// research

Research & Technical Reports.

Systems work that we can show our working for. Each report ships with the source that produced its numbers, and every table in it reproduces offline on your own machine — including the rows where the result went against us.

// latest

Current Report

DCR-TR-2026-01 21 August 2026 LLM systems 17 pages

Dynamic Context Runtime: Bounded Attention over Unbounded History

Language models degrade as stale and superseded material piles up in their context. Making the window bigger does not fix it. This report describes a runtime that keeps history unbounded and attention bounded — storing everything as immutable spans and a typed provenance graph, then assembling a small working set each turn by solving a knapsack under an explicit token budget. It includes a zero-dependency Rust implementation, an ablation that names which mechanisms are actually carrying the result, and a scaling table that holds at 4.19 million tokens of history on a lexically varied corpus — alongside an account of why the same measurement on a repetitive one would have meant considerably less.

  • 4.19Mtokens of history
  • 259tokens per query
  • 7 / 7probes answered
  • 48,651state nodes
Download PDF Read in browser Source & experiments

// also featured

Passwordless Login: the Double-TOTP Study

Feasibility study 31 August 2026 Authentication 2 × 8 digits

Can Two 30-Second Codes Replace Your Password?

A feasibility study of passwordless login built entirely from two independent 8-digit, 30-second TOTP layers — no password, no shared secret typed anywhere. Both codes must be correct inside the same window and the fail-anywhere-restart rule squares the effective code space: at the NIST attempt cap the expected time to break the pair is roughly 950,000 years, and the paper is as clear about where that number leaks (clock drift triples the odds; a synced seed compromise voids it) as it is about where it holds.

  • 100Mvalues per 30-second window
  • 1×10⁻¹²odds per double-layer cycle
  • ~950,000 yrsexpected crack time (NIST cap)
  • 2 × 30 slayers — fail anywhere, restart
Read the full study Download PDF More resources

// index

All Reports

Aug 2026 DCR-TR-2026-01

Dynamic Context Runtime: Bounded Attention over Unbounded History

A representation ladder, a provenance graph, and a budgeted context planner. All seven evaluation probes answered from a 200-token window over a 27,362-token transcript, with the working set flat across a 33× growth in history.

Read report
Aug 2026 Authentication

Can Two 30-Second Codes Replace Your Password?

Two 30-second 8-digit 2FA codes instead of a password: the brute-force math, the fail-anywhere-restart rule, and the honest gaps. Feasibility research from Cyber Sec.

Read article

// how we publish

What a report from us contains

01

Numbers you can re-run

Every table and figure lists the command that produced it. The experiments run offline with no API key and no network, so a reader can reproduce or refute them in minutes rather than trusting a screenshot.

02

The source, not a description of it

The implementation behind each report is released with it. A claim about a system that nobody else can run is marketing, and we would rather be checked than believed.

03

Negative results kept in

Ablations that show a mechanism earning nothing stay in the paper. A results table with no losing rows usually means the experiment was designed after the conclusion.

04

Limits stated in the paper

Each report carries an explicit threats-to-validity section saying what its evidence cannot support — synthetic data, missing comparisons, measurement shortcuts. The caveats are not in the footnotes.

// collaboration

Working on something adjacent?

If you are building on this work, found a flaw in it, or want research done against your own systems, get in touch. Critique is more useful to us than citation.

Get in touch Services
↑