{
 "about": "Deviations from the frozen Phase 3 plan and other decisions taken during Phase 4, each with its effect on the results and when it was taken relative to the results it could touch. None changes a confirmatory estimator, population, threshold or rule; the four frozen files are asserted unchanged by every Phase 4 script.",
 "items": [
  {
   "id": "DV1",
   "kind": "correction of a Phase 3 count",
   "what": "The plan says V2 drops the runs whose remark records an early stop (9 laboratory runs). The Phase 3 rule matched the substring 'early stop', which also caught the remark 'Tool breakage but no early stopped' (imi_vmx30ui run 17). Phase 4 uses remarks that start with 'Early stopped': 8 laboratory runs, 7 of them on imi_vm20i.",
   "effect": "none on any result: V2 runs on imi_vm20i, whose set of early-stopped runs is the same under both rules (asserted in r01_confirm.py)",
   "when": "before the confirmatory run (r01_confirm.py asserts the equality before computing)",
   "tokens": [
    "facts_early_runs_vm20i",
    "facts_early_runs_lab"
   ]
  },
  {
   "id": "DV2",
   "kind": "build environment",
   "what": "The plan states that EPS figures are converted by epstopdf under pdflatex and that this path was checked in this workspace. Ghostscript and the epstopdf script were in fact missing; the Phase 3 check tested only that pdflatex existed. Phase 4 installed ghostscript 10.02.1 and texlive-font-utils from the Ubuntu archive.",
   "effect": "none on the analysis; the build path of the plan now works",
   "when": "before the first manuscript build"
  },
  {
   "id": "DV3",
   "kind": "build environment",
   "what": "The template ZIP could not be downloaded from this environment (the egress policy refused the Springer Nature file server). Adam downloaded the same package from the same page and link and placed it in the connected folder; its size equals the printed 880.68 KB, and its sha256 is recorded in inputs/sources4.json.",
   "effect": "none",
   "when": "start of Phase 4"
  },
  {
   "id": "DV4",
   "kind": "figure style",
   "what": "The guidelines recommend Helvetica or Arial lettering. Neither font is installed; the figures use Liberation Sans, which is metrically identical to Arial, embedded in every EPS file (Type 42).",
   "effect": "lettering shape only",
   "when": "figure build"
  },
  {
   "id": "DV5",
   "kind": "figure numbering",
   "what": "Plan Fig. 2 (clipping distributions) and plan Fig. 3 (example spectra) swap numbers, because the Methods cite the spectra before the Results cite the clipping figure and the guidelines require figures to be cited in consecutive order. Contents, sizes and the registered heights (fig2_height_mm for the clipping figure, fig3_height_mm for the spectra) are unchanged; the values that record the spectra choice keep their plan names (fig3_*).",
   "effect": "numbering only",
   "when": "manuscript drafting"
  },
  {
   "id": "DV6",
   "kind": "figure presentation",
   "what": "Fig. 4 shows the AUC on an axis logarithmic in one minus the AUC, with AUC tick labels, so that differences near 1 remain visible next to the design-variable models.",
   "effect": "presentation only",
   "when": "figure build, after the AUCs were known"
  },
  {
   "id": "DV7",
   "kind": "exploratory scope made explicit",
   "what": "E4 ('every feature set') was computed for the design-variable models too, including the model without tool family under the family-held-out scheme; E7 (materials held out) was computed for the high-rate, MTConnect, no-Sound0 and clipping sets. Cells where a held-out family or material could enter only through a term never seen in training are marked not defined.",
   "effect": "adds exploratory cells to Table 3; no confirmatory quantity involved",
   "when": "before the exploratory run"
  },
  {
   "id": "DV8",
   "kind": "exploratory presentation choice",
   "what": "E1 describes how the Sound0 level flattens by the median RMS in bins of the full-scale share with edges at d9_threshold, 10 percent and 30 percent (e1_bin_edge_mid, e1_bin_edge_high). The two upper edges were chosen after the full-scale share quantiles by label had been printed in Phase 4.",
   "effect": "descriptive text only; no inference rests on the bins",
   "when": "during the exploratory run"
  },
  {
   "id": "DV9",
   "kind": "implementation detail left open by the plan",
   "what": "A1 recomputes every confirmatory statistic with each run left out, with the Holm step applied to the three predictions of each leave-one-run-out set; E8 fits the transferred MTConnect model on every labelled path of imi_vm20i; E9's baseline pairs the Acc0 line of each path with the Sound2 line of every other path of the same run among the paths where both lines are present.",
   "effect": "fixes details the plan did not spell out; stated in the Methods",
   "when": "before the respective runs"
  },
  {
   "id": "DV10",
   "kind": "reference style switch",
   "what": "Adam chose the journal's reference format (Q1). The template's sn-basic.bst sorts the list alphabetically by default; its header says to comment out the two SORT lines for a list numbered consecutively, which the guidelines ask for and which IJAMT's own article U1 shows. The submitted sn-basic.bst differs from the template's file in those two lines only (asserted by the check).",
   "effect": "reference list in order of first citation",
   "when": "first build"
  },
  {
   "id": "DV11",
   "kind": "figure sizes in the review PDF",
   "what": "The template's two-column layout has columns of 76 mm and a text width of 160 mm, narrower than the 84 mm and 174 mm of the published layout that the guidelines give. The EPS files keep the published sizes and 8 pt lettering; the manuscript scales them to the column or text width, so the review PDF shows the lettering slightly smaller.",
   "effect": "presentation in the review PDF only",
   "when": "first build"
  },
  {
   "id": "DV12",
   "kind": "title page",
   "what": "The class's ORCID command needs a logo file that the December 2024 package does not ship, so the ORCIDs from the pipeline are listed as text under 'ORCID identifiers' at the end of the manuscript and will be entered in the submission system. Departments are left out: Adam does not know them and the guidelines mark the department as optional.",
   "effect": "placement of the ORCIDs",
   "when": "drafting"
  },
  {
   "id": "DV13",
   "kind": "added analysis (post hoc, exploratory)",
   "what": "AUC of the Sound0 full-scale share itself, with no model and no split, for labels 0 and 1 on imi_vm20i and on imi_vmx30ui, each with a family-stratified run-bootstrap interval (boot_b resamples, seed_analysis, predictions fixed by construction); no p-value.",
   "why": "Review item R1: the imi_vmx30ui clipping AUC of the Phase 2 screen (s1_auc_imi_vmx30ui_clip_run) came from an unweighted logistic model pooled over run folds, whose fold intercepts follow each fold's base rate; the draft used it as evidence that clipping does not transfer. A statistic without fitting describes the clipping relation on both machines without that artefact.",
   "effect": "Replaces the two sentences that cited the screen value; exploratory, enters no verdict.",
   "when": "after the independent review; decided before the value was computed"
  },
  {
   "id": "DV14",
   "kind": "added analysis (post hoc, uncertainty gauge)",
   "what": "Delete-one-run jackknife standard errors of h1_drop, h2_drop and h3_kappa from the refits already made for check A1 (outputs/a1_leave_one_run_out.csv, all 34 runs), set beside the standard deviation of the registered bootstrap draws, recomputed from the saved scores and indicators with the registered weight matrix.",
   "why": "Review item R3: the plan (M6) says that holding the fitted predictions fixed leaves out the variability of refitting but registered no gauge of its size.",
   "effect": "Reported in the Results and the Limitations as a rough gauge (a few runs dominate it); the registered classes are unchanged.",
   "when": "after the independent review; decided before the values were computed"
  },
  {
   "id": "DV15",
   "kind": "planned outputs missing from the first draft",
   "what": "E1 asked for clipping by machine, sound channel and label, E5 for a supplementary table of the off-harmonic shares by label, channel and setting, and E8 for clipping, the Sound2 off-harmonic shares and tool identity on imi_vmx30ui. The first draft printed Sound0 clipping by label on imi_vm20i only, compared the two machines on shares that included label 2 paths, had no E5 table, and left out the imi_vmx30ui Sound2 shares and tool identity. They are now in the text and in a new Online Resource 4 (ESM_4.csv), with the machine comparison restricted to labels 0 and 1.",
   "why": "Review items R4 and R11.",
   "effect": "Completes the planned exploratory outputs; the values were computed in Phase 2 (seen) or by r03_explore.py, except the per-label shares above 1 percent on imi_vmx30ui and for Sound1 and Sound2, which are counted from the same path table.",
   "when": "after the independent review"
  },
  {
   "id": "DV16",
   "kind": "added analysis (post hoc, exploratory)",
   "what": "H3 agreement against a null of independence within runs: the frozen q_common.kappa_null_w and kappa_max_w with the run as the stratum, the threshold rule at thr_kappa_frac, and a family-stratified run-bootstrap interval of the excess of kappa over that null; no p-value.",
   "why": "Review item R6: the registered null conditions on two spindle-speed classes and removes almost nothing (h3_kappa_null); speed within a class, depths of cut, material and amplitude stay uncontrolled. A null within runs holds every run-level setting fixed.",
   "effect": "Reported beside H3 as an exploratory check; the class of H3 does not depend on it.",
   "when": "after the independent review; decided before the value was computed"
  },
  {
   "id": "DV17",
   "kind": "added analysis (post hoc, descriptive)",
   "what": "Frequencies of the dominant off-harmonic lines of Acc0 and Sound2 on imi_vm20i: the quartiles among paths whose indicator is on, by label, and the Spearman rank correlation between the line frequency and the commanded spindle speed among label 1 paths; these summaries and no others. The per-path line frequencies are added to Online Resource 3.",
   "why": "Review item R6 and self-audit item SA4: lines that sit near one frequency at every spindle speed point to a structural mode of the machine, which chatter and forced excitation of that mode share.",
   "effect": "Descriptive; enters no verdict.",
   "when": "after the independent review; decided before the values were computed"
  },
  {
   "id": "DV18",
   "kind": "reporting",
   "what": "The manuscript calls the predictions pre-specified rather than pre-registered, and the title no longer says 'a pre-registered analysis'. The plan, the analysis functions, the estimators and the rule evaluator were frozen with SHA-256 digests before any confirmatory statistic was computed, but the digests are held by the authors and no third party holds a time-stamped copy, so a reader can check what was frozen but not when.",
   "why": "Review item R9.",
   "effect": "Wording of the title, abstract and Methods; the freeze itself is unchanged and described as it was done.",
   "when": "after the independent review"
  }
 ]
}