F682 — Deep Coherence Corpus
F682 covers “Deep Coherence Corpus”. The canonical record classifies it as mixed, based on 7 source artifacts and 1 original source figure. The highlighted measurements are Correction Total = 50, Gate Sentence Rescue Used = No, and Family Counts Correction Feedback = 24.
Purpose and research question
F682 compared two Deep Generation coherence candidates on separate 300-prompt evaluations. The full candidate passed its recorded gate at 269/300 useful responses, while the narrow candidate failed at 203/300; neither candidate was promoted to the runtime.
Method
The full evaluation measured 300 prompts across open, follow-up, correction, conversation, and controlled groups and recorded 269 useful, 9 off-domain, 5 repetitive, and 0 pseudo-word outputs. The narrow evaluation used the same five group labels and 300-prompt size but recorded 203 useful, 47 off-domain, 3 repetitive, and 39 pseudo-word outputs. Training records identify separate full and narrow candidates; both record candidate_promoted=false and no checkpoint or frozen-state change.
Documentation basis
7 source artifacts and 1 phase-specific figure are recorded for this phase.
Evidence highlights
- Full candidate useful: 269/300
- Full candidate verdict: PASS
- Narrow candidate useful: 203/300
- Narrow candidate verdict: FAIL
- Candidates promoted: false
- Cold reload: not recorded
Recorded evidence
| Correction Total | 50 |
|---|---|
| Gate Sentence Rescue Used | No |
| Family Counts Correction Feedback | 24 |
| Gate Candidate Promoted | No |
| Families Correction Feedback | 10 |
| Correction Useful | 42 |
| Correction Off Domain | 0 |
| Correction Repetition | 0 |
Public evidence and source figures
Original source figure · 1600 × 900 px
These are original source figures from the existing public evidence set, not editorial HTML visualizations.
Result and interpretation
F682 remains Mixed. The full candidate's evaluation records PASS with a 0.8967 useful rate, while the narrow candidate records FAIL with a 0.6767 useful rate. No source declares one evaluation to supersede the other, no operational promotion occurred, and no complete phase-level terminal decision resolves the conflict.
Key results
- Correction Total
- 50
- Gate Sentence Rescue Used
- No
- Family Counts Correction Feedback
- 24
- Gate Candidate Promoted
- No
- Families Correction Feedback
- 10
- Correction Useful
- 42
- Correction Off Domain
- 0
- Correction Repetition
- 0
Evidence sources
Values on this page are reproduced from the archived phase record and the source artifacts below. No new benchmark was run to create this page.
- DEEP COHERENCE CORPUS 682 (JSON)
- DEEP GENERATION COHERENCE 682 NARROW TEST (JSON)
- DEEP GENERATION COHERENCE 682 NARROW TEST (MD)
- DEEP GENERATION COHERENCE 682 (JSON)
- DEEP GENERATION COHERENCE 682 (MD)
- DEEP GENERATION TRAIN 682 NARROW TEST (JSON)
- DEEP GENERATION TRAIN 682 (JSON)
Research topics
Language · Memory · Training · Self-learning