Corpus
Corpus Composition
Distribution, split balance, and the separability diagnostic rendered from committed results JSON.
Results cover the dev split only, with two configurations reported: hosted large and local. A single prompt version was used with no iteration against dev results. The held out test split has not been run. The hosted redacted run was attempted and did not complete.
Summary
- Threads
- 200
- Loops
- 273
- Zero-loop threads
- 40 (20.0% of threads)
- Dev / test
- 80 / 120 (40.0% / 60.0%)
- Loops per thread
- 1.36
- Loops per non-empty thread
- 1.71
Additional Findings
- Deadlines resolved to a date: 163/273 (59.7%).
- Off-thread counterparties: 11/273 delegation targets never sent a message.
Separability
- Balanced accuracy
- 0.52
- Permutation null mean
- 0.50
- Permutation null p95
- 0.55
- p-value
- 0.22
- Split
- dev
- Threads with / without loops
- 65 / 15
Residual features toward a loop
| Token | Weight |
|---|
| can | 1.34 |
|---|
| deta | 1.11 |
|---|
| one | 1.11 |
|---|
| send | 1.03 |
|---|
| tha | 0.93 |
|---|
| ok | 0.93 |
|---|
Residual features toward no loop
| Token | Weight |
|---|
| anything | -2.98 |
|---|
| been | -2.76 |
|---|
| looks | -2.76 |
|---|
| if | -2.62 |
|---|
| worth | -2.06 |
|---|
| saw | -2.06 |
|---|
Leak History
Real organisation names appeared in early threads.
Replaced them with roles and added the no-real-organisations rule.
Accepted directives were evidenced inconsistently.
Clarified that the acceptor's own commitment wins when it exists.
The negative-thread cue test was English-centric.
Moved from literal keywords to construction families across English, Hindi-English, and Tamil-English.
Deadline-certainty distribution shifted between batch 0 and batch 1.
Audited sampled loops, corrected 3 labels, and recorded the remaining shift as thread-writing drift.