Direction, meetings, decisions
Cross-cutting notes: meetings with Maria, phase decisions, open questions. The story of each individual result lives in that result's own Notes tab on the Results page.
8 Jul 2026 · decisionsDiscrepancy logged · scale-up green-lit
Same MSHT20 model, same 648 real DIS points, same theory and covariance — three standard uncertainty methods give incompatible error bands. Agreement only on the gluon at mid-x (MC vs Bayes widths within 1.8×, centrals <1σ). Off-constraint: MC replica bands 0.6–10× the Bayesian width with centrals up to ~7σ apart (initialization-dependent). Hessian: curvature at the minimum indefinite — 11 of 52 eigen-directions with ≈zero/negative curvature, bands 50–10⁷× Bayesian, unusable. Cause: 648 DIS points constrain only ~11–14 of 52 parameter directions. Open: whether full-DIS (~3,300 pts, running) or parameter freezing restores the approximate methods. Details: PPDF-14, PPDF-15.
Decisions (Mukesh, 8 Jul):
1 · Tier-2: GO — DY/W/Z fits (54 datasets, already prefetched). Discipline unchanged: closure gate
first, then real data. Queues after PPDF-17.
2 · Fitted charm: GO — the 61-parameter variant (free c⁺ at Q₀ = 1.65, theory 40000000, FPPDF
fitted-charm convention; parity re-check on the charm block before any fit).
3 · Upstream Colibri issue: HELD (public-facing).
8 Jul 2026 · phase completeMSHT20 phase done — discussion agenda for Maria
Everything from Maria's email is implemented and verified, ~30 hours end-to-end:
1 · MSHT20 model in Colibri — ported from FPPDF's implementation as she suggested, four
analytic sum rules, proven exact by a 96-check parity gate (PPDF-10).
2 · Closure at 52 parameters — clean drill 0.0074, noisy drill 0.979 ≈ 1; by-product: our DIS subset
constrains only ~11 of 52 directions (PPDF-11).
3 · First Bayesian MSHT20 posterior on real data — χ²/N = 1.002, logZ = −381.9 ± 0.3; resolves the old
χ²/N ≈ 5.6 puzzle as pure model rigidity (PPDF-12).
4 · Her final goal, the three-method comparison — Bayes ≈ MC where data constrains (gluon: widths within
2×, centrals <1σ); replicas initialization-shaped off-constraint; the Hessian breaks down outright
(indefinite curvature, ~38/52 flat directions)
(PPDF-13
PPDF-14
PPDF-15).
In flight tonight: full-DIS scale-up (19 datasets) — the L1 gate plus a sealed real-data fit on the Nebius L40S. Directly tests whether more data rescues the replica and Hessian methods.
Discussion points for the call: (a) the χ²/N floor resolution and what it implies for toy-model studies; (b) is the Hessian breakdown at this dataset:parametrization ratio worth writing up — and does MSHT-style parameter freezing restore it? (c) fair prior design for grid models on real data; (d) positivity settings for comparability with production fits; (e) fitted-charm variant: theory 40000000 vs 40001000 (Q₀ = 1 GeV); (f) the Colibri Hessian-path XLA issue (three compile-bound routes at 52 params) — worth an upstream issue/PR.
7 Jul 2026 · meetingMaria: graduate from the toy model → MSHT20 in Colibri
Direction for the next phase (decoded from meeting notes; her detailed email pending): our Les Houches parametrization is a toy — the real next step is implementing the MSHT20 parametrization (52 parameters, Chebyshev form; the shape used by one of the big-three global-fit groups) as a Colibri model. First pass on the same SLAC+BCDMS dataset for a direct evidence comparison against PPDF-7; then extend to LHC/Tevatron data.
Homework done same day: digested the brand-new FPPDF tool (arXiv:2602.07118) — MSHT-style Hessian fits on the NNPDF stack. It ships the exact parametrization spec and starting parameters we need, and it means a Bayesian MSHT20 posterior doesn't exist publicly: that's our natural contribution lane. Execution plan: JAX port of the parametrization → numerical parity test against FPPDF's own implementation → L0/L1 closure → Bayesian fit on the baseline → evidence verdict. 52-dim sampling likely needs the GPU box.
Questions carried into the discussion: the χ²/N ≈ 5.6 floor both our models hit on SLAC+BCDMS (global fits get ≈1 — cuts? deuteron corrections? theory settings?); fair prior design for grid models; whether grid models should carry sum rules.
3 Jul 2026 · phase wrapReal-data campaign complete — evidence referee works, verdict provisional
Three real-data fits done: baseline PPDF-7, the disqualified-but-instructive grid v1 PPDF-8 (two standing lessons: always configure the slice sampler above ~10 params; flavour masks zero FK columns), and the first genuine evidence comparison PPDF-9. Big takeaways: the evidence prices missing physics, not parameter count, and Bayes-factor verdicts inherit the prior — fair contests need designed priors. Goal 3 of the project (evidence-based model comparison on real data) demonstrated end-to-end.
2 Jul 2026 · phase wrapMachinery validated: full Table 3.1 replication in one day
Bootstrapped the project from Maria's references, built the local Colibri stack, and completed the entire validation ladder: clean drill PPDF-1, our own truth-in-band check PPDF-2, noisy drill PPDF-3, and the three-engine cross-checks PPDF-4 PPDF-5 PPDF-6. Decision locked: never point an untested instrument at real data — and by evening, the instrument had earned it. First real-data fit launched the same night.