FJ9.8 — Publication Composites

Date: 2026-08-23.

A figure must not only reproduce the right numbers. Its caption must reproduce the right interpretation.

STATUS

PASSED. FJ9.8a–9.8e closed. Four main figures build from live data objects, export as PDF/SVG/PNG from a fresh process, and every caption is validated against required and forbidden claims before the figure is drawn.

ENVIRONMENT

WSL Ubuntu-Baru, juliaup override 1.11.3
Pkg.test with ddm-ref + ddm-torch + MCTS.jl : 193 testsets, 148 758 assertions
publication build (fresh process)           : PUBLICATION_EXIT=0
                                              ENV_MODEL_CALLS=0
                                              PYTHON_MODULES=none
                                              MCTS_LOADED=false

FJ9.8a — INVENTORY BEFORE LAYOUT

19 artefacts, classified before any figure was designed, and probed for existence rather than assumed:

The full table is artifacts/fj9/publication/inventory.md.

All four FJ9.7 videos are SUPPLEMENTARY. A printed paper has to stand on its own, so nothing in the main body depends on being able to play a file. A test asserts every video is supplementary and every main-figure artefact is assigned to a numbered figure.

FJ9.8b — COMPOSITE, NOT SCREENSHOT ASSEMBLY

WorldScene / ProjectionScene / PolicySlice / SearchSnapshot /
EpisodeDiagnostics / progress bins
        ↓
PanelSpec (payload + provenance + semantics)
        ↓
PublicationComposite (panels + layout + caption + metadata + fingerprint)
        ↓
Makie -> PDF / SVG / PNG

PanelSpec.payload holds a live data object. Nothing loads a PNG. That is what makes the PDF and SVG genuinely vector, keeps one font stack and one axis style across all panels, and — the reason that matters most here — lets the caption be generated from the same objects the panels were drawn from.

Panels are placed by an explicit layout of (row, col, rowspan, colspan). The default is a two-column grid; figures whose panels differ in natural shape override it. Figure 4 gives its summary table a full-width short row, because a wide text table in a square cell is mostly whitespace.

FJ9.8c — THE FOUR MAIN FIGURES

Figure 1 — Model and representation

Latent world state; the 15-entry privileged projection; FJ10's per-component observability classification. The caption states that the projection is a privileged model state and not an observation, and reports that 6 of 15 components are sensor-estimable while 2 (sigma_stop, stop_hold_progress) are the agent's own memory with no physical counterpart.

Figure 2 — Policy structure and ambiguity

Recomputed live from the slice objects, not copied from FJ9.3's prose:

cells                     3 721
tied cells                1 174
decisive cells            2 547
distinct tabular states      23

The caption carries the three qualifiers that make the panel honest: feature space, not guaranteed reachable, and the fixed context (v = 0.1; tile = STRAIGHT; d_stop = none; sigma_stop = false; duck = NONE). Tied cells are marked with a hatch marker rather than a second colour scale, so the ambiguity survives greyscale.

Figure 3 — Search behaviour

Both snapshots come from the same frozen state, and figure_search refuses to build if the two state fingerprints differ — a test constructs a mismatched snapshot and asserts the ArgumentError.

state fingerprint       65ecc28dcdc8e0c5   (identical for both)

MCTS     7 root actions   mean 5.14   median 2   max 24
DPW     24 root actions   mean 1.46   median 1   max  3

The caption says both planners leave many candidates lightly explored, and that MCTS concentrates more evidence on its preferred discrete action while DPW spreads the same budget across many continuous actions. It does not say DPW is random, and "is random", "dpw is worse" and "optimal action" are all in that figure's forbidden list. The figure demonstrates evidence dilution; it does not license a claim about search quality.

Figure 4 — Episode behaviour and failure mechanisms

The frozen six-solver table; TD3's stagnation; DPW's deterioration; realised generative calls by episode progress.

Every number in the caption is computed from the payload at build time: full stop at decision 28, 123 of 150 decisions in the stop zone, mean speed 0.005 m/s, 1019 → 835 → 940 → 602 → 195. Selection rules reach the reader, not just the seeds:

td3     :first_stagnation            seed 1001
dpw@1k  :median_length_terminating   seed 1002
mcts@1k :median_return               seed 1018

The TD3 and DPW panels each use two stacked axes, m/s over m, with the stop-zone flag shaded across both. The first version put v, d_stop and sigma_stop on one axis — a regression of the FJ9.6 unit rule that flattened both real quantities against a flag pinned at 1.0.

FJ9.8d — CAPTIONS ARE CHECKED LIKE DATA

Each figure has a CaptionRule of required and forbidden phrases. render_composite validates before drawing and throws on failure: a figure whose caption fails its rule is a build failure, not a copy-editing note.

The forbidden list is the half that matters. FJ9.6 found a false sentence living beside correct numbers, and a rule that only required content could never have caught it. Figure 4 forbids "never reaches a stop sign" and requires "reaches the stop sign", "full stop", "never proceeds", "horizon" and "passed_stops = 0". It also forbids "combined score" and "overall ranking", carrying FJ8.4b's no-composite-ranking rule into the figure layer.

A test constructs a composite carrying the old false claim and asserts it is rejected, and asserts the exported figure4.caption.txt contains the corrected claim and not the old one.

FJ9.8e — EXPORT

Each also writes figureN.caption.txt with the caption and the full machine-readable provenance block.

[x] same composite renders fresh-process
[x] no Python                      PYTHON_MODULES=none
[x] no MCTS required               MCTS_LOADED=false
[x] no environment step            ENV_MODEL_CALLS=0, measured
[x] source fingerprints embedded   in metadata AND in the caption file
[x] text legible at column size    captions wrapped; panel fonts 7-10 pt
[x] no clipped titles/footers
[x] no overlapping legends
[x] all units visible              one axis per unit throughout
[x] raster panels high-resolution  px_per_unit = 4
[x] greyscale/print sanity

Greyscale. Colour is never the only channel. Action regions carry contour boundaries; tied cells carry a hatch marker; time series use black with distinct linestyles; the stop zone is a grey band; events are dashed rules with text labels; bin counts are printed as n=. Sequential surfaces use viridis, which is monotonic in luminance and therefore survives a greyscale print as a legible gradient.

THE TD3 WEIGHTS CACHE

Figure 2 needs the TD3 actor. Extracting it from the frozen .pt requires torch, which would put Python on the publication path. So the extraction is a separate provisioning step — tools/run_export_weights.sh, run once — which writes the checkpoint's exact arrays to artifacts/fj9/weights/{td3,sac}/. The publication build reads them with the native reader FJ7 validated layer by layer, and then asserts PYTHON_MODULES=none. The .pt checkpoints are read-only inputs and were not written to.

FILES

NEW DEFECTS

Four of my own, all in the composite renderer, all found by looking at the rendered output.

The worst was a regression of my own rule: Figure 4's episode panels put m/s, m and a 0/1 flag on one axis, exactly the mistake FJ9.6 spent a gate establishing must not happen. The flag sat at 1.0 and compressed both physical quantities into the bottom of the panel. Split into stacked per-unit axes with the flag shaded.

The caption ran off the right edge unwrapped, losing the end of the sentence — which for a caption means losing the clause that qualifies the claim. Makie does not reflow, so wrap_text now hard-wraps on word boundaries in the core.

The outer grid used a doubled column index left over from when colorbars lived in the outer layout; combined with content-driven column sizing, every figure collapsed into a narrow band with wide empty margins. Colorbars now live in each panel's own nested grid, and columns are sized explicitly.

Two smaller ones: SearchNode.action is deliberately Any (the snapshot is solver-neutral), so grouping root visits by action needed string(...) rather than a String key; and the action-plane panel read a action_value field that does not exist, instead of filtering on action isa DuckieAction.

SCIENTIFIC INTERPRETATION

The gate's own principle turned out to be testable, which was not obvious at the start. "The caption must reproduce the right interpretation" sounds like a style rule; implemented as required/forbidden phrase lists checked at build time, it is a constraint with the same standing as a numerical tolerance, and it fails the build the same way.

That matters because of what FJ9.6 found. The defect there was not a wrong number — every number was right — but a sentence that misdescribed them, and it survived because nothing in the pipeline treated prose as checkable. Figure 4 now cannot be built with that sentence in it, and cannot be built without the corrected one.

NEXT GATE

FJ9.9 — final reproducibility closure: fresh environment, full test run, every publication artefact re-rendered, fingerprints verified, no optional-backend leaks, documentation links checked, and the structured test reporter that has been carried as tooling debt since FJ9.0.