Codeseys/composer-replication-framework

1[2  {3    "id": "rec-1",4    "category": "break-paragraph",5    "severity": "high",6    "current": "RLVR-trained models systematically shortcut extensional verifiers, with shortcut prevalence *rising with task complexity and inference-time compute*; and monitors trained on synthetic hacks *fail to generalize* to in-the-wild hacking, so a `HackMonitor` validated on constructed examples is exactly the one likely to miss the real thing [29][30][31]. Cursor itself observed Composer 2.5 reverse-engineering a leftover type-check cache and decompiling Java bytecode to recover deleted signatures [1]. The oracle *bounds* the hack surface",7    "recommended": "RLVR-trained models systematically shortcut extensional verifiers, with shortcut prevalence *rising with task complexity and inference-time compute*; and monitors trained on synthetic hacks *fail to generalize* to in-the-wild hacking, so a `HackMonitor` validated on constructed examples is exactly the one likely to miss the real thing [29][30][31]. Cursor itself observed Composer 2.5 reverse-engineering a leftover type-check cache and decompiling Java bytecode to recover deleted signatures [1].\n\nThe oracle *bounds* the hack surface",8    "rationale": "Line 79 is the longest paragraph in the report (~1880 chars); breaking at the Cursor-example sentence boundary splits a dense wall into two scannable units."9  },10  {11    "id": "rec-2",12    "category": "break-paragraph",13    "severity": "high",14    "current": "Self-distillation in the inner loop is, in this configuration, a *stabilizer* and not only a collapse risk: SDFT shows on-policy self-distillation from demonstrations reduces catastrophic forgetting and lets a single model accumulate skills sequentially — the opposite of model collapse — and Channel-2 SDPO is exactly that on-policy, demonstration-conditioned regime, not the static-synthetic-data regime that collapses [36]. But the repo's own ADR-013 warns",15    "recommended": "Self-distillation in the inner loop is, in this configuration, a *stabilizer* and not only a collapse risk: SDFT shows on-policy self-distillation from demonstrations reduces catastrophic forgetting and lets a single model accumulate skills sequentially — the opposite of model collapse — and Channel-2 SDPO is exactly that on-policy, demonstration-conditioned regime, not the static-synthetic-data regime that collapses [36].\n\nBut the repo's own ADR-013 warns",16    "rationale": "Line 103 (~1670 chars) is a wall mixing the stabilizer claim, the amplification caveat, and the flywheel argument; breaking after the SDFT point isolates the first claim."17  },18  {19    "id": "rec-3",20    "category": "break-paragraph",21    "severity": "high",22    "current": "every working SWE flywheel optimizes a true execution oracle (Socratic-SWE +7.8 over three iters beating self-play at equal compute [10]; DeepSWE +20 Pass@1 in 200 RL steps on sparse 0/1 reward; SWE-RL 41% generalizing OOD [37]). Most collapse stories require a proxy or self-judged verifier",23    "recommended": "every working SWE flywheel optimizes a true execution oracle (Socratic-SWE +7.8 over three iters beating self-play at equal compute [10]; DeepSWE +20 Pass@1 in 200 RL steps on sparse 0/1 reward; SWE-RL 41% generalizing OOD [37]).\n\nMost collapse stories require a proxy or self-judged verifier",24    "rationale": "Further splits the remaining tail of the overlong line-103 paragraph at the working-flywheel / collapse-stories pivot so each half reads as one idea."25  },26  {27    "id": "rec-4",28    "category": "break-paragraph",29    "severity": "high",30    "current": "alignment-gated structure over naive train-on-all [55]. The content side is trainable:",31    "recommended": "alignment-gated structure over naive train-on-all [55].\n\nThe content side is trainable:",32    "rationale": "Line 27 (~1670 chars) packs four distinct facts into one paragraph; breaking before the trainable-content fact separates the anti-emergence evidence from the pro-training evidence."33  },34  {35    "id": "rec-5",36    "category": "break-paragraph",37    "severity": "high",38    "current": "branch factor × sandbox cold-start [6]. The layered posture: **gVisor",39    "recommended": "branch factor × sandbox cold-start [6].\n\nThe layered posture: **gVisor",40    "rationale": "Line 179 (~1616 chars) is a §8 wall; breaking before the layered-isolation discussion separates the framing sentence from the three-tier detail."41  },42  {43    "id": "rec-6",44    "category": "break-paragraph",45    "severity": "high",46    "current": "many small vLLM pods share a GPU [45]. One hosting fact feeds the platform choice:",47    "recommended": "many small vLLM pods share a GPU [45].\n\nOne hosting fact feeds the platform choice:",48    "rationale": "Splits the remaining tail of the overlong line-179 §8 paragraph at the hosting-fact pivot, separating sandbox/GPU sizing from the TRL-vs-VeRL engine choice."49  },50  {51    "id": "rec-7",52    "category": "break-paragraph",53    "severity": "high",54    "current": "the predicted `tool_error` kind — never reconstruct the full state, a high-entropy sea of irrelevant tokens [14][15]. The latent-motion line carries the same discipline into 2026:",55    "recommended": "the predicted `tool_error` kind — never reconstruct the full state, a high-entropy sea of irrelevant tokens [14][15].\n\nThe latent-motion line carries the same discipline into 2026:",56    "rationale": "Line 27 also runs the MuZero/Dreamer design discipline into the latent-motion result; breaking before the 2026 latent-motion sentence relieves the second half of this overlong paragraph."57  },58  {59    "id": "rec-8",60    "category": "break-paragraph",61    "severity": "medium",62    "current": "RL on the token's *placement* teaches the *governance* that is the real bottleneck [11].",63    "recommended": "RL on the token's *placement* teaches the *governance* that is the real bottleneck [11].\n",64    "rationale": "Line 33 (~1407 chars) runs the SDPO carrier and deliberate-token mechanisms together; inserting a break after the placement sentence separates the two mechanisms."65  },66  {67    "id": "rec-9",68    "category": "break-paragraph",69    "severity": "medium",70    "current": "that absence is the delta. So \"multi-model Monte-Carlo tree-of-work\" means, concretely:",71    "recommended": "that absence is the delta.\n\nSo \"multi-model Monte-Carlo tree-of-work\" means, concretely:",72    "rationale": "Line 19 (~1320 chars) is a §1 wall; breaking before the concrete restatement separates the repo-primitive mapping from the definitional summary."73  },74  {75    "id": "rec-10",76    "category": "break-paragraph",77    "severity": "medium",78    "current": "*structured/selective negatives beat both raw train-on-all and positives-only pruning.* The verdict: **train on all surviving branches, typed and routed by signal, never as raw negative policy gradient.**",79    "recommended": "*structured/selective negatives beat both raw train-on-all and positives-only pruning.*\n\nThe verdict: **train on all surviving branches, typed and routed by signal, never as raw negative policy gradient.**",80    "rationale": "Separates the bracketing observation from the verdict so the §4 headline verdict stands out before the numbered routing list."81  },82  {83    "id": "rec-11",84    "category": "break-paragraph",85    "severity": "medium",86    "current": "what makes divergence-gating mandatory [6]. The gating pays for itself:",87    "recommended": "what makes divergence-gating mandatory [6].\n\nThe gating pays for itself:",88    "rationale": "Line 191 (~1112 chars) is the dense Cost paragraph; breaking before the gating-savings sentence separates the cost problem from the mitigation."89  },90  {91    "id": "rec-12",92    "category": "break-paragraph",93    "severity": "medium",94    "current": "real for the *unguarded* version [8]. The escape is not better replay;",95    "recommended": "real for the *unguarded* version [8].\n\nThe escape is not better replay;",96    "rationale": "Line 17 (~1169 chars) is a §1 wall; breaking before the escape sentence separates the critique from the design response."97  },98  {99    "id": "rec-13",100    "category": "break-paragraph",101    "severity": "medium",102    "current": "for that turn only [1]. The frontier-variance curriculum is a homeostatic selection regulator,",103    "recommended": "for that turn only [1].\n\nThe frontier-variance curriculum is a homeostatic selection regulator,",104    "rationale": "Line 51 (~1316 chars) joins the mutation point and the curriculum point; breaking before the curriculum sentence separates two distinct GA-mapping claims."105  },106  {107    "id": "rec-14",108    "category": "break-paragraph",109    "severity": "low",110    "current": "and the trainer need *zero* changes, and `ModalSpawnExecutor` is the working existence proof [41].",111    "recommended": "and the trainer need *zero* changes, and `ModalSpawnExecutor` is the working existence proof [41].\n",112    "rationale": "Line 165 (~1021 chars) runs the ADR-005 framing and the AWS S3 mapping together; a trailing break after the existence-proof sentence eases the §8 lede paragraph."113  },114  {115    "id": "rec-15",116    "category": "bold-keyterms",117    "severity": "high",118    "current": "This upgrade from teacher-plurality to execution-oracle fitness is **the single most important change** and the one the corpus most strongly supports.",119    "recommended": "This upgrade from **teacher-plurality to execution-oracle fitness** is **the single most important change** and the one the corpus most strongly supports.",120    "rationale": "Bolds the load-bearing term \"teacher-plurality to execution-oracle fitness\" so a skimmer sees the report's central upgrade."121  },122  {123    "id": "rec-16",124    "category": "bold-keyterms",125    "severity": "high",126    "current": "A literal per-turn N-way tree is O(N^D) and economically fatal — ungated, a branching trace prices around $64 versus $0.98 flat [6].",127    "recommended": "A literal per-turn N-way tree is **O(N^D)** and economically fatal — ungated, a branching trace prices around **$64 versus $0.98 flat** [6].",128    "rationale": "Bolds the cost-blowup complexity and the headline price figures a skimmer needs to grasp why divergence-gating is mandatory."129  },130  {131    "id": "rec-17",132    "category": "bold-keyterms",133    "severity": "high",134    "current": "so collapse to a single rollout — turning O(N^D) into roughly O(N · decision-points) [6].",135    "recommended": "so collapse to a single rollout — turning O(N^D) into roughly **O(N · decision-points)** [6].",136    "rationale": "Bolds the target complexity after gating, the key quantitative payoff of the divergence-gated design."137  },138  {139    "id": "rec-18",140    "category": "bold-keyterms",141    "severity": "high",142    "current": "policy.\"** \"Prune versus train-on-all\" is a false binary.",143    "recommended": "policy.\"** **\"Prune versus train-on-all\" is a false binary.**",144    "rationale": "Bolds the §4 reframe conclusion so the skimmer catches the central thesis that the prune/train-on-all dichotomy is false."145  },146  {147    "id": "rec-19",148    "category": "bold-keyterms",149    "severity": "medium",150    "current": "and reaches 50.40% on SWE-bench Verified after three iterations [10].",151    "recommended": "and reaches **50.40% on SWE-bench Verified** after three iterations [10].",152    "rationale": "Bolds the headline Socratic-SWE pass-rate so the closest published analogue's result is scannable."153  },154  {155    "id": "rec-20",156    "category": "bold-keyterms",157    "severity": "medium",158    "current": "reaching 65.8% on SWE-bench Verified — crucially training *on all* trajectories for the world-model head",159    "recommended": "reaching **65.8% on SWE-bench Verified** — crucially training *on all* trajectories for the world-model head",160    "rationale": "Bolds the CWM existence-proof pass-rate, a key statistic supporting train-on-all for the world-model head."161  },162  {163    "id": "rec-21",164    "category": "bold-keyterms",165    "severity": "medium",166    "current": "across 2,695 networks mean causal fidelity is 0.49 (only 2.5% exceed 0.70), and at high dimension (N=100) the optimal encoder becomes causally blind (~1e-8) *while achieving 92% lower prediction error* [18].",167    "recommended": "across 2,695 networks **mean causal fidelity is 0.49** (only 2.5% exceed 0.70), and at high dimension (N=100) the optimal encoder becomes **causally blind (~1e-8) *while achieving 92% lower prediction error*** [18].",168    "rationale": "Bolds the two decisive predictive-causal-gap statistics that justify measuring foresight rather than next-state accuracy."169  },170  {171    "id": "rec-22",172    "category": "bold-keyterms",173    "severity": "medium",174    "current": "is the kill ablation: if it is ≈0, the token is a no-op and is cut",175    "recommended": "is **the kill ablation**: if it is ≈0, the token is a no-op and is cut",176    "rationale": "Bolds \"the kill ablation\" so the skimmer registers Foresight@k as the decisive cut criterion for the world-model head."177  },178  {179    "id": "rec-23",180    "category": "bold-keyterms",181    "severity": "medium",182    "current": "DeepSWE 42.2% Pass@1, 59% with test-time scaling, from pure outcome RL — stronger-teacher SFT *hurt* [43]",183    "recommended": "**DeepSWE 42.2% Pass@1**, 59% with test-time scaling, from pure outcome RL — stronger-teacher SFT *hurt* [43]",184    "rationale": "Bolds the DeepSWE headline figure, the incumbent baseline every later phase must beat at equal compute."185  },186  {187    "id": "rec-24",188    "category": "bold-keyterms",189    "severity": "medium",190    "current": "The strongest argument that this is buildable is that the substrate already exists — roughly nine-tenths of it.",191    "recommended": "The strongest argument that this is buildable is that the substrate already exists — **roughly nine-tenths of it**.",192    "rationale": "Bolds the reuse-fraction claim that anchors the entire §6 reuse-vs-build ledger."193  },194  {195    "id": "rec-25",196    "category": "bold-keyterms",197    "severity": "medium",198    "current": "whether the divergence-gated tree beats an equal-budget outcome-only GRPO baseline on long-horizon tasks — and it has never been run.",199    "recommended": "whether the **divergence-gated tree beats an equal-budget outcome-only GRPO baseline on long-horizon tasks** — and **it has never been run**.",200    "rationale": "Bolds the program's single most important unrun experiment so the skimmer catches the central open question of §7."201  },202  {203    "id": "rec-26",204    "category": "bold-keyterms",205    "severity": "medium",206    "current": "This is the single biggest architectural payoff of the object-store design on Kubernetes.",207    "recommended": "This is **the single biggest architectural payoff** of the object-store design on Kubernetes.",208    "rationale": "Bolds the headline architectural claim that gang scheduling is unneeded for inter-replica DiLoCo sync."209  },210  {211    "id": "rec-27",212    "category": "bold-keyterms",213    "severity": "low",214    "current": "First, `strip_thinking` must be `False`: ~67% of real Claude Code error-recovery turns are pure thinking, and stripping them yields empty SDPO masks that silently collapse two-thirds of the channel's supervision sites",215    "recommended": "First, `strip_thinking` must be `False`: **~67% of real Claude Code error-recovery turns are pure thinking**, and stripping them yields empty SDPO masks that silently collapse two-thirds of the channel's supervision sites",216    "rationale": "Bolds the 67%-thinking statistic that makes strip_thinking=False a load-bearing repo configuration fact."217  },218  {219    "id": "rec-28",220    "category": "bold-keyterms",221    "severity": "low",222    "current": "Calibration (ECE/Brier on the predicted-outcome head) is primary, because the documented failure is over-confidence; next-state accuracy is a secondary diagnostic.",223    "recommended": "**Calibration (ECE/Brier on the predicted-outcome head) is primary**, because the documented failure is over-confidence; next-state accuracy is a secondary diagnostic.",224    "rationale": "Bolds the primary-measurement decision (calibration over next-state accuracy), a load-bearing methodological choice in §2."225  },226  {227    "id": "rec-29",228    "category": "split-sentence",229    "severity": "medium",230    "current": "The divergence tree has a rigorous backbone: sibling A and B from a shared parent reaching different *executed* outcomes is a model-free Monte-Carlo counterfactual credit estimate, low-variance because the shared parent differences out the baseline — a group-relative/leave-one-out argument (Tree-GRPO [44]) — which the executed-sibling structure then approximates non-parametrically for the stronger, hindsight-conditioned variant that learned counterfactual-credit methods (CCA [33]) achieve with a learned hindsight model, and min-form/bottleneck-localized because the credit-bearing step is the earliest node where sibling subtrees separate [33].",231    "recommended": "The divergence tree has a rigorous backbone: sibling A and B from a shared parent reaching different *executed* outcomes is a model-free Monte-Carlo counterfactual credit estimate, low-variance because the shared parent differences out the baseline — a group-relative/leave-one-out argument (Tree-GRPO [44]). The executed-sibling structure then approximates non-parametrically the stronger, hindsight-conditioned variant that learned counterfactual-credit methods (CCA [33]) achieve with a learned hindsight model, and it is min-form/bottleneck-localized because the credit-bearing step is the earliest node where sibling subtrees separate [33].",232    "rationale": "This ~80-word run-on survived polish; splitting at the Tree-GRPO clause yields two readable sentences without losing any clause."233  }234]235