akribia

A unified computational model of precision-weighted Bayesian inference across autism, ADHD, and PPCS

One inference engine. Three miscalibrations. Same math.

▶ Try the live demo — the full dashboard runs in your browser, no install. · ⬇ Download the desktop app

This is not a diagnostic tool

akribia is a computational-psychiatry research replication and exploration tool. It implements published mathematical models of precision-weighted Bayesian inference to illustrate theoretical mechanisms proposed in the literature for autism, ADHD, and PPCS. It is not a diagnostic tool, has not been validated against patient data, and should not be used to assess, diagnose, or characterize any individual's condition — including the author's. The autism precision-weighting literature in particular is actively contested (see docs/THEORY.md); this project models competing hypotheses, not settled fact.

akribia runs entirely on synthetic, simulated data. It does not collect, store, transmit, or process any personal, behavioral, or health information.

Desktop app — the interactive explorer

The dark dashboard — baseline vs. a comorbid profile across all five tasks. Open it live and drag the precision sliders.

akribia ships a rich, dark, interactive desktop app (Tauri + React + Tailwind): pick a profile in the sidebar, drag a precision lever, and watch every task's behaviour update live against the neurotypical baseline — all computed by the same validated Rust core, no Python required.

Install (no build needed): grab the installer for your OS from the Releases page — .msi/.exe (Windows), .dmg (macOS), .AppImage/.deb (Linux). These are built automatically by GitHub Actions when a v* tag is pushed.

Installers are unsigned. Windows: click "More info" → "Run anyway" on the SmartScreen prompt. macOS: right-click the app → Open (or xattr -d com.apple.quarantine <file>).

Or build / run it yourself:

cd gui
npm install
npm run tauri dev      # live dev window with hot reload
npm run tauri build    # installer lands in gui/src-tauri/target/release/bundle/

Requires Node, a current stable Rust toolchain, and WebView2 (preinstalled on Windows 10/11). The Python research layer (below) stays available for notebooks, sweeps and validation.

See the thesis demonstrated in ~90 seconds

→ notebooks/00_golden_path.ipynb — baseline vs. one profile vs. one plot. The single image below is the whole idea: the same engine and the same task, under two precision parameterizations, produce different perceptual behaviour.

The neurotypical baseline "sees" a Kanizsa illusory triangle (high illusion-susceptibility score); autism_weak_prior caps prior precision, so absent local evidence dominates and the illusion weakens — the literature's reduced illusion finding, reproduced.

Quickstart

git clone https://github.com/Londopy/akribia.git && cd akribia
docker compose -f .devcontainer/docker-compose.yml up -d   # or: maturin develop && pip install -e ".[dev]"
python -m akribia.tasks.illusion_task --profile autism_weak_prior --plot

Runs the illusory-contour task under the weak-prior autism profile and saves a comparison plot against baseline. Swap --profile for any entry in the Profile Catalog. No Rust toolchain? It still runs — the package falls back to a pure-Python core (python -c "from akribia import core; print(core.BACKEND)"). Or launch the interactive app: cd gui && npm run tauri dev.

The thesis

Predictive coding treats the brain as a hierarchical inference machine: a prediction meets evidence, the mismatch is a prediction error, and that error is weighted by precision (inverse variance — how much the system trusts the signal) before updating beliefs. The same precision-weighting math, miscalibrated at different points in the hierarchy, produces phenotypically distinct conditions:

akribia implements one core engine with pluggable "lesion profiles" (one PrecisionProfile dataclass, six levers), plus a literature-grounded comorbidity (AuDHD) mode — because co-occurrence is common and the more realistic case to model.

Theory

Each condition's module reproduces specific, pre-registered predictions from the literature (encoded as tests/test_predictions.py):

• Autism — weak priors reduce illusion susceptibility; HIPPEA (inflexible precision) produces a transient reconvergence delay after a context switch.
• ADHD — steep discounting collapses the delay-discounting AUC; unstable reward gain produces erratic learning.
• PPCS — an impaired forward-model update rate leaves a persistent vestibular mismatch that does not habituate.
• AuDHD — a non-additive signature: slow recovery (autism inertia) AND erratic recovery (ADHD gain noise), distinct from the average of the two.

Full literature review, with the competing hypotheses and the framework-level critique of the Bayesian-brain paradigm itself, in docs/THEORY.md and the Wiki.

Architecture

            core/  (Rust — the inference math, fast, no GC pauses)
   kalman.rs · hgf.rs · td_learning.rs · forward_model.rs · error.rs
         │ PyO3 (akribia._core)              │ rlib (direct link)
         ▼                                   ▼
   akribia/ (Python orchestration)     gui/src-tauri (Tauri/Rust)
   profiles · tasks · viz · validation  React + Tailwind + shadcn/Radix
         │ every task emits schemas/task_result.json
         ▼
   viz (plots) · validation (per-parameter recovery) · GUI (display)

The Rust core is the numerical engine; the Python layer is the research surface (notebooks, sweeps, CI-enforced validation); the optional Tauri GUI links the same Rust crate directly for live, interactive exploration. The pure-Python fallback core mirrors the Rust math so the project runs with or without a Rust toolchain. See docs/architecture.md for the ADR log and rationale.

Installation / dev environment

One command: open the repo in VS Code Dev Containers / GitHub Codespaces and the devcontainer builds the Rust+Python toolchain and runs maturin develop automatically. See CONTRIBUTING.md for manual setup and the extension points (adding a profile/task). pre-commit install runs the same checks CI runs.

Validation & benchmarks (honestly reported)

• Per-parameter recovery (docs/LIMITATIONS.md): discount_factor and prior_precision_cap recover cleanly (corr ≈ 1.0); precision_flexibility is weakly identified (corr ≈ 0.52) and reported as such. The CI gate is defined only on the reliably-recoverable parameters.
• Two independent core implementations (Rust + Python) agree to ~1e-16.
• Performance (docs/BENCHMARKS.md): the boundary-free Rust Kalman core is ~5.6× faster than pure Python (measured, not asserted), with an honest note about PyO3 boundary overhead on single calls.

Related work

• TAPAS (Mathys et al.) — the reference HGF/computational-psychiatry toolbox. akribia's contribution is the cross-condition profile framework (autism/ADHD/PPCS under one engine) with a comorbidity mode, not a novel filtering algorithm. TAPAS is GPL and is compared against, never linked or copied from (see docs/architecture.md §5).
• PyHGF / pymdp — Python-native HGF / active-inference libraries; viable reference oracles.

Accessibility (spec 9)

Plots use the Okabe-Ito colorblind-safe palette and pair colour with distinct line styles/markers (never colour alone), so figures read in grayscale and colorblind vision. Theory pages open with a plain-language paragraph before the math; jargon is defined in docs/GLOSSARY.md. The Tauri GUI is built on Radix UI, whose ARIA compliance is real accessibility infrastructure.

Roadmap

Predictive coding's reach extends well past these three conditions. Each slots into the same profile architecture (a new profiles/<condition>_<mechanism>.py + a literature-grounded parameterization + a demonstrating task): schizophrenia (aberrant precision in hierarchical message passing), anxiety (overestimated threat precision), depression (biased reward valuation), addiction (pathological cue RPE). akribia "happens to start with three conditions relevant to the author," not "models the author."

Links

• Wiki — theory deep-dive & FAQ · docs/THEORY.md
• docs/LIMITATIONS.md · docs/architecture.md
• CONTRIBUTING.md · SECURITY.md · CHANGELOG.md
• CITATION.cff — GitHub renders a "Cite this repository" button.

GUI installers are unsigned. macOS Gatekeeper / Windows SmartScreen will block them by default. macOS: right-click → Open, or xattr -d com.apple.quarantine <file>. Windows: "More info" → "Run anyway".

License

MIT © Londopy. akribia (ἀκρίβεια) — exactness, precision.