versioned preprints / not peer reviewed
Each paper is a complete scholarly manuscript with an abstract, related work, formal framing, methods, results or preregistered protocol, limitations, availability statement, and verified references. HTML is the canonical reading edition; the matching PDF is the stable downloadable edition.
Publication-grade does not mean peer reviewed. Claims remain bounded by the exact evidence named in each manuscript, and negative or unrun results stay visible.
No research papers loaded
The paper index is unavailable; the reviewed session archive remains available below.
06 / SESSION ARCHIVE
Search the working record.
No matching session
Try a broader term or clear the active filters.
READING GUIDE
Evidence is a first-class output.
Cards marked BOUNDED EVIDENCE report what a named test or analytic check established. They intentionally avoid upgrading fixture validity into model utility. Cards marked UNCONFIRMED are live hypotheses or non-claim metadata pointers. Cards marked BLOCKED name a missing dependency, while CLOSED / NEGATIVE means the tested version was ruled out or narrowed.
Read the reviewed findings ledger