generation run 2: report

.gitattributes

@@ -48,3 +48,7 @@ generation/gen-qwen3b-s30-c20.json filter=lfs diff=lfs merge=lfs -text
 report/gen-qwen3b-s30-c20/figures/gen1-canon-violation.png filter=lfs diff=lfs merge=lfs -text
 report/gen-qwen3b-s30-c20/figures/gen2-final.png filter=lfs diff=lfs merge=lfs -text
 generation/gen-v3-hybrid.json filter=lfs diff=lfs merge=lfs -text
+report/gen-v3-hybrid/figures/gen2-final.png filter=lfs diff=lfs merge=lfs -text
+report/pdnc-qwen7b/figures/fig1-main-results.png filter=lfs diff=lfs merge=lfs -text
+report/pdnc-qwen7b/figures/fig1b-ladder.png filter=lfs diff=lfs merge=lfs -text
+report/pdnc-qwen7b/figures/fig3-paired-deltas.png filter=lfs diff=lfs merge=lfs -text

report/gen-v3-hybrid/generation.md

@@ -0,0 +1,76 @@
+# Generation: consistency at length
+
+30 stories × 6 memory conditions × 20 chapters, Qwen/Qwen2.5-3B-Instruct. Canon is stated in chapter 1 only; from chapter 2 each condition keeps whatever its memory carries. Violations are deterministic string tests against facts we planted, so no model or human judgement enters the measurement.
+
+## Means over stories
+
+| condition | n | on_premise | violation_rate | retention | restatement_rate | canon_capture | mean_first_violation | total_chars | prompt_tokens_last |
+|---|---|---|---|---|---|---|---|---|---|
+| base:none | 30 | 1.000 | 0.504 | 0.496 | 0.842 | 0.000 | 5.089 | 68,306 | 112 |
+| base:last-chapter | 30 | 1.000 | 0.308 | 0.692 | 0.838 | 0.000 | 2.665 | 76,320 | 1,009 |
+| base:full-context | 30 | 1.000 | 0.329 | 0.671 | 0.838 | 0.000 | 2.209 | 76,299 | 5,398 |
+| base:dsg-state | 30 | 0.998 | 0.588 | 0.412 | 0.871 | 0.342 | 4.127 | 69,623 | 357 |
+| base:dsg-hybrid | 30 | 1.000 | 0.379 | 0.621 | 0.838 | 0.487 | 2.350 | 75,937 | 1,194 |
+| base:dsg-hybrid-repair | 30 | 1.000 | 0.367 | 0.633 | 0.842 | 0.512 | 2.703 | 76,028 | 1,196 |
+
+## Paired comparisons (bootstrap over stories)
+
+| comparison | metric | mean Δ | 95% CI | excludes 0 | wins | n |
+|---|---|---|---|---|---|---|
+| base:dsg-state vs base:none | violation_rate | +0.0833 | [+0.0250, +0.1417] | **yes** | 16–6 | 30 |
+| base:dsg-state vs base:none | restatement_rate | +0.0292 | [+0.0125, +0.0500] | **yes** | 7–0 | 30 |
+| base:dsg-state vs base:none | mean_first_violation | -0.9618 | [-2.0806, +0.0936] | no | 13–17 | 30 |
+| base:dsg-state vs base:none | total_chars | +1316.4333 | [+247.5300, +2454.6150] | **yes** | 20–10 | 30 |
+| base:dsg-state vs base:none | prompt_tokens_last | +244.7000 | [+232.1325, +257.0667] | **yes** | 30–0 | 30 |
+| base:dsg-state vs base:full-context | violation_rate | +0.2583 | [+0.1917, +0.3250] | **yes** | 24–0 | 30 |
+| base:dsg-state vs base:full-context | restatement_rate | +0.0333 | [+0.0125, +0.0542] | **yes** | 8–0 | 30 |
+| base:dsg-state vs base:full-context | mean_first_violation | +1.9182 | [+0.4666, +2.9475] | **yes** | 27–1 | 30 |
+| base:dsg-state vs base:full-context | total_chars | -6675.9667 | [-7807.8433, -5398.0908] | **yes** | 2–28 | 30 |
+| base:dsg-state vs base:full-context | prompt_tokens_last | -5040.6667 | [-5139.2683, -4936.0650] | **yes** | 0–30 | 30 |
+| base:dsg-hybrid vs base:dsg-state | violation_rate | -0.2083 | [-0.2625, -0.1542] | **yes** | 1–25 | 30 |
+| base:dsg-hybrid vs base:dsg-state | restatement_rate | -0.0333 | [-0.0542, -0.0125] | **yes** | 0–8 | 30 |
+| base:dsg-hybrid vs base:dsg-state | mean_first_violation | -1.7771 | [-2.9766, -0.0603] | **yes** | 5–24 | 30 |
+| base:dsg-hybrid vs base:dsg-state | total_chars | +6314.2667 | [+5097.3242, +7426.5717] | **yes** | 28–2 | 30 |
+| base:dsg-hybrid vs base:dsg-state | prompt_tokens_last | +836.5333 | [+820.9333, +852.1667] | **yes** | 30–0 | 30 |
+| base:dsg-hybrid vs base:last-chapter | violation_rate | +0.0708 | [+0.0208, +0.1208] | **yes** | 18–5 | 30 |
+| base:dsg-hybrid vs base:last-chapter | restatement_rate | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 30 |
+| base:dsg-hybrid vs base:last-chapter | mean_first_violation | -0.3150 | [-1.7378, +0.5578] | no | 14–7 | 30 |
+| base:dsg-hybrid vs base:last-chapter | total_chars | -383.0333 | [-1157.6058, +402.6342] | no | 13–17 | 30 |
+| base:dsg-hybrid vs base:last-chapter | prompt_tokens_last | +185.0333 | [+164.3667, +211.3667] | **yes** | 30–0 | 30 |
+| base:dsg-hybrid vs base:full-context | violation_rate | +0.0500 | [+0.0000, +0.1000] | no | 15–7 | 30 |
+| base:dsg-hybrid vs base:full-context | restatement_rate | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 30 |
+| base:dsg-hybrid vs base:full-context | mean_first_violation | +0.1411 | [-1.7200, +1.9906] | no | 13–11 | 30 |
+| base:dsg-hybrid vs base:full-context | total_chars | -361.7000 | [-1362.7575, +603.0817] | no | 13–17 | 30 |
+| base:dsg-hybrid vs base:full-context | prompt_tokens_last | -4204.1333 | [-4300.3333, -4099.7658] | **yes** | 0–30 | 30 |
+| base:dsg-hybrid vs base:none | violation_rate | -0.1250 | [-0.1875, -0.0625] | **yes** | 3–16 | 30 |
+| base:dsg-hybrid vs base:none | restatement_rate | -0.0042 | [-0.0167, +0.0083] | no | 1–2 | 30 |
+| base:dsg-hybrid vs base:none | mean_first_violation | -2.7389 | [-3.9506, -1.3644] | **yes** | 3–26 | 30 |
+| base:dsg-hybrid vs base:none | total_chars | +7630.7000 | [+6956.7317, +8310.2542] | **yes** | 30–0 | 30 |
+| base:dsg-hybrid vs base:none | prompt_tokens_last | +1081.2333 | [+1067.4333, +1095.9333] | **yes** | 30–0 | 30 |
+| base:dsg-hybrid-repair vs base:dsg-hybrid | violation_rate | -0.0125 | [-0.0458, +0.0125] | no | 2–2 | 30 |
+| base:dsg-hybrid-repair vs base:dsg-hybrid | restatement_rate | +0.0042 | [+0.0000, +0.0125] | no | 1–0 | 30 |
+| base:dsg-hybrid-repair vs base:dsg-hybrid | mean_first_violation | +0.3528 | [-0.0444, +0.9667] | no | 3–2 | 30 |
+| base:dsg-hybrid-repair vs base:dsg-hybrid | total_chars | +90.9667 | [-305.5017, +448.0333] | no | 14–5 | 30 |
+| base:dsg-hybrid-repair vs base:dsg-hybrid | prompt_tokens_last | +2.2667 | [-8.1000, +13.9000] | no | 11–9 | 30 |
+
+## Continuity errors on real generated text
+
+| condition | ch | fact | contradicting text |
+|---|---|---|---|
+| base:none | 20 | `s01f03` | …tranquility as she chatted with the villagers, easing their fears and doubts. Silas, the blacksmith, worked tirelessly to repair t… |
+| base:full-context | 20 | `s01f00` | …the solitary figure of Thaddeus, hunched over his journal, his thoughts lost in a sea of grey skies and distant horizons. Ceridwen… |
+| base:full-context | 20 | `s04f00` | …As Thaddeus toiled to clear the silt from the locks, his grey eyes scanned the waterway with practiced care. The lantern of Ceridw… |
+| base:full-context | 20 | `s05f00` | …mean, it’s still floating? Has it been here all along?" Maud’s eyes met Thaddeus’s, her hazel gaze filled with a… |
+| base:full-context | 20 | `s07f00` | …nbury’s sake.” Odell, his green eyes wide with curiosity and a hint of nervousness, joined them. “What if it’s a trap? Wh… |
+| base:none | 20 | `s09f07` | …inst the town's neglect." Tobias, the town's blacksmith, appeared at the door, his face grave. "We must also consider the town's r… |
+| base:full-context | 20 | `s09f02` | …soot and sweat. His weathered face was a testament to the many years he had spent forging iron, but now, his gaze was more often c… |
+| base:none | 20 | `s11f03` | …or. With the help of his friends Ceridwen, the shrewd shopkeeper, and Hesper, the skilled blacksmith, they had managed to repair a… |
+| base:full-context | 20 | `s11f00` | …king metal filled the air, mingling with the crisp spring air. Every once in a while, his green eyes would stray to the village we… |
+| base:none | 20 | `s12f03` | …showcasing the local heritage and the unique features of our cathedral city." Silas, the blacksmith’s apprentice who had grown up … |
+| base:full-context | 20 | `s13f00` | …tep echoing off the wooden floor. Silas noticed Miriam sitting quietly in the corner, her hazel eyes fixed on the auctioneer. She … |
+| base:full-context | 20 | `s14f05` | …ly audible. "But he must be careful. The journey is long and perilous." Maud nodded, her green eyes narrowing. "Yes, Ambrose. We'l… |
+
+## Figures
+
+![canon](figures/gen1-canon-violation.png)
+![final](figures/gen2-final.png)

report/pdnc-qwen7b/results.md

@@ -0,0 +1,216 @@
+# Dynamic Story Graph — results
+
+Books: **28** · policies: window-only, append-only, dsg-merge, dsg-eager, dsg-full, retrospective · scoring time: 51441.4s
+
+## Means over books
+
+| policy | conll_f1 | mention_acc | mention_acc_answered | b3_f1_multi | coverage | fragmentation | conflation | inconsistent_slot_rate | speaker_acc | speaker_acc_matched | violations_per_100w | monotone_fraction | rollback | live_entities |
+|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+| window-only | 0.454 | 0.415 | 0.955 | 0.535 | 0.748 | 0.572 | 0.005 | 0.110 | 0.099 | 0.668 | 527551.5 | 1.000 | 0.000 | 968.5 |
+| append-only | 0.664 | 0.568 | 0.900 | 0.697 | 0.747 | 0.225 | 0.153 | 0.371 | 0.097 | 0.661 | 16501.7 | 1.000 | 0.000 | 223.2 |
+| dsg-merge | 0.662 | 0.563 | 0.882 | 0.695 | 0.747 | 0.214 | 0.187 | 0.377 | 0.096 | 0.650 | 16757.4 | 0.996 | 4.393 | 218.8 |
+| dsg-eager | 0.662 | 0.563 | 0.882 | 0.695 | 0.747 | 0.214 | 0.187 | 0.000 | 0.096 | 0.650 | 41.225 | 0.969 | 7.107 | 218.8 |
+| dsg-full | 0.659 | 0.570 | 0.896 | 0.690 | 0.747 | 0.235 | 0.138 | 0.000 | 0.097 | 0.659 | 1.465 | 0.975 | 2.607 | 270.5 |
+| retrospective | 0.657 | 0.598 | 0.894 | 0.689 | 0.747 | 0.235 | 0.139 | 0.000 | 0.098 | 0.664 | 14.286 | 0.970 | 2.500 | 312.4 |
+
+## Paired comparisons (bootstrap over books)
+
+| comparison | metric | mean Δ | 95% CI | excludes 0 | wins | n |
+|---|---|---|---|---|---|---|
+| append-only vs window-only | conll_f1 | +0.2091 | [+0.1815, +0.2395] | **yes** | 28–0 | 28 |
+| append-only vs window-only | mention_acc | +0.1523 | [+0.1238, +0.1825] | **yes** | 28–0 | 28 |
+| append-only vs window-only | mention_acc_answered | -0.0548 | [-0.0764, -0.0355] | **yes** | 2–23 | 28 |
+| append-only vs window-only | speaker_acc_matched | -0.0069 | [-0.0278, +0.0160] | no | 9–15 | 28 |
+| append-only vs window-only | violations_per_100w | -511049.7959 | [-728981.0248, -327298.4080] | **yes** | 0–28 | 28 |
+| append-only vs window-only | inconsistent_slot_rate | +0.2613 | [+0.2325, +0.2905] | **yes** | 28–0 | 28 |
+| append-only vs window-only | rollback | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| append-only vs window-only | monotone_fraction | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| append-only vs window-only | fragmentation | -0.3473 | [-0.4214, -0.2800] | **yes** | 0–28 | 28 |
+| append-only vs window-only | conflation | +0.1477 | [+0.1004, +0.2027] | **yes** | 26–0 | 28 |
+| dsg-merge vs append-only | conll_f1 | -0.0020 | [-0.0081, +0.0048] | no | 3–14 | 28 |
+| dsg-merge vs append-only | mention_acc | -0.0048 | [-0.0163, +0.0030] | no | 11–6 | 28 |
+| dsg-merge vs append-only | mention_acc_answered | -0.0178 | [-0.0354, -0.0049] | **yes** | 6–14 | 28 |
+| dsg-merge vs append-only | speaker_acc_matched | -0.0110 | [-0.0269, +0.0017] | no | 6–6 | 28 |
+| dsg-merge vs append-only | violations_per_100w | +255.6209 | [+34.2114, +523.2930] | **yes** | 9–1 | 28 |
+| dsg-merge vs append-only | inconsistent_slot_rate | +0.0065 | [+0.0008, +0.0137] | **yes** | 9–2 | 28 |
+| dsg-merge vs append-only | rollback | +4.3929 | [+3.2500, +5.8571] | **yes** | 27–0 | 28 |
+| dsg-merge vs append-only | monotone_fraction | -0.0043 | [-0.0055, -0.0032] | **yes** | 0–27 | 28 |
+| dsg-merge vs append-only | fragmentation | -0.0112 | [-0.0266, -0.0006] | **yes** | 0–4 | 28 |
+| dsg-merge vs append-only | conflation | +0.0349 | [+0.0191, +0.0539] | **yes** | 17–0 | 28 |
+| dsg-eager vs dsg-merge | conll_f1 | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| dsg-eager vs dsg-merge | mention_acc | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| dsg-eager vs dsg-merge | mention_acc_answered | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| dsg-eager vs dsg-merge | speaker_acc_matched | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| dsg-eager vs dsg-merge | violations_per_100w | -16716.1291 | [-24119.1620, -10583.9412] | **yes** | 0–28 | 28 |
+| dsg-eager vs dsg-merge | inconsistent_slot_rate | -0.3773 | [-0.4083, -0.3474] | **yes** | 0–28 | 28 |
+| dsg-eager vs dsg-merge | rollback | +2.7143 | [+1.8571, +3.6429] | **yes** | 22–0 | 28 |
+| dsg-eager vs dsg-merge | monotone_fraction | -0.0269 | [-0.0311, -0.0232] | **yes** | 0–28 | 28 |
+| dsg-eager vs dsg-merge | fragmentation | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| dsg-eager vs dsg-merge | conflation | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| dsg-full vs dsg-eager | conll_f1 | -0.0029 | [-0.0096, +0.0038] | no | 10–9 | 28 |
+| dsg-full vs dsg-eager | mention_acc | +0.0072 | [+0.0017, +0.0137] | **yes** | 12–4 | 28 |
+| dsg-full vs dsg-eager | mention_acc_answered | +0.0139 | [+0.0051, +0.0242] | **yes** | 16–2 | 28 |
+| dsg-full vs dsg-eager | speaker_acc_matched | +0.0086 | [-0.0020, +0.0202] | no | 7–3 | 28 |
+| dsg-full vs dsg-eager | violations_per_100w | -39.7595 | [-66.9078, -16.7992] | **yes** | 1–11 | 28 |
+| dsg-full vs dsg-eager | inconsistent_slot_rate | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| dsg-full vs dsg-eager | rollback | -4.5000 | [-5.9286, -3.3929] | **yes** | 0–27 | 28 |
+| dsg-full vs dsg-eager | monotone_fraction | +0.0064 | [+0.0050, +0.0077] | **yes** | 27–1 | 28 |
+| dsg-full vs dsg-eager | fragmentation | +0.0215 | [+0.0120, +0.0316] | **yes** | 13–0 | 28 |
+| dsg-full vs dsg-eager | conflation | -0.0494 | [-0.0745, -0.0264] | **yes** | 0–19 | 28 |
+| retrospective vs dsg-full | conll_f1 | -0.0019 | [-0.0073, +0.0026] | no | 5–7 | 28 |
+| retrospective vs dsg-full | mention_acc | +0.0280 | [+0.0154, +0.0431] | **yes** | 25–2 | 28 |
+| retrospective vs dsg-full | mention_acc_answered | -0.0022 | [-0.0144, +0.0128] | no | 8–16 | 28 |
+| retrospective vs dsg-full | speaker_acc_matched | +0.0049 | [-0.0018, +0.0124] | no | 11–4 | 28 |
+| retrospective vs dsg-full | violations_per_100w | +12.8205 | [+0.0000, +32.7839] | no | 2–0 | 28 |
+| retrospective vs dsg-full | inconsistent_slot_rate | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| retrospective vs dsg-full | rollback | -0.1071 | [-0.4643, +0.1429] | no | 2–2 | 28 |
+| retrospective vs dsg-full | monotone_fraction | -0.0055 | [-0.0071, -0.0040] | **yes** | 2–26 | 28 |
+| retrospective vs dsg-full | fragmentation | -0.0000 | [-0.0054, +0.0058] | no | 3–4 | 28 |
+| retrospective vs dsg-full | conflation | +0.0007 | [-0.0059, +0.0081] | no | 5–7 | 28 |
+| dsg-full vs append-only | conll_f1 | -0.0049 | [-0.0152, +0.0049] | no | 11–13 | 28 |
+| dsg-full vs append-only | mention_acc | +0.0023 | [-0.0113, +0.0135] | no | 15–6 | 28 |
+| dsg-full vs append-only | mention_acc_answered | -0.0040 | [-0.0251, +0.0145] | no | 15–8 | 28 |
+| dsg-full vs append-only | speaker_acc_matched | -0.0024 | [-0.0209, +0.0144] | no | 10–7 | 28 |
+| dsg-full vs append-only | violations_per_100w | -16500.2678 | [-23868.1494, -10409.4737] | **yes** | 0–28 | 28 |
+| dsg-full vs append-only | inconsistent_slot_rate | -0.3708 | [-0.4033, -0.3397] | **yes** | 0–28 | 28 |
+| dsg-full vs append-only | rollback | +2.6071 | [+1.7857, +3.5357] | **yes** | 22–0 | 28 |
+| dsg-full vs append-only | monotone_fraction | -0.0249 | [-0.0288, -0.0213] | **yes** | 0–28 | 28 |
+| dsg-full vs append-only | fragmentation | +0.0104 | [-0.0039, +0.0233] | no | 11–2 | 28 |
+| dsg-full vs append-only | conflation | -0.0146 | [-0.0471, +0.0173] | no | 11–12 | 28 |
+| dsg-full vs window-only | conll_f1 | +0.2042 | [+0.1760, +0.2339] | **yes** | 28–0 | 28 |
+| dsg-full vs window-only | mention_acc | +0.1546 | [+0.1253, +0.1851] | **yes** | 27–1 | 28 |
+| dsg-full vs window-only | mention_acc_answered | -0.0588 | [-0.0825, -0.0382] | **yes** | 3–22 | 28 |
+| dsg-full vs window-only | speaker_acc_matched | -0.0093 | [-0.0279, +0.0095] | no | 11–13 | 28 |
+| dsg-full vs window-only | violations_per_100w | -527550.0637 | [-749945.1216, -340553.8576] | **yes** | 0–28 | 28 |
+| dsg-full vs window-only | inconsistent_slot_rate | -0.1095 | [-0.1219, -0.0974] | **yes** | 0–28 | 28 |
+| dsg-full vs window-only | rollback | +2.6071 | [+1.7857, +3.5357] | **yes** | 22–0 | 28 |
+| dsg-full vs window-only | monotone_fraction | -0.0249 | [-0.0288, -0.0213] | **yes** | 0–28 | 28 |
+| dsg-full vs window-only | fragmentation | -0.3370 | [-0.4100, -0.2699] | **yes** | 0–28 | 28 |
+| dsg-full vs window-only | conflation | +0.1331 | [+0.0993, +0.1722] | **yes** | 26–0 | 28 |
+| dsg-full vs dsg-merge | conll_f1 | -0.0029 | [-0.0096, +0.0038] | no | 10–9 | 28 |
+| dsg-full vs dsg-merge | mention_acc | +0.0072 | [+0.0017, +0.0137] | **yes** | 12–4 | 28 |
+| dsg-full vs dsg-merge | mention_acc_answered | +0.0139 | [+0.0051, +0.0242] | **yes** | 16–2 | 28 |
+| dsg-full vs dsg-merge | speaker_acc_matched | +0.0086 | [-0.0020, +0.0202] | no | 7–3 | 28 |
+| dsg-full vs dsg-merge | violations_per_100w | -16755.8887 | [-24171.8070, -10626.9035] | **yes** | 0–28 | 28 |
+| dsg-full vs dsg-merge | inconsistent_slot_rate | -0.3773 | [-0.4083, -0.3474] | **yes** | 0–28 | 28 |
+| dsg-full vs dsg-merge | rollback | -1.7857 | [-3.2857, -0.3571] | **yes** | 8–19 | 28 |
+| dsg-full vs dsg-merge | monotone_fraction | -0.0206 | [-0.0247, -0.0167] | **yes** | 0–28 | 28 |
+| dsg-full vs dsg-merge | fragmentation | +0.0215 | [+0.0120, +0.0316] | **yes** | 13–0 | 28 |
+| dsg-full vs dsg-merge | conflation | -0.0494 | [-0.0745, -0.0264] | **yes** | 0–19 | 28 |
+| dsg-full vs retrospective | conll_f1 | +0.0019 | [-0.0026, +0.0073] | no | 7–5 | 28 |
+| dsg-full vs retrospective | mention_acc | -0.0280 | [-0.0431, -0.0154] | **yes** | 2–25 | 28 |
+| dsg-full vs retrospective | mention_acc_answered | +0.0022 | [-0.0128, +0.0144] | no | 16–8 | 28 |
+| dsg-full vs retrospective | speaker_acc_matched | -0.0049 | [-0.0124, +0.0018] | no | 4–11 | 28 |
+| dsg-full vs retrospective | violations_per_100w | -12.8205 | [-32.7839, +0.0000] | no | 0–2 | 28 |
+| dsg-full vs retrospective | inconsistent_slot_rate | +0.0000 | [+0.0000, +0.0000] | no | 0–0 | 28 |
+| dsg-full vs retrospective | rollback | +0.1071 | [-0.1429, +0.4643] | no | 2–2 | 28 |
+| dsg-full vs retrospective | monotone_fraction | +0.0055 | [+0.0040, +0.0071] | **yes** | 26–2 | 28 |
+| dsg-full vs retrospective | fragmentation | +0.0000 | [-0.0058, +0.0054] | no | 4–3 | 28 |
+| dsg-full vs retrospective | conflation | -0.0007 | [-0.0081, +0.0059] | no | 7–5 | 28 |
+| dsg-eager vs append-only | conll_f1 | -0.0020 | [-0.0081, +0.0048] | no | 3–14 | 28 |
+| dsg-eager vs append-only | mention_acc | -0.0048 | [-0.0163, +0.0030] | no | 11–6 | 28 |
+| dsg-eager vs append-only | mention_acc_answered | -0.0178 | [-0.0354, -0.0049] | **yes** | 6–14 | 28 |
+| dsg-eager vs append-only | speaker_acc_matched | -0.0110 | [-0.0269, +0.0017] | no | 6–6 | 28 |
+| dsg-eager vs append-only | violations_per_100w | -16460.5083 | [-23824.8619, -10369.4142] | **yes** | 0–28 | 28 |
+| dsg-eager vs append-only | inconsistent_slot_rate | -0.3708 | [-0.4033, -0.3397] | **yes** | 0–28 | 28 |
+| dsg-eager vs append-only | rollback | +7.1071 | [+5.5714, +8.9286] | **yes** | 28–0 | 28 |
+| dsg-eager vs append-only | monotone_fraction | -0.0312 | [-0.0354, -0.0274] | **yes** | 0–28 | 28 |
+| dsg-eager vs append-only | fragmentation | -0.0112 | [-0.0266, -0.0006] | **yes** | 0–4 | 28 |
+| dsg-eager vs append-only | conflation | +0.0349 | [+0.0191, +0.0539] | **yes** | 17–0 | 28 |
+
+## Where the reader's model is refined versus corrected
+
+Spearman correlation between position in the book and the share of update operations of each kind. The shuffled row permutes reading order, which preserves state growth exactly while destroying narrative order: a trend that survives the shuffle is an artefact of the state getting larger, one that disappears is a property of the text.
+
+| ordering | operation | rho | p | first bin | last bin |
+|---|---|---|---|---|---|
+| narrative | elaboration | -0.084 | 0.72 | 0.107 | 0.112 |
+| narrative | revision | +0.313 | 0.18 | 0.001 | 0.003 |
+| shuffled | elaboration | +0.684 | 0.00088 | 0.057 | 0.100 |
+| shuffled | revision | +0.325 | 0.16 | 0.001 | 0.004 |
+
+## Identity-resolution latency (measurement instrument)
+
+How long a novel holds a figure unnamed before the reader is told who they are, by narrative person. Unpaired bootstrap on the group means; the first-person group is small, so read the interval, not the point.
+
+| metric | first-person | third-person | Δ | 95% CI | excludes 0 |
+|---|---|---|---|---|---|
+| mean_latency | 0.074 (n=7) | 0.089 (n=21) | -0.016 | [-0.091, +0.067] | no |
+| median_latency | 0.042 (n=7) | 0.050 (n=21) | -0.008 | [-0.072, +0.065] | no |
+| unresolved_fraction | 0.964 (n=7) | 0.970 (n=21) | -0.006 | [-0.037, +0.021] | no |
+| max_latency | 0.198 (n=7) | 0.278 (n=21) | -0.080 | [-0.274, +0.129] | no |
+
+## One character, traced (dsg-full)
+
+### AHandfulOfDust
+
+**Tony Last** — what the reader believes, and when.
+
+| read at | predicate | value | held until | status | source |
+|---|---|---|---|---|---|
+| 0.03 | child_of | Aunt Frances | — | believed | reported |
+| 0.03 | emotion | vexation | window 13 | superseded | narrated |
+| 0.03 | knows | damp had penetrated into one corner of t | — | believed | reported |
+| 0.03 | occupation | detective | — | believed | reported |
+| 0.03 | parent_of | Brenda | — | believed | reported |
+| 0.03 | resides_in | Hetton | window 33 | superseded | reported |
+| 0.03 | travelling_to | the church bells | window 15 | superseded | narrated |
+| 0.05 | other:believes | Brenda | — | believed | narrated |
+| 0.07 | possesses | the house and lots of other things | — | believed | narrated |
+| 0.07 | knows | Brenda | — | believed | reported |
+| 0.07 | knows | Pigstanton | — | believed | reported |
+| 0.08 | knows | Mrs Beaver | — | believed | reported |
+
+### APassageToIndia
+
+**Dr. Aziz** — what the reader believes, and when.
+
+| read at | predicate | value | held until | status | source |
+|---|---|---|---|---|---|
+| 0.02 | child_of | unnamed wife | window 37 | retracted | narrated |
+| 0.02 | possesses | two boys and their sister | — | believed | reported |
+| 0.02 | resides_in | India | window 35 | superseded | narrated |
+| 0.02 | other:has_a_good | memory | — | believed | reported |
+| 0.02 | other:is_worried | about the appeal | — | believed | reported |
+| 0.02 | other:prefers | themes of the decay of Islam and the bre | — | believed | reported |
+| 0.03 | other:believes | there is a message | — | believed | reported |
+| 0.03 | other:believes | the servant is withholding the message o | — | believed | reported |
+| 0.03 | other:fears | a gross snub | — | believed | reported |
+| 0.03 | travelling_to | Mau | — | believed | reported |
+| 0.03 | knows | Old Callendar | — | believed | narrated |
+| 0.03 | knows | the servant | — | believed | narrated |
+
+
+## Revision events on real text (dsg-full)
+
+| op | book | position | change | evidence |
+|---|---|---|---|---|
+| `revise` | AHandfulOfDust | 0.62 | Winnie.child_of: 'Milly' → 'Tony Last' | …right?" "Love's young dream," said Milly. "All right, then. I'll ring the bell." When the tray had been brought, Tony got out of bed and put on his th… |
+| `revise` | AHandfulOfDust | 0.43 | John.child_of: 'Mrs Beaver' → 'Tony Last' | …Now she stood with her eyes at an arrow slit gazing rather resentfully at the animated scene below. "It's all a lot of nonsense of Ben Hacket's," she … |
+| `merge_provisional` | AHandfulOfDust | 0.87 | absorbed e00796: link 'the woman' -> 'Brenda' | …rly, before you've made any plans." "Yes." "Then good night, bless you." Beaver went up the two flights of stairs again, and got into bed. * * * * * "… |
+| `merge_provisional` | AHandfulOfDust | 0.87 | absorbed e00423: link 'the doctor' -> 'Dr Messinger' | …ly of them to be scared by a toy. I suppose the others are following." "Yes, I expect so. Try and sit still." "Damned fool, being frightened of a toy … |
+| `merge_provisional` | AHandfulOfDust | 0.29 | absorbed e00282: link 'the Old Hundredth' -> 'Hetton Abbey' | …clock, though the wind had got up, fluttering the papers on which the order of the service was printed and once threatening to unveil the memorial pre… |
+| `merge_provisional` | AHandfulOfDust | 0.29 | absorbed e00283: link 'the kindly young lady' -> 'Mrs Beaver' | …had telegraphed to Brenda asking her to dine. "Mrs Jimmy Deane's very upset that she couldn't get you for to-night. I didn't give away what you were d… |
+| `supersede` | AHandfulOfDust | 0.97 | Tony Last.emotion: 'awful' → 'cordiality' | …sunset to dawn with the small wick glowing in the beef fat and the palm thatch overhead dimly discernible; but Tony lived in quiet confidence and expe… |
+| `supersede` | AHandfulOfDust | 0.85 | Mrs Beaver.travelling_to: 'Brenda's house' → 'America' | …t's quite light," she added. "After five. Oughtn't we to go to bed?" "Yes." "Only another week and then all the parties will be over," said Beaver. "Y… |
+| `supersede` | AHandfulOfDust | 0.85 | Brenda.emotion: 'sad' → 'doesn't want to go' | …t's quite light," she added. "After five. Oughtn't we to go to bed?" "Yes." "Only another week and then all the parties will be over," said Beaver. "Y… |
+| `supersede` | AHandfulOfDust | 0.85 | Tony Last.location: 'on a ship' → 'the river Waurupang' | …t's quite light," she added. "After five. Oughtn't we to go to bed?" "Yes." "Only another week and then all the parties will be over," said Beaver. "Y… |
+| `supersede` | AHandfulOfDust | 0.85 | Dr Messinger.emotion: 'queer' → 'not desperate' | …t's quite light," she added. "After five. Oughtn't we to go to bed?" "Yes." "Only another week and then all the parties will be over," said Beaver. "Y… |
+| `supersede` | AHandfulOfDust | 0.77 | Jock Grant-Menzies.occupation: 'speaker_of_speech' → 'member_of' | …e chief want talk to you." At last, very shyly, a little old woman emerged, clad in the filthy calico gown that was kept for use in the presence of st… |
+| `supersede` | AHandfulOfDust | 0.67 | Tony Last.resides_in: 'Brighton' → 'Hetton' | …left his greatcoat in the cabin and went down to dinner. It was the first evening at sea. Tony sat at the captain's table, but the captain was on the … |
+| `supersede` | AHandfulOfDust | 0.66 | Tony Last.location: 'in the hall' → 'on a ship' | …ted the old boy to turn up like that?" asked Polly Cockpurse. "Now I understand why they keep going on in the papers about divorce law reform," said V… |
+| `supersede` | AHandfulOfDust | 0.59 | Tony Last.resides_in: 'Jock's house' → 'Brighton' | …ir, I don't want to appear unobliging, but it's not part of my duties." He seemed to be wavering but Blenkinsop interposed. "Quite out of the question… |
+| `supersede` | AHandfulOfDust | 0.53 | Tony Last.emotion: 'wishes' → 'awful' | …on't go on." So Tony stopped and after a time said, "So you're going to Veronica's to-morrow?" "Mmmm." "I think I will ask Jock to come." "Yes, I shou… |
+| `supersede` | AHandfulOfDust | 0.47 | Brenda.resides_in: 'Hetton' → 'Sussex Gardens' | …been in London such a lot. I'm afraid that's going to hurt her." "You can't ever tell what's going to hurt people." "But, you see, I know Brenda so we… |
+| `supersede` | AHandfulOfDust | 0.39 | Tony Last.emotion: 'hates' → 'wishes' | …ked?" "Well, you _were_, of course, but I refused for you. You know you always hate staying away." "I wouldn't mind coming." "Oh, darling, I wish I'd … |
+| `supersede` | AHandfulOfDust | 0.33 | Brenda.location: 'London' → 'Hetton' | …John. "You ungrateful little bastard, that's a lousy way to speak of your dad." "And you ought not to say bastard or lousy in front of me, nanny says … |
+| `supersede` | AHandfulOfDust | 0.33 | Tony Last.emotion: 'low' → 'hates' | …John. "You ungrateful little bastard, that's a lousy way to speak of your dad." "And you ought not to say bastard or lousy in front of me, nanny says … |
+
+## Figures
+
+![main](figures/fig1-main-results.png)
+![ladder](figures/fig1b-ladder.png)
+![prefix](figures/fig2-prefix-curve.png)
+![paired](figures/fig3-paired-deltas.png)
+![revision](figures/fig4-revision-profile.png)
+![violations](figures/fig5-violation-growth.png)