Skip to content
Rutger Hemrika
Menu

Eight predictions, graded in public

In May 2026 I published a framework scoring ten programming languages for the AI era, and eight falsifiable predictions tied to its scores. This page tracks them until they are graded. Every prediction has a date, a measurable check and a condition that would prove it wrong.

8
open
0
held
0
missed
  1. P1Open

    Pydantic maintainership, or Python dependency risk falls

    Either Pydantic has at least three active maintainers, each merging a pull request in the prior six months, or Python’s ecosystem-dependency-risk score moves toward 2.0 in the next major revision.

    Tied to
    Python, ecosystem dependency risk: 2.5
    Grading date
    How it will be checked
    Six-month merged-PR count per maintainer on pydantic/pydantic.
    Falsified if
    Pydantic stays under three active maintainers and Python’s dependency-risk score does not fall.
  2. P2Open

    C++ keeps losing greenfield share in security-sensitive domains

    C++’s share of new software in defence, finance and similar domains keeps declining year over year.

    Tied to
    C++, weighted total 2.46, rank 10
    Grading date
    How it will be checked
    Year-over-year greenfield signal in the CNCF, JetBrains, RedMonk and GitHub Octoverse surveys.
    Falsified if
    C++ greenfield share holds steady or rises in those surveys.
  3. P3Open

    TypeScript structured-output verification improves

    TypeScript’s structured-output-native score rises toward 3.0 in the next major revision, driven by satisfies-style verification, the type-test ecosystem and erasable-syntax ergonomics.

    Tied to
    TypeScript, structured-output native: 2.5
    Grading date
    How it will be checked
    A review of the dimension, plus TC39 type-stripping progress and community verification idioms.
    Falsified if
    The score stays at 2.5 or falls.
  4. P4Open

    A first-party .NET SDK for Anthropic models

    Microsoft or Anthropic ships an official .NET SDK for Anthropic models with a stable release, which would raise .NET’s AI-native score by at least 0.5.

    Tied to
    .NET, AI-systems native: 4.0
    Grading date
    How it will be checked
    A repository under microsoft/ or anthropics/ with Microsoft engineering attribution and a stable release.
    Falsified if
    No first-party .NET SDK reaches a stable release by the date.
  5. P5Open

    Rust compile-time ergonomics improve

    Rust’s AI-agent operability rises from 4.0 to 4.5 as the parallel front-end stabilises and a fast linker becomes the stable default.

    Tied to
    Rust, AI-agent operability: 4.0
    Grading date
    How it will be checked
    rustc release notes and cold and incremental rebuild benchmarks.
    Falsified if
    Either change is missing from stable rustc, or the score stays at 4.0.
  6. P6Open

    Java Panama and Loom mature

    Java’s machine-cognition and agent-operability scores each rise 0.5 after the Vector API becomes final and virtual threads become production defaults.

    Tied to
    Java, machine cognition 4.0, agent operability 3.5
    Grading date
    How it will be checked
    OpenJDK release notes and virtual-thread defaults in Spring Boot and Quarkus.
    Falsified if
    Panama and Loom slip, or ship without the scores rising. If they slip, Java’s reachability is revised down.
  7. P7Open

    Swift stays capped outside Apple’s platforms

    Swift’s AI-ecosystem score does not improve materially and stays at or below 2.5, because its non-Apple agent-framework reach does not grow.

    Tied to
    Swift, AI-systems ecosystem: 2.0
    Grading date
    How it will be checked
    A catalogue of Swift AI packages, and whether a Swift agent framework with non-Apple CI exists.
    Falsified if
    The score rises above 2.5. A jump to 3.5 or more weakens the thesis that steward concentration caps reach.
  8. P8Open

    Elixir ships a stable set-theoretic type spec, or is revised down

    Elixir’s set-theoretic types reach a stable specification, or Elixir’s reachability score is revised down.

    Tied to
    Elixir, reachability: 2.5
    Grading date
    How it will be checked
    elixir-lang.org release notes, stable type-system documentation, uptake in Phoenix and Ecto.
    Falsified if
    No stable spec is published and the score is not revised down.

Why publish this

A prediction has a date stamp, a measurable check and a state of the world that falsifies it. An opinion has none of those, which is why opinions cost the writer nothing and predictions cost everything. If four or more of these miss, the framework gets a structural revision, not a score tweak.

Calibration rule

A pass rate of at least 60% counts as useful forecasting. Four or more misses out of eight trigger a structural revision of the framework, not score tweaks.

Source and method: v0.6, 2026-05-02, GitHub.

Explore the scorecardRead the predictions article