BLOCKING: Pin confidence-threshold gating rules for Call 1 pre-population and review gate #9

New Issue

2026-09-01T22:09:15Z

Ty commented

2026-09-01 22:09:15 +00:00

Summary

The plan of record for Call 1 confidence handling is not captured in the docs at all. Without an explicit doc change, a merge of docs/refactor-v2 would carry forward the permissive behavior ("ready: values from Call 1, manual, or mix") and the plan of record would be lost. This issue is blocking for POC.

Plan of record (authoritative)

Above 0.7 (configurable threshold): Call 1 pre-populates species, wood categories, edge type, figure, grade, and maps character/inclusions/voids/checks to feat-* tags. All pre-populated values are mutable and overridable by the user in the review screen.
Below 0.7: Fields are not pre-populated (left empty). The user is gated before continuing and must manually select the mandatory fields — at minimum species and one figure attribute — before the flow can proceed to publishable/ready state.
Inference remains assist-only. Manual entry is always authoritative. The full suite must still pass with inference OFF.

Current doc state (gap)

AGENTS.md §6 mentions the 0.7 threshold (configurable) but only says it "lives in IMPL Phase 5 / PROMPTS." It does not state the above/below behavior.
PROMPTS.md Call 1 schema returns confidence scores but says nothing about gating or pre-population rules.
CONTENT-WOO.md §2.3 describes deterministic name-matching for categories/tags and says every assignment is overridable, but never states that Call 1 vision results are the primary pre-population source, nor the below-threshold empty + gate behavior.
ARCHITECTURE.md status machine says "ready: mandatory fields filled (manual and/or Call 1)" but does not define the gate.
Issue #8 covers making Call 1 the primary pre-population source; this issue covers the threshold-specific gating semantics.

Required doc changes (drive the change)

AGENTS.md §6 states the above/below-0.7 rules verbatim (or references a single canonical section).
PROMPTS.md Call 1 notes state: ≥ threshold → pre-populate (mutable); < threshold → empty + mandatory-field gate.
CONTENT-WOO.md §2.3 states Call 1 results are the primary source for species/categories/tags/attributes when confidence ≥ threshold, with manual override always available.
ARCHITECTURE.md status machine or clarifications note the review gate for below-threshold mandatory fields: species + ≥1 figure attribute required before ready/publishable.
No doc implies below-threshold fields are auto-filled or that inference gates publish.
IMPL-PLAN.md Phase 5 acceptance criteria include the gate behavior as testable.

Non-goals

Changing the 0.7 default or making it user-configurable in the UI (follow-up).
Graduated confidence tiers beyond the single threshold.
Implementing the gate in code (docs-only for this issue; implementation follows the docs).

## Summary The plan of record for Call 1 confidence handling is not captured in the docs at all. Without an explicit doc change, a merge of docs/refactor-v2 would carry forward the permissive behavior ("ready: values from Call 1, manual, or mix") and the plan of record would be lost. This issue is **blocking for POC**. ## Plan of record (authoritative) 1. **Above 0.7 (configurable threshold):** Call 1 pre-populates species, wood categories, edge type, figure, grade, and maps character/inclusions/voids/checks to `feat-*` tags. All pre-populated values are **mutable and overridable** by the user in the review screen. 2. **Below 0.7:** Fields are **not pre-populated** (left empty). The user is **gated before continuing** and must manually select the mandatory fields — at minimum **species** and **one figure attribute** — before the flow can proceed to publishable/ready state. 3. Inference remains **assist-only**. Manual entry is always authoritative. The full suite must still pass with inference OFF. ## Current doc state (gap) - `AGENTS.md` §6 mentions the 0.7 threshold (configurable) but only says it "lives in IMPL Phase 5 / PROMPTS." It does not state the above/below behavior. - `PROMPTS.md` Call 1 schema returns confidence scores but says nothing about gating or pre-population rules. - `CONTENT-WOO.md` §2.3 describes deterministic name-matching for categories/tags and says every assignment is overridable, but never states that Call 1 vision results are the primary pre-population source, nor the below-threshold empty + gate behavior. - `ARCHITECTURE.md` status machine says "ready: mandatory fields filled (manual and/or Call 1)" but does not define the gate. - Issue #8 covers making Call 1 the primary pre-population source; this issue covers the **threshold-specific gating semantics**. ## Required doc changes (drive the change) - [ ] `AGENTS.md` §6 states the above/below-0.7 rules verbatim (or references a single canonical section). - [ ] `PROMPTS.md` Call 1 notes state: ≥ threshold → pre-populate (mutable); < threshold → empty + mandatory-field gate. - [ ] `CONTENT-WOO.md` §2.3 states Call 1 results are the primary source for species/categories/tags/attributes when confidence ≥ threshold, with manual override always available. - [ ] `ARCHITECTURE.md` status machine or clarifications note the review gate for below-threshold mandatory fields: species + ≥1 figure attribute required before ready/publishable. - [ ] No doc implies below-threshold fields are auto-filled or that inference gates publish. - [ ] `IMPL-PLAN.md` Phase 5 acceptance criteria include the gate behavior as testable. ## Non-goals - Changing the 0.7 default or making it user-configurable in the UI (follow-up). - Graduated confidence tiers beyond the single threshold. - Implementing the gate in code (docs-only for this issue; implementation follows the docs).

Ty changed title from ~~Docs: Pin confidence-threshold gating rules for Call 1 pre-population and review gate~~ to BLOCKING: Pin confidence-threshold gating rules for Call 1 pre-population and review gate

2026-09-01 22:11:34 +00:00

Ty commented

2026-09-01 22:11:36 +00:00

Escalated to blocking per plan of record. Reconciliation check: the refactor branch does not capture the 0.7 gating rules, so a merge would lose them. This issue now owns driving the explicit doc change — see updated acceptance criteria.

Escalated to **blocking** per plan of record. Reconciliation check: the refactor branch does not capture the 0.7 gating rules, so a merge would lose them. This issue now owns driving the explicit doc change — see updated acceptance criteria.

Ty added this to the POC milestone 2026-09-01 23:49:48 +00:00

grokbot_ty referenced this issue from a commit

2026-09-02 00:29:12 +00:00

docs(#9): Call 1 confidence gate ≥0.7 prefill, <0.7 empty + mandatory gate

grokbot_ty referenced this issue from a commit

2026-09-02 02:40:41 +00:00

docs: split list-ready into capture, submit, and confidence-gate

Ty commented

2026-09-03 02:20:27 +00:00

The docs on refactor-v2 correctly state the confidence-threshold gating rules (above 0.7 pre-populates, below 0.7 empty + mandatory gate, inference assist-only). However, the same rule is repeated in 8 places (AGENTS, ARCHITECTURE, CONTENT-WOO, DATA-MODEL, IMPL-PLAN, OPENAPI, PROMPTS, UX). This needs a cleanup pass to consolidate into a single canonical owner with references elsewhere. The issue cannot close until the duplication is resolved. See issues #5 and #8 comments for the pattern.

Ty referenced this issue

2026-09-03 02:36:57 +00:00

Test plan: POC acceptance test cases for all milestone issues #19

Ty commented

2026-09-03 02:54:07 +00:00

Resolution: consolidated confidence gating into AGENTS.md §6 as canonical owner.

Changes across 8 files:

AGENTS.md §6 — full canonical rule: ≥ 0.7 → pre-populate (mutable), < 0.7 → empty + mandatory gate, assist-only, never gates publish.
AGENTS.md §7 — removed inline "Vision may prefill..." → reference §6.
CONTENT-WOO.md §2.3 — replaced full above/below rule → reference §6.
ARCHITECTURE.md §3.4 — changed reference from CONTENT-WOO → AGENTS.md §6.
ARCHITECTURE.md §6.3 — removed inline review gate → reference §6.
IMPL-PLAN.md Phase 5 — removed inline rule + acceptance criteria duplication → reference §6.
PROMPTS.md Call 1 notes — changed reference from CONTENT-WOO → AGENTS.md §6.
UX.md — replaced inline rule and failure path → reference §6.

Verification: grep for 0.7 across live docs returns only AGENTS.md §6 (canonical), OPENAPI schema example value (grade_confidence: 0.75), and a UX out-of-scope mention. No duplication of the rule remains.

Test cases noted on issue:

Phase 5 acceptance criteria reference AGENTS.md §6
≥ 0.7: fields pre-fill, editable, override sticks
< 0.7: fields stay empty, mandatory fields block ready
No doc/code path auto-fills below-threshold or treats inference as publish gate
Inference OFF: app works end-to-end, zero calls to inference endpoint

This issue is resolved. Ready to close.

**Resolution: consolidated confidence gating into `AGENTS.md` §6 as canonical owner.** Changes across 8 files: - **`AGENTS.md` §6** — full canonical rule: ≥ 0.7 → pre-populate (mutable), < 0.7 → empty + mandatory gate, assist-only, never gates publish. - **`AGENTS.md` §7** — removed inline "Vision may prefill..." → reference §6. - **`CONTENT-WOO.md` §2.3** — replaced full above/below rule → reference §6. - **`ARCHITECTURE.md` §3.4** — changed reference from CONTENT-WOO → AGENTS.md §6. - **`ARCHITECTURE.md` §6.3** — removed inline review gate → reference §6. - **`IMPL-PLAN.md` Phase 5** — removed inline rule + acceptance criteria duplication → reference §6. - **`PROMPTS.md` Call 1 notes** — changed reference from CONTENT-WOO → AGENTS.md §6. - **`UX.md`** — replaced inline rule and failure path → reference §6. **Verification:** grep for `0.7` across live docs returns only AGENTS.md §6 (canonical), OPENAPI schema example value (`grade_confidence: 0.75`), and a UX out-of-scope mention. No duplication of the rule remains. **Test cases noted on issue:** - Phase 5 acceptance criteria reference AGENTS.md §6 - ≥ 0.7: fields pre-fill, editable, override sticks - < 0.7: fields stay empty, mandatory fields block ready - No doc/code path auto-fills below-threshold or treats inference as publish gate - Inference OFF: app works end-to-end, zero calls to inference endpoint This issue is resolved. Ready to close.

Ty closed this issue

2026-09-03 02:54:09 +00:00

Ty reopened this issue

2026-09-03 02:58:32 +00:00

Ty closed this issue

2026-09-03 02:58:54 +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#9