| … | … | @@ -19,3 +19,10 @@ full table in its `summary.txt`. Control R0 PASS: all Test A arms reproduce `res |
| 19 | 19 | 2. **Per-occurrence labels fragment level 2.** C-JT has 49K span types vs 27K for exchange classes and 24K for |
| 20 | 20 | gold: the same phrase appears under more label sequences, so counts thin out and type recall drops. |
| 21 | 21 | 3. Better layer-1 class metrics (v2) did not carry into level 2. Gold classes with the old parse remain best. |
| 22 | ||
| 23 | --- | |
| 24 | ||
| 25 | **Correction (2026-10-18, after Test B):** the learned layer-1 parse scores word pairs symmetrically, so its arc | |
| 26 | directions are a tie-break artifact (76–94% of heads on the left vs 34% gold). The directional "complete assembly" | |
| 27 | (`comp`, HARD/SOFT) numbers with a learned parse inherit that bias. See `RESULTS-B.md`. Gold-tree controls and | |
| 28 | `conn` numbers are unaffected. |
| … | … | @@ -23,3 +23,10 @@ At 30K, recall / precision with the top 3K types, G: CNT 0.950 / 0.128, HARD 0.8 |
| 23 | 23 | ranking at 3K it covers 0.883). But it does not help find more of them than counting does. |
| 24 | 24 | 4. Caveat on the measure: R@B asks whether a phrase's *type* is in the searched set, not whether a phrase was |
| 25 | 25 | chosen. Choosing units (Test B) remains the open step, and that is where precision should matter. |
| 26 | ||
| 27 | --- | |
| 28 | ||
| 29 | **Correction (2026-10-18, after Test B):** the learned layer-1 parse scores word pairs symmetrically, so its arc | |
| 30 | directions are a tie-break artifact (76–94% of heads on the left vs 34% gold). The directional "complete assembly" | |
| 31 | (`comp`, HARD/SOFT) numbers with a learned parse inherit that bias. See `RESULTS-B.md`. Gold-tree controls and | |
| 32 | `conn` numbers are unaffected. |
| … | … | @@ -55,3 +55,10 @@ Run `results/1792340082` (harness `854d470`, 10 seeds, tiers 3K/10K/30K). Scored |
| 55 | 55 | - **Several parses per sentence** (Linas's convention): a span is a candidate if it is an assembly in any of |
| 56 | 56 | k parses. Tests whether the loss is parse noise that multiple linkages average out. |
| 57 | 57 | - Test B (choosing units) still needs a unit criterion. |
| 58 | ||
| 59 | --- | |
| 60 | ||
| 61 | **Correction (2026-10-18, after Test B):** the learned layer-1 parse scores word pairs symmetrically, so its arc | |
| 62 | directions are a tie-break artifact (76–94% of heads on the left vs 34% gold). The directional "complete assembly" | |
| 63 | (`comp`, HARD/SOFT) numbers with a learned parse inherit that bias. See `RESULTS-B.md`. Gold-tree controls and | |
| 64 | `conn` numbers are unaffected. |