Prompt changes for call #1 #86

New Issue

2026-09-21T05:04:28Z

grokbot_ty commented

2026-09-21 05:04:28 +00:00

Data-populated prose is wrong

The text prompt sent to the model for Call 1 (species/wood/edge/figure) carries misleading vocabulary in the taxonomy section it populates from the synced Woo lists.

Problem

The prompt's 'SYNCED TAXONOMY (only legal choices)' block is keyed species_leaves / wood_category_leaves and the surrounding prose uses the word "leaves" to describe the selectable items. "Leaves" is internal tree-structure vocabulary (parent vs. leaf category) and doesn't belong in the user-facing / model-facing data description. The populated prose reads as if it's describing a tree, not a flat list of valid choices.

What it should be

Drop the "leaves" vocabulary entirely. Just present the synced lists as the legal choices (e.g. species / wood_categories / edge_terms / figure_terms).
General cleanup of the taxonomy prose: plain, neutral wording that names the lists and says only those ids/names are legal choices.
No functional change to which ids are offered (8 species, 9 wood categories, edge/figure terms) — wording only.

Scope

backend/inference.py: CALL1_SEED prompt text + _taxonomy_lists key names (species_leaves -> species, wood_category_leaves -> wood_categories) and the prompt's reference to them.
Keep the flat *_confidence schema and the grade exclusion (METADATA-40) as-is.

This is a prompt-wording cleanup, not a behavior change. Filed from live UAT review of the actual rendered prompt (issue #77 follow-on).

## Data-populated prose is wrong The text prompt sent to the model for Call 1 (species/wood/edge/figure) carries misleading vocabulary in the taxonomy section it populates from the synced Woo lists. ### Problem The prompt's 'SYNCED TAXONOMY (only legal choices)' block is keyed `species_leaves` / `wood_category_leaves` and the surrounding prose uses the word **"leaves"** to describe the selectable items. "Leaves" is internal tree-structure vocabulary (parent vs. leaf category) and doesn't belong in the user-facing / model-facing data description. The populated prose reads as if it's describing a tree, not a flat list of valid choices. ### What it should be - Drop the "leaves" vocabulary entirely. Just present the synced lists as the legal choices (e.g. `species` / `wood_categories` / `edge_terms` / `figure_terms`). - General cleanup of the taxonomy prose: plain, neutral wording that names the lists and says only those ids/names are legal choices. - No functional change to which ids are offered (8 species, 9 wood categories, edge/figure terms) — wording only. ### Scope - `backend/inference.py`: `CALL1_SEED` prompt text + `_taxonomy_lists` key names (`species_leaves` -> `species`, `wood_category_leaves` -> `wood_categories`) and the prompt's reference to them. - Keep the flat `*_confidence` schema and the grade exclusion (METADATA-40) as-is. This is a prompt-wording cleanup, not a behavior change. Filed from live UAT review of the actual rendered prompt (issue #77 follow-on).

Ty referenced this issue

2026-09-21 05:16:07 +00:00

inference_log must not store base64/binary image payloads #87

grokbot_ty referenced this issue from a commit

2026-09-21 05:54:42 +00:00

fix(inference): #86 tighten Call 1 prompt and closed lists

grokbot_ty commented

2026-09-21 05:56:18 +00:00

Fixed in 7287a55 on main.

Locked Call 1 expert seed (cats before edge; species=highest conf; mill-count edges; Wild for cookies/burl; multi figure; feature_tag_ids only; no grade / free-text bags)
Closed LISTS keys; feat- stripped for model display names
Gate + Review UI apply feature_tag_ids
Seed auto-activates when still seed-authored
Live UAT Call 1 returned new schema (feature_tag_ids, multi figure); gated species@0.6 correctly
Tests: 76 backend passed

UAT-verified on deploy stack. Closing.

Fixed in `7287a55` on main. - Locked Call 1 expert seed (cats before edge; species=highest conf; mill-count edges; Wild for cookies/burl; multi figure; `feature_tag_ids` only; no grade / free-text bags) - Closed `LISTS` keys; `feat-` stripped for model display names - Gate + Review UI apply `feature_tag_ids` - Seed auto-activates when still seed-authored - Live UAT Call 1 returned new schema (`feature_tag_ids`, multi figure); gated species@0.6 correctly - Tests: 76 backend passed UAT-verified on deploy stack. Closing.

grokbot_ty closed this issue

2026-09-21 05:56:18 +00:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: Ty_Tech/SlabUploader#86