Temporal Dynamics Visualization Lens
Philosophical Mode: Temporal
Primary Question: "Are training dynamics shown clearly and honestly?"
Focus: Axis Scaling (linear vs log), Smoothing Disclosure, Epoch/Step Alignment,
Run Aggregation (mean + variance bands), Early-Stopping Markers,
Wall-Clock vs Step-Count X-Axis
Arguments
/autoskillit:vis-lens-temporal [context_path] [experiment_plan_path]
- context_path (optional positional arg 1) — Absolute path to a lens context file
containing IV/DV tables, H0/H1 hypotheses, controlled variables, and success criteria.
If provided, read this file before beginning analysis to obtain structured context.
If omitted, discover context by exploring the CWD.
- experiment_plan_path (optional positional arg 2) — Absolute path to the full
experiment plan. If provided, read for complete experimental methodology and design.
If omitted, locate the experiment plan by exploring the CWD.
When to Use
- Reviewing training curves, learning curves, or any metric-vs-step/epoch plots
- Checking whether x-axis units are consistent across compared runs
- Evaluating whether smoothing is disclosed and appropriate
- Planning multi-run aggregation with variance bands
- User invokes
/autoskillit:vis-lens-temporal
Critical Constraints
NEVER:
- Modify any source code files
- Do not litter the codebase with useless comments, TODO markers, or explanatory annotations — the skill output and diagram speak for themselves
- Create files outside
{{AUTOSKILLIT_TEMP}}/vis-lens-temporal/
- Omit the CRITICAL flag when n_seeds == 1 for training curves — single-seed variance is unquantifiable
- Apply smoothing without disclosing the smoothing window or method
- Mix epoch-count and step-count x-axes on the same multi-run comparison without alignment
ALWAYS:
CRITICAL: if n_seeds == 1 for any training curve, flag as CRITICAL — single-seed training curves cannot demonstrate stability or convergence robustness
Disclose smoothing: state the EMA α or window size in the figure caption or axis label
When comparing runs with different batch sizes or learning rate schedules, align on
wall-clock time OR total gradient steps (not raw epochs), and document the choice
Use log-scale y-axis when loss spans more than one order of magnitude
Mark early-stopping epoch/step as a vertical dashed line with label
BEFORE creating any diagram, LOAD the /autoskillit:mermaid skill using the Skill tool — this is MANDATORY
If the Skill tool cannot be used (disable-model-invocation) or refuses this invocation, do NOT proceed with diagram creation. Abort this step and omit the diagram from output.
Write output to {{AUTOSKILLIT_TEMP}}/vis-lens-temporal/vis_spec_temporal_{YYYY-MM-DD_HHMMSS}.md (relative to the current working directory)
After writing the file, emit the structured output token as literal plain text with no
markdown formatting on the token name (the adjudicator performs a regex match):
diagram_path = /absolute/path/to/{{AUTOSKILLIT_TEMP}}/vis-lens-temporal/vis_spec_temporal_{...}.md
Analysis Workflow
Step 0: Parse optional arguments
If positional arg 1 (context_path) is provided and the file exists, read it to obtain
IV/DV tables, H0/H1 hypotheses, controlled variables, and success criteria. If positional
arg 2 (experiment_plan_path) is provided and exists, read the experiment plan for full
methodology. Use this structured context as the foundation for Steps 1–4; skip the CWD
exploration for these fields if the context file supplies them.
Step 1: Inventory Training Curves
Scan experiment plan, context file, and codebase for:
Learning and Loss Curves
- Find all learning curves, loss curves, metric-vs-step plots
- Look for:
train_loss, val_loss, reward, accuracy_vs_epoch, loss_curve, plt.plot
Seed Count
- Find n_seeds for each training run
- Look for:
n_seeds, num_seeds, seeds, SEEDS, random_state, runs
Smoothing Calls
- Detect whether smoothing is applied and whether it is disclosed
- Look for:
smooth, ema, rolling_mean, gaussian_filter, savgol_filter, EMA α parameters
X-Axis Type
- Determine whether x-axis is epoch-count, step-count, or wall-clock time
- Look for:
epochs, steps, global_step, time_elapsed, x-axis labels
Early Stopping
- Detect early-stopping usage and whether it is marked on the plot
- Look for:
early_stopping, patience, best_epoch, EarlyStopping
Step 2: Determine Axis Scaling
For each loss or metric curve, check the range:
- If loss spans more than one order of magnitude (max/min > 10): recommend log-scale y-axis
- If loss is bounded (e.g., accuracy 0–1): linear scale is acceptable
- Document the recommendation with the detected range
Step 3: Alignment Check
For all multi-run comparisons:
- Verify that all compared runs use the same x-axis unit (epoch vs step vs time)
- Flag mismatches as WARNING: "Runs use mixed x-axis units — align on gradient steps or wall-clock time"
- Check batch size and learning rate schedule consistency across compared runs
Step 4: Emit yaml:figure-spec Blocks
For each figure, emit one yaml:figure-spec fenced block with the stat_overlay
variance band filled in. Then LOAD /autoskillit:mermaid and create a temporal flow
diagram showing x-axis unit → scaling choice → smoothing annotation → variance band → verdict.
Output Template
# Temporal Dynamics Spec: {System / Experiment Name}
**Lens:** Temporal Dynamics (Temporal)
**Question:** Are training dynamics shown clearly and honestly?
**Date:** {YYYY-MM-DD}
**Scope:** {What was analyzed}
**n_seeds detected:** {N}
## Temporal Audit Summary
| Figure | n_seeds | x_axis | y_scale | smoothing | early_stop_marked | Status |
|--------|---------|--------|---------|-----------|-------------------|--------|
| {fig-01} | 1 | epoch | linear | none | no | CRITICAL |
| {fig-02} | 5 | step | log | EMA α=0.9 | yes | OK |
## Figure Specs
```yaml
# yaml:figure-spec — canonical schema (spec_version: "1.0")
figure_id: "fig-02-loss-curve"
figure_title: "Training Loss vs Gradient Steps"
spec_version: "1.0"
chart_type: "line"
chart_type_fallback: "scatter"
perceptual_justification: "Log-scale y-axis spans 2 orders of magnitude; variance band shows run stability."
data_source: "results/loss_curves.csv"
data_mapping:
x: "global_step"
y: "train_loss"
color: "run_id"
size: ""
facet: ""
layout:
width_inches: 6.0
height_inches: 4.0
dpi: 300
stat_overlay:
type: "band"
measure: "CI95"
n_seeds: 5
annotations: ["log-scale y; EMA α=0.9 disclosed; early-stop at step 4200"]
anti_patterns: ["ap-missing-variance-band"]
palette: "okabe-ito"
format: "pdf"
target_dpi: 300
library: "matplotlib"
report_section: "Section 3 Training"
priority: "P1"
placement_tier: "main"
conflicts: []
metadata:
created_by: "vis-lens-temporal"
reviewed_by: ""
last_updated: "{YYYY-MM-DD}"
Temporal Dynamics Diagram
%%{init: {'flowchart': {'nodeSpacing': 50, 'rankSpacing': 60, 'curve': 'basis'}}}%%
flowchart TB
%% CLASS DEFINITIONS %%
classDef cli fill:#1a237e,stroke:#7986cb,stroke-width:2px,color:#fff;
classDef stateNode fill:#004d40,stroke:#4db6ac,stroke-width:2px,color:#fff;
classDef handler fill:#e65100,stroke:#ffb74d,stroke-width:2px,color:#fff;
classDef output fill:#00695c,stroke:#4db6ac,stroke-width:2px,color:#fff;
classDef detector fill:#b71c1c,stroke:#ef5350,stroke-width:2px,color:#fff;
subgraph XAxis ["X-AXIS UNIT"]
X1["epoch / step / wall-clock<br/>━━━━━━━━━━<br/>{alignment status}"]
end
subgraph Scale ["Y-AXIS SCALING"]
Y1["linear / log<br/>━━━━━━━━━━<br/>{loss range: {min}–{max}}"]
end
subgraph Smooth ["SMOOTHING"]
S1["none / EMA / rolling<br/>━━━━━━━━━━<br/>{α or window disclosed: yes/no}"]
end
subgraph Variance ["VARIANCE BAND"]
V1["CI95 / SD band<br/>━━━━━━━━━━<br/>n_seeds = {N}"]
end
subgraph Verdict ["VERDICT"]
VD1["{OK / WARNING / CRITICAL}<br/>━━━━━━━━━━<br/>{reason}"]
end
X1 --> Y1
Y1 --> S1
S1 --> V1
V1 --> VD1
class X1 stateNode;
class Y1 cli;
class S1 handler;
class V1 output;
class VD1 detector;
Color Legend:
| Color |
Category |
Description |
| Dark Teal |
X-Axis |
Unit choice and alignment status |
| Dark Blue |
Y-Axis |
Scaling decision based on loss range |
| Orange |
Smoothing |
Disclosure status |
| Teal |
Variance |
Band type and seed count |
| Red |
Verdict |
OK / WARNING / CRITICAL assessment |
---
## Pre-Diagram Checklist
Before creating the diagram, verify:
- [ ] LOADED `/autoskillit:mermaid` skill using the Skill tool
- [ ] Using ONLY classDef styles from the mermaid skill (no invented colors)
- [ ] Diagram will include a color legend table
- [ ] Every CRITICAL (n_seeds == 1) training curve is flagged
- [ ] Every smoothing call has its parameters disclosed in the figure spec
- [ ] Early-stopping markers are noted for all curves with early stopping
1---2name: vis-lens-temporal3categories: [vis-lens]4activate_deps: [mermaid]5description: "Create Temporal Dynamics visualization planning spec showing axis scaling (linear vs log), smoothing disclosure, epoch/step alignment, run aggregation (mean + variance bands), early-stopping markers, and wall-clock vs step-count x-axis. Temporal lens answering \"Are training dynamics shown clearly and honestly?\""6hooks:7 PreToolUse:8 - matcher: "*"9 hooks:10 - type: command11 command: "echo 'Temporal Lens - Analyzing training curve representation...'"12 once: true13metadata:14 source: https://skillmd.com/skills/talont-org/vis-lens-temporal15---1617# Temporal Dynamics Visualization Lens1819**Philosophical Mode:** Temporal20**Primary Question:** "Are training dynamics shown clearly and honestly?"21**Focus:** Axis Scaling (linear vs log), Smoothing Disclosure, Epoch/Step Alignment,22 Run Aggregation (mean + variance bands), Early-Stopping Markers,23 Wall-Clock vs Step-Count X-Axis2425## Arguments2627`/autoskillit:vis-lens-temporal [context_path] [experiment_plan_path]`2829- **context_path** (optional positional arg 1) — Absolute path to a lens context file30 containing IV/DV tables, H0/H1 hypotheses, controlled variables, and success criteria.31 If provided, read this file before beginning analysis to obtain structured context.32 If omitted, discover context by exploring the CWD.33- **experiment_plan_path** (optional positional arg 2) — Absolute path to the full34 experiment plan. If provided, read for complete experimental methodology and design.35 If omitted, locate the experiment plan by exploring the CWD.3637## When to Use3839- Reviewing training curves, learning curves, or any metric-vs-step/epoch plots40- Checking whether x-axis units are consistent across compared runs41- Evaluating whether smoothing is disclosed and appropriate42- Planning multi-run aggregation with variance bands43- User invokes `/autoskillit:vis-lens-temporal`4445## Critical Constraints4647**NEVER:**48- Modify any source code files49- Do not litter the codebase with useless comments, TODO markers, or explanatory annotations — the skill output and diagram speak for themselves50- Create files outside `{{AUTOSKILLIT_TEMP}}/vis-lens-temporal/`51- Omit the CRITICAL flag when n_seeds == 1 for training curves — single-seed variance is unquantifiable52- Apply smoothing without disclosing the smoothing window or method53- Mix epoch-count and step-count x-axes on the same multi-run comparison without alignment5455**ALWAYS:**56- CRITICAL: if `n_seeds == 1` for any training curve, flag as **CRITICAL** — single-seed training curves cannot demonstrate stability or convergence robustness57- Disclose smoothing: state the EMA α or window size in the figure caption or axis label58- When comparing runs with different batch sizes or learning rate schedules, align on59 wall-clock time OR total gradient steps (not raw epochs), and document the choice60- Use log-scale y-axis when loss spans more than one order of magnitude61- Mark early-stopping epoch/step as a vertical dashed line with label62- BEFORE creating any diagram, LOAD the `/autoskillit:mermaid` skill using the Skill tool — this is MANDATORY63- If the Skill tool cannot be used (disable-model-invocation) or refuses this invocation, do NOT proceed with diagram creation. Abort this step and omit the diagram from output.64- Write output to `{{AUTOSKILLIT_TEMP}}/vis-lens-temporal/vis_spec_temporal_{YYYY-MM-DD_HHMMSS}.md` (relative to the current working directory)65- After writing the file, emit the structured output token as **literal plain text** with no66 markdown formatting on the token name (the adjudicator performs a regex match):6768 ```69 diagram_path = /absolute/path/to/{{AUTOSKILLIT_TEMP}}/vis-lens-temporal/vis_spec_temporal_{...}.md70 ```7172---7374## Analysis Workflow7576### Step 0: Parse optional arguments7778If positional arg 1 (context_path) is provided and the file exists, read it to obtain79IV/DV tables, H0/H1 hypotheses, controlled variables, and success criteria. If positional80arg 2 (experiment_plan_path) is provided and exists, read the experiment plan for full81methodology. Use this structured context as the foundation for Steps 1–4; skip the CWD82exploration for these fields if the context file supplies them.8384### Step 1: Inventory Training Curves8586Scan experiment plan, context file, and codebase for:8788**Learning and Loss Curves**89- Find all learning curves, loss curves, metric-vs-step plots90- Look for: `train_loss`, `val_loss`, `reward`, `accuracy_vs_epoch`, `loss_curve`, `plt.plot`9192**Seed Count**93- Find n_seeds for each training run94- Look for: `n_seeds`, `num_seeds`, `seeds`, `SEEDS`, `random_state`, `runs`9596**Smoothing Calls**97- Detect whether smoothing is applied and whether it is disclosed98- Look for: `smooth`, `ema`, `rolling_mean`, `gaussian_filter`, `savgol_filter`, EMA α parameters99100**X-Axis Type**101- Determine whether x-axis is epoch-count, step-count, or wall-clock time102- Look for: `epochs`, `steps`, `global_step`, `time_elapsed`, x-axis labels103104**Early Stopping**105- Detect early-stopping usage and whether it is marked on the plot106- Look for: `early_stopping`, `patience`, `best_epoch`, `EarlyStopping`107108### Step 2: Determine Axis Scaling109110For each loss or metric curve, check the range:111- If loss spans more than one order of magnitude (max/min > 10): recommend log-scale y-axis112- If loss is bounded (e.g., accuracy 0–1): linear scale is acceptable113- Document the recommendation with the detected range114115### Step 3: Alignment Check116117For all multi-run comparisons:118- Verify that all compared runs use the same x-axis unit (epoch vs step vs time)119- Flag mismatches as WARNING: "Runs use mixed x-axis units — align on gradient steps or wall-clock time"120- Check batch size and learning rate schedule consistency across compared runs121122### Step 4: Emit yaml:figure-spec Blocks123124For each figure, emit one `yaml:figure-spec` fenced block with the `stat_overlay`125variance band filled in. Then LOAD `/autoskillit:mermaid` and create a temporal flow126diagram showing x-axis unit → scaling choice → smoothing annotation → variance band → verdict.127128---129130## Output Template131132```markdown133# Temporal Dynamics Spec: {System / Experiment Name}134135**Lens:** Temporal Dynamics (Temporal)136**Question:** Are training dynamics shown clearly and honestly?137**Date:** {YYYY-MM-DD}138**Scope:** {What was analyzed}139**n_seeds detected:** {N}140141## Temporal Audit Summary142143| Figure | n_seeds | x_axis | y_scale | smoothing | early_stop_marked | Status |144|--------|---------|--------|---------|-----------|-------------------|--------|145| {fig-01} | 1 | epoch | linear | none | no | CRITICAL |146| {fig-02} | 5 | step | log | EMA α=0.9 | yes | OK |147148## Figure Specs149150```yaml151# yaml:figure-spec — canonical schema (spec_version: "1.0")152figure_id: "fig-02-loss-curve"153figure_title: "Training Loss vs Gradient Steps"154spec_version: "1.0"155chart_type: "line"156chart_type_fallback: "scatter"157perceptual_justification: "Log-scale y-axis spans 2 orders of magnitude; variance band shows run stability."158data_source: "results/loss_curves.csv"159data_mapping:160 x: "global_step"161 y: "train_loss"162 color: "run_id"163 size: ""164 facet: ""165layout:166 width_inches: 6.0167 height_inches: 4.0168 dpi: 300169stat_overlay:170 type: "band"171 measure: "CI95"172 n_seeds: 5173annotations: ["log-scale y; EMA α=0.9 disclosed; early-stop at step 4200"]174anti_patterns: ["ap-missing-variance-band"]175palette: "okabe-ito"176format: "pdf"177target_dpi: 300178library: "matplotlib"179report_section: "Section 3 Training"180priority: "P1"181placement_tier: "main"182conflicts: []183metadata:184 created_by: "vis-lens-temporal"185 reviewed_by: ""186 last_updated: "{YYYY-MM-DD}"187```188189## Temporal Dynamics Diagram190191```mermaid192%%{init: {'flowchart': {'nodeSpacing': 50, 'rankSpacing': 60, 'curve': 'basis'}}}%%193flowchart TB194 %% CLASS DEFINITIONS %%195 classDef cli fill:#1a237e,stroke:#7986cb,stroke-width:2px,color:#fff;196 classDef stateNode fill:#004d40,stroke:#4db6ac,stroke-width:2px,color:#fff;197 classDef handler fill:#e65100,stroke:#ffb74d,stroke-width:2px,color:#fff;198 classDef output fill:#00695c,stroke:#4db6ac,stroke-width:2px,color:#fff;199 classDef detector fill:#b71c1c,stroke:#ef5350,stroke-width:2px,color:#fff;200201 subgraph XAxis ["X-AXIS UNIT"]202 X1["epoch / step / wall-clock<br/>━━━━━━━━━━<br/>{alignment status}"]203 end204205 subgraph Scale ["Y-AXIS SCALING"]206 Y1["linear / log<br/>━━━━━━━━━━<br/>{loss range: {min}–{max}}"]207 end208209 subgraph Smooth ["SMOOTHING"]210 S1["none / EMA / rolling<br/>━━━━━━━━━━<br/>{α or window disclosed: yes/no}"]211 end212213 subgraph Variance ["VARIANCE BAND"]214 V1["CI95 / SD band<br/>━━━━━━━━━━<br/>n_seeds = {N}"]215 end216217 subgraph Verdict ["VERDICT"]218 VD1["{OK / WARNING / CRITICAL}<br/>━━━━━━━━━━<br/>{reason}"]219 end220221 X1 --> Y1222 Y1 --> S1223 S1 --> V1224 V1 --> VD1225226 class X1 stateNode;227 class Y1 cli;228 class S1 handler;229 class V1 output;230 class VD1 detector;231```232233**Color Legend:**234| Color | Category | Description |235|-------|----------|-------------|236| Dark Teal | X-Axis | Unit choice and alignment status |237| Dark Blue | Y-Axis | Scaling decision based on loss range |238| Orange | Smoothing | Disclosure status |239| Teal | Variance | Band type and seed count |240| Red | Verdict | OK / WARNING / CRITICAL assessment |241```242243---244245## Pre-Diagram Checklist246247Before creating the diagram, verify:248249- [ ] LOADED `/autoskillit:mermaid` skill using the Skill tool250- [ ] Using ONLY classDef styles from the mermaid skill (no invented colors)251- [ ] Diagram will include a color legend table252- [ ] Every CRITICAL (n_seeds == 1) training curve is flagged253- [ ] Every smoothing call has its parameters disclosed in the figure spec254- [ ] Early-stopping markers are noted for all curves with early stopping