docs: record probsie refactor and chart direction

README.md

@@ -4,6 +4,11 @@
 
 ## probably a mini oracle. definitely a mini chatbot
 
+> **Current refactor:** Probaboracle is in a clean-baseline refactor. The
+> Beta `6.0` pulse is preserved as a diagnostic snapshot, not active baseline
+> proof. The next comparable evidence gate is a new fixed-prompt `eval-pulse`
+> from the proper-config candidate.
+
 Probaboracle is a small, local, agent-backed CLI mini chatbot using the **[Polinko research model](https://github.com/tryskian/polinko)**.
 
 It only accepts four question types:
@@ -63,6 +68,24 @@ healthy.
   - the clean-baseline candidate keeps the same pulse method before the next
     beta-boundary decision
 
+## Data Viz Direction
+
+Probsie charts follow the eval shape first. Shared chart families are useful
+only when the data shape naturally matches.
+
+The initial visualisation set is:
+
+- pulse charts:
+  - stacked horizontal bars for `anchor`, `counted_seam`, and `excluded_noise`
+  - grouped or faceted pulse comparison for snapshot versus clean baseline
+- detail table:
+  - row id, prompt, output, pulse label, reason, and seam note below the chart
+- row and lens charts:
+  - row-level `pass / fail / pending` stack by prompt type
+  - prompt-by-lens table heatmap
+  - fail-family horizontal bars
+  - correction slope only when true before/after pairs exist
+
 ## Run It
 
 ```sh
@@ -91,6 +114,10 @@ make check
   - beta map and research reading path
 - [docs/research/070_CB-CLEAN_BASELINE_RESET.md](./docs/research/070_CB-CLEAN_BASELINE_RESET.md)
   - current reset boundary, docs cleanup, and first local pulse plan
+- [docs/diagrams/EVAL_CHART.md](./docs/diagrams/EVAL_CHART.md)
+  - current static eval chart contract
+- [docs/diagrams/PIPELINE.md](./docs/diagrams/PIPELINE.md)
+  - public generation and eval-shape diagrams
 - [docs/runtime/templates/README.md](./docs/runtime/templates/README.md)
   - public templates for future research docs and pulse reports
 - [docs/governance/DECISIONS.md](./docs/governance/DECISIONS.md)

docs/governance/DECISIONS.md

@@ -1181,3 +1181,35 @@ If a decision crosses layers, say so plainly instead of flattening the method in
   scratch material or older row-level wording. Putting the templates under
   runtime docs keeps the process discoverable while preserving the public
   research lane for actual findings.
+
+## D-062: Probsie chart types follow eval shape first
+
+- Date: `2026-05-25`
+- Category: `eval_quality`, `workflow_environment`
+- Tags: `data_viz`, `observable_plot`, `pulse_method`, `chart_types`
+- Provenance: `human-led visualisation planning with Codex chart-type review`
+- Decision:
+  - choose chart types from Probsie's eval data shape before aligning with
+    neighbouring toy repos
+  - use shared chart families only when the data shape naturally matches:
+    - bars for counts
+    - stacked bars for part-to-whole eval labels
+    - table heatmaps for matrix-style lens/status reads
+    - slope charts only for real before/after correction pairs
+  - keep the initial Probsie chart set focused on:
+    - `eval-pulse-stack`: stacked horizontal bar chart
+    - `pulse-comparison`: grouped or faceted stacked horizontal bars
+    - `eval-detail-table`: detail table below the chart
+    - `row-verdict-stack`: stacked bar chart by prompt type
+    - `lens-table-heatmap`: table heatmap
+    - `fail-family-bars`: horizontal bar chart
+    - `correction-slope`: optional slope chart when correction pairs exist
+  - treat small multiples as a layout choice inside comparison views, not as a
+    standalone chart family
+  - treat lollipop styling as an option inside `fail-family-bars`, not as a
+    separate chart family
+  - keep Sankey-style method storytelling out of the initial eval-data set
+- Why: Probsie's data is primarily an eval-method surface: row-level verdicts,
+  sidecar lenses, fixed-prompt pulse labels, and pulse-level verdicts. The
+  chart set should make that method legible without forcing another toy's
+  shape onto Probsie.