partonmap

A working lab on the structure of the proton.

A working research lab · Bayesian proton structure

What's inside the proton?
Measured, not assumed.

Every LHC prediction stands on a map of the proton's interior that no one can calculate — it has to be inferred from four decades of collision data, and at today's precision the error bar on that map is the product. This lab fits the map with explicit Bayesian statistics, keeps every number traceable to a script, and publishes the failures next to the wins.

See the results → Library Study tutorials
u u d the proton: quarks, gluons, and a sea
15
registered results — plot · code · notes · provenance
6/6
Colibri validation table replicated
χ²/N = 1.00
MSHT20 on real data — the 5.6 floor solved
96
parity checks on the MSHT20 port

Running now: Maria's deliverable is DONE: Hessian vs MC vs Bayesian on MSHT20 real data — only the Bayesian bands survive. Full-DIS gate + sealed GPU fit in flight.

The journey — from toy to the real thing

Phase 1 · complete · 2 Jul

Prove the machinery

Never point an untested instrument at the sky. We replicated the Colibri paper's entire validation table — planted-proton drills clean and noisy, truth recovery at 100% of checkpoints, and all three inference engines agreeing where they must and splitting exactly where theory predicts (PPDF-1…PPDF-6).

Phase 2 · complete · 3 Jul

Meet real data — and get judged

First genuine Bayesian fits of SLAC + BCDMS measurements. The evidence referee disqualified a model for missing physics (ΔlogZ ≈ −800), preferred 13 honest parameters over 36 straitjacketed ones (ΔlogZ ≈ −109), and both models hit an unexplained χ²/N ≈ 5.6 floor — now a live question with Maria's group (PPDF-7…PPDF-9).

Phase 3 · in flight

The real-world parametrization: MSHT20

The 52-parameter form used across LHC phenomenology, ported to JAX and parity-verified against its reference implementation (96 checks — PPDF-10), now running its closure drills. Destination: the first Bayesian MSHT20 posterior, then Maria's goal — Hessian vs replicas vs Bayesian uncertainties on the same real data.

Phase 4 · next

Scale up, contribute back

Full DIS, then LHC and Tevatron datasets on GPU; evidence-refereed model ladders with fair priors; findings back into the HEP-PBSP collaboration.

Latest from the lab

8 Jul 2026 — Maria's email confirms the plan line-for-line and sharpens the goal: three-method uncertainty comparison on MSHT20. Port passed its 96-check parity gate the same night — including a bug only catchable by testing beyond publication values. Notes →

3 Jul 2026 — The evidence prices missing physics, not parameter count: our flavour-blind grid model lost ~800 units of logZ despite 11 extra parameters. The lesson →

How this lab works

Results are cards

Every claim is a numbered PPDF-n card with three tabs — Explanation, Code, Notes — plus plot, numbers, and provenance. IDs never change; failures stay published as lessons.

Trust is earned

Parity gates before fits, closure tests before real data, priors declared before results exist. The wiki's "Trust, earned" cluster is the house constitution.

Everything traces

The canonical ledger lives in the repo (RESULTS_LEDGER.md); the site is generated from it. Deep tutorials, concepts, papers and notes all live in the Library.

partonmap.com · a proton-structure research lab · Mukesh Bansal, with Maria Ubiali's HEP-PBSP toolkit built on Colibri · NNPDF · UltraNest · JAX