Existing: 16 cells ~985 tok (61.5/cell, 6/6 PASS); 64 cells ~3,980 tok (62/cell, drift +0.5%, 7/7 PASS); Kubota 6x6 run 36/36 cells, gates 1-6 closed. UPDATE 2026-11-19: the independent A/B re-run below has now been run and measured with a real tokenizer (Section 2). Do NOT use the published Linux A/B pair as evidence — audit found B is A re-rendered. ## 1. Experimental Setup (existing runs, from the OpenVersions corpus) - Source run 1 — Kubota 6x6 Discovery Matrix, second pass, with the anti-hang protocol named explicitly: 36/36 cells completed, Stages I-VI, [GATE k/6 COMPLETE] markers at every stage boundary, every cell in the fixed schema (Formal Definition, 4-Step Construction, Post-Oct 2026 Novelty / Why AI Sees It). - Source run 2 — Kubota 6x6 Discovery Matrix, first pass: 36/36 cells completed without the protocol being named as a separate Part 0. - Demo grids cited in formal-names-for-context-window-strategies: 4x4 consensus grid: 16 cells, ~985 tokens total, ~61.5 tokens/cell, 6/6 PASS. 8x8 consensus grid: 64 cells, ~3,980 tokens total, ~62 tokens/cell, per-cell band 60.8 -> 62.8, Stage-8 drift vs S1 +0.5%, 7/7 PASS. Prose-baseline equivalence (hypothetical, NOT measured): ~3,200-4,000 tokens (4x4) and ~13,000-16,000 tokens (8x8), claimed ~3.3-4.0x compression. Signal-to-scaffold ~88-89% payload. Scaling rule: T_total ~ 62 x N, budget max_tokens = 65 x N. - Token figures in all existing runs are labelled "Estimated BPE Tokens (cl100k/o200k)". No external token-counter benchmark was run in the source manuscripts — this was the main evidence gap, closed by the measured re-run in Section 2 (2026-11-19). ## 2. Independent A/B re-run — MEASURED 2026-11-19 (real tokenizer) Design: four independent generations by four isolated agents, none of which could see any other arm's output. Identical content task per pair; the ONLY difference is the protocol instruction. Control = full prose (Definition / numbered Execution flow / Edge cases, no limits, no skeleton, no checkpoints). Protocol = the exact MDL prompt (skeleton line; one line per cell, Def <=15 words, Edge <=12 words; [CP k/K] checkpoint after each stage). Grid 1: distributed consensus, 6 stages x 6 = 36 cells. Grid 2: Linux kernel components, 8 categories x 10 = 80 cells (the exact task of the published A/B pair this re-run replaces). Measurement: tiktoken cl100k_base (o200k_base within 0.5% on every run), script measure.py, full cell-level data in measurements.json. A cell counts complete only if all three parts are present and non-empty. | Run | Complete | Total tok | Mean/cell | Median | Min-Max | Drift (last vs first stage) | File-writes to finish | |---|---|---|---|---|---|---|---| | control36 | 36/36 | 15,872 | 438.9 | 437.5 | 365-507 | +9.88% | 3 | | protocol36 | 36/36 | 1,784 | 45.3 | 44.5 | 40-50 | -0.45% | 1 | | control80 | 80/80 | 27,360 | 340.8 | 336.0 | 295-409 | +13.36% | 4 | | protocol80 | 80/80 | 3,673 | 43.1 | 43.0 | 39-49 | +2.34% | 1 | - Compression vs the measured prose control: 8.90x (36-cell), 7.45x (80-cell). Total-token saving 88.8% and 86.6%. (The source paper's 3.3-4.0x figure was against a hypothetical prose baseline, not a measured one.) - Drift: the protocol passes the paper's own <3% per-cell drift target on both grids (-0.45%, +2.34%); the control fails it on both (+9.88%, +13.36%). - Word-bound compliance (protocol arms, measured, not self-reported): Def <=15w: 36/36 and 80/80 (median 9 words, max 10). Edge <=12w: 36/36 and 80/80 (median 7 and 6 words, max 8). Contrast the published compact view: EDGE median 64 words vs a 12-word bound. - Scaffold overhead (skeleton + checkpoint lines): 117 of 1,784 tokens (6.6%) at 36 cells; 146 of 3,673 (4.0%) at 80 cells. - Truncation: NO run truncated in this environment. The honest difference is completion shape: both protocol runs finished in a single file-write; both controls needed sequential appends (3 and 4 writes) to finish. No claim about a per-turn token ceiling is made or tested here. - Scaling: measured protocol cost is ~43-45 tokens/cell, i.e. T_total ~ 45 x N in this schema, below the source paper's 62 x N estimate (its cells carried longer Steps segments). Control cost is task-dependent (438.9 vs 340.8/cell). Resume test — MEASURED 2026-11-19: a fifth, fresh agent received ONLY the skeleton line and the marker [CP 3/6 | NEXT: S4|I1], with no access to stages 1-3 content. Result: started exactly at [S4|I1], produced exactly 18/18 cells for stages 4-6, ended [S6|I6] + [CP 6/6 | COMPLETE], 3/3 checkpoint lines, 0 duplicated coordinates, 0 repeated stages, in 1 write. Resumed cells cost 45.3 tokens/cell on average — identical to the original protocol36 mean (45.3). Word bounds 18/18 on both Def and Edge. Negative example (not evidence): the Linux-kernel A/B pair published on OpenVersions. Audit (provenance-formalization-and-independent-audit, 2026-11-18) found definitions, flows and edge cases byte-identical (80/80/80) between control A and treatment B; B is A re-rendered, not an independent run, and its compact view violates the protocol's own bounds. The re-run above replaces it. ## 3. Baselines and literature anchors (citation-checked in the audit) - Skeleton-of-Thought: Ning et al., 2023 — arXiv 2307.15337 - MemGPT: Packer et al., 2023 — arXiv 2310.08560 - Lost in the Middle: Liu et al., 2023 — arXiv 2307.03172 - Plan-and-Solve: Wang et al., 2023 — arXiv 2305.04091