graziul/tokamark-artifacts

TokaMark Working Artifacts

Documentation and tooling supporting a contribution to the TokaMark benchmark: diagnostics, errata, specifications, a protocol-matched metric suite, and the governance that keeps them honest.

Version: 0.6

STATE UPDATE 2026-10-11 (v0.6) - the front door. This version adds the START HERE path and the doc map, and retires the hand-maintained version table (it had drifted - the X87 class; 8 of 13 rows were stale). Every document version now lives in STATUS.md, which is GENERATED from the ledger - never copied into a table again. Also corrected in this pass: STAGE-1 has PASSED (our harness reproduces the shipped Mean baseline: NRMSE 0.92713 vs published 0.9267), the windower commit is PINNED (Q6, 1a200f45...294), the lcurve/ split is an artifact (lcurve/split.json; X50 closed), and the magnetics chi^2 gate SEPARATES at full budget (T-9: 15 shots / 199 frames).

Status: working draft. Research notes, not peer-reviewed claims. Every measured quantity is tagged and traceable. Licence: CC BY-SA 4.0, matching the upstream provenance (X9). Canonical ledgers: errata and open questions - governance.md 5 - results - results.md (RES-##) - tasks - tasks.md (T-##) - identifier glossary - tasks.md 2.

Read this first

Four headline claims this set previously carried are withdrawn, downgraded or reversed. They are listed here because a reader who finds a stale copy needs to know which numbers are gone.

1. "Error is core-concentrated by 12.47x" - WITHDRAWN. No candidate mask reproduces it from the stored residual map, and the mask was never recorded. The binned profile is ~2.06x raw and ~1.20x after normalising by local sigma, which is U-shaped and worst at the extreme edge. (X30, X31) [X30]
2. "TokaMind halves the CNN." False. The CNN's Task 1-3 value is 0.1481, not 0.255 - a number appearing in neither paper. The gain is 17.6%. (X18)
3. "There is no streaming path." False. The toolkit documents S3 + fsspec and ships an IterableDataset. What is missing is Hub presentation. (X22)
4. "The truncation floor is ~0.0427 at k=7." The correction was the error: at k=7 the value is 0.0913, and 0.0427 belongs to k~9-11. (X40) [X40]

Every correction is in governance.md 5 with its evidence. The ledger is append-only - it records what we believed and why it was wrong.

---

START HERE

If you are new to this repository, read in this order (plain language, no tools needed):

1. this file - what this is and what is in it
2. roadmap.md sections 0-2 - the orientation table, what the dataset is, the task map
3. datasheet.md sections 1 and 3 - what is in the data, and the code book
4. results.md - the answers ledger (RES-##), negatives included
5. W1-MEMO.md - how to think about this data, in plain words, written for a newcomer

Want to check our work instead of reading prose? See How to run the checks below: one command, and n_red = 0 means every gate in this repository is green.

The doc map

A view, not a source: versions live in STATUS.md (GENERATED from the ledger); ownership and identifier series live in governance.md 4.5 and tasks.md 2.

doc	what it is	audience
roadmap.md	orientation and entry point	everyone
datasheet.md	the data: code book, task contracts, descriptives, corrected baselines	everyone
results.md	the results ledger (RES-##), each row with its counter-argument and falsifier	everyone
errata.md	upstream-facing corrections and undocumented data properties	everyone
W1-MEMO.md	interpretation: what the diagnostics license, and how to hold the data in your head	everyone
evaluation_protocol.md	the upstream metric, read from source and implemented (M1-M6)	working
spec_dataset.md	the critical path: provenance lock, frozen split, metric module, harness, card sample (D1-D6)	working
SPEC-P1.md, SPEC-P2.md	the two model specs: grey-box response model; Grad-Shafranov-structured decoder	working
governance.md	rules, the append-only errata ledger (X##), open questions (Q##) - canonical	working
hygiene.md	the data-hygiene register (H##/HG##): what to check, in what order, at what cost	working
tasks.md	the live task register (T-##) and the identifier-series index	working
dev_roadmap.md	gates, workstreams, cost ledger, risk register	working
handoff.md	session-to-session state transfer	working
STATUS.md, SESSION-BRIEF.md	GENERATED doc registry and continuation brief - never hand-edit	generated
friction.md, d5_filing.md, HF-SCIENCE.md	upstream asks - filed by the owner (the agent has no GitHub access)	owner
tools/docs/	working plans and ledgers (next.md, critical_paths.md, refactor.md, organization_guards.md, sop.md)	working
tools/	the code: tokamark_core (the one core), thin analysis tools, the health gate	working

Why this exists

TokaMark is a 14-task benchmark over 11,573 MAST shots (58 archives, 565 GB).

1. Presentation. The Hub viewer cannot render the upstream archive (0 configs, 0 rows). A documented S3 path exists (X22), but it is invisible from the card - so we built a small, renderable card sample (D5) and a provenance-locked split artifact (D2).
2. Evaluation - the part that survived. The flux-map task (1-3) has no physics-consistency metric, and its aggregate NRMSE is tail-dominated and variance-dominated: the top 1% of windows carry 47.9% of test MSE (X36), and corr(per-pixel RMSE, per-pixel target sigma) = 0.970 (X31). A single NRMSE largely reports the field's own variance and which shots were drawn. This is an argument about metric design, and it is the strongest thing this project has.
3. Physics consistency. The magnetics chi^2 gate (T-9) scores whether a predicted flux map explains the measured magnetics - the one thing aggregate pixel error cannot see. At full budget it separates the reference from every trivial arm and from the coil-only vacuum field (stage2/p6_gate.json).

Our harness reproduces the shipped baselines end-to-end (STAGE-1 passed 2026-10-03), so every comparison we make is protocol-matched or labelled a pilot.

What was found

Finding	Value	Confidence
psi target is a low-dimensional family	corpus median k99 = 5, median PR 1.222; batch 0 recomputed k99 = 7, PR 1.82499	[M] corpus-wide
The oracle rank-7 floor and the gap to it	floor 0.0913 at k=7 (sigma_meas); fitted ridge 0.838 -> a 9.2x gap. The bound uses true coefficients, so it is not a target - it says the error lives in predicting ~7 coefficients	[M] batch 0
NRMSE is tail-dominated	top 1% of windows -> 47.9% of test MSE	[M]
NRMSE is variance-dominated	corr(RMSE, local sigma) = 0.970; normalised spread 1.20x, U-shaped, worst at the edge	[M]
Error in derived quantities is EDGE-concentrated	flux-surface displacement 0.024 -> 0.687 cells as t goes 0.1 -> 0.9; core:edge 0.035	[M] pilot
Energy and geometry disagree in sign	median: elementwise energy 0.43-0.54x (better) vs level-set displacement 1.17-1.47x worse. Only the median phrasing is quotable	[M] pilot
The magnetics chi^2 gate SEPARATES	full budget, 15 shots / 199 frames: reference chi^2 2.424; every trivial arm 7.3-29.5x worse; the C0 reconstruction 2.58x the reference and closer than a constant	[M] stage2/p6_gate.json (live)
EFIT frames entirely NaN	29.4% mean (19.6-66.7%)	[M] corpus-wide
Dead input channels	1 of 115 dead in every batch; the rest batch-specific. No corpus-wide mask exists	[M] corpus-wide
Effective input dimension	batch-dependent: 4,415 (batch 0) vs 5,490 (12-batch subset)	[M]
An NRMSE is identified by its sigma AND its row set	official random-split sigma 0.054930705723548384; our pooled caches differ by row set (X91) - never quote an NRMSE without both	[M]
psi units	Wb / rad, in the zarr attrs but not TokaMark's stats file	[M]
Licence	card CC BY 4.0; data attrs CC BY-SA 4.0	[M]
Unscored signals present	j_tor (65,65,T); vloop_dynamic/vloop_static - unselected, not absent	[M]
Strongest baseline, Task 1-3	TokaMind FT-Base 0.122, 6,927,799 params, random split, 9,270 shots. CNN 0.1481. FT-Tiny ties 0.122	[P] from source
Classical baseline is per-task	CNN+LSTM beats CNN on 9 of 14; on 2-3 the bar is 0.1263	[P]
Split matters more than model family	1-3 goes 0.1481 -> 0.5016 (CNN), random -> temporal	[P]

Tags: [M] measured - [R] repo - [P] paper - [I] inference (carries a falsifier) - [U] unverified.

Scope caveat. Availability, dead-channel and manifold figures are corpus-wide. The region/estimator figures rest on a 122-shot pilot (X2, X35) and are labelled as such - [M] for that slice, never corpus-level claims. Batch 0 is the 1.7th percentile on Thomson availability and the 98.3rd on manifold width.

How to run the checks

This repository audits itself. The audit is one command (Python 3.10+; installs nothing beyond huggingface_hub, numpy, scipy):

from huggingface_hub import snapshot_download
import sys
root = snapshot_download("graziul/tokamark-artifacts", repo_type="dataset",
                         allow_patterns=["tools/**", "*.md"])
sys.path.insert(0, root + "/tools")
import tokamark_bootstrap as B
B.setup(root, need=("health",), verbose=True)   # raises, naming what is missing
import health
rep = health.main(upload=False)
print(rep["n_red"])   # MUST be 0: all 14 checks green

The 14 checks include: every module compiles and passes its own self-test (96 modules), every file name conforms to the naming rule (L4), every doc's header agrees with its machine-checked marker (L5 - the staleness kill-switch), every cited artifact exists, generated docs equal a fresh render, and the coordination state machine is licensed. health/health.json in the runs repo holds the last uploaded report.

For contributors

Five rules are already enforced, not stated: a new tool imports tokamark_core and defines self_test(); a new metric gets a metrics_index.json entry and one implementation; a new root document is registered in the ledger in the same commit; module edits ship through commit_verified (staged, compiled, self-tested, refused otherwise) and doc edits through patch() (live read, unique anchors); verification is a fresh process. The one-page recipe is in tools/analysis/README.md; the constraint list with each guard is tools/docs/organization_guards.md.

Evidence base

Measured quantities live in graziul/tokamark-p2-runs. Highlights: w1/ (structure, units, licence attrs, the batch-0 target spectrum), sweep/ (corpus-wide availability and dead-channel records), q11/-q15/ (geometry, sigma, tail, spectrum, estimator 2x2), stage1/ and stage2/ (the harness reproduction, the magnetics gate, the metric falsifiers), records/ (the machine-readable RES-## records), metrics_index.json (the metric suite, made runnable), lcurve/ (the 12-shard windowed cache).

lcurve/ caveats before citing: the split is lcurve/split.json (held-out shards 2 and 9); Y is present for the held-out shards (it is a training bundle, not a held-out evaluation set); these windows predate the windower pin, so they are ours, not upstream's; per-shard window counts spread 1.9x; lcurve_fits.json does not exist (the producing job was cancelled).

Upstream

• Dataset: https://huggingface.co/datasets/UKAEA-IBM-STFC/tokamark-v1
• Toolkit (task YAMLs, splits CSV, evaluator, metadata): https://github.com/UKAEA-IBM-STFC-Fusion-FMs/tokamark
• Baseline (ships no checkpoints or predictions): https://github.com/UKAEA-IBM-STFC-Fusion-FMs/tokamark_baseline
• Model: https://huggingface.co/UKAEA-IBM-STFC/tokamind-base-v2
• Paper: arXiv:2602.10132 - Foundation model: arXiv:2602.15084

Citation

Cite the upstream TokaMark and TokaMind papers. These are working notes on top of them, not a substitute.