|
| 1 | +# ECP-8 Sealed Development Log |
| 2 | + |
| 3 | +The ECP-8 confirmatory split is sealed. This log records every development arm, |
| 4 | +including negative results, before any possible confirmatory access. |
| 5 | + |
| 6 | +## Batch 1 preregistration — Global channel capacity |
| 7 | + |
| 8 | +Preregistered: July 18, 2026 at 19:58:58 UTC<br> |
| 9 | +Seed: `11`<br> |
| 10 | +Test unsealed: **no**<br> |
| 11 | +Split SHA-256: `241acd2ee42ff658ddf0c8c897d4e4df0f52a18098b3ef68c009c74c2134592a` |
| 12 | + |
| 13 | +| Arm | Configuration | Raw SHA-256 | Effective SHA-256 | |
| 14 | +|---|---|---|---| |
| 15 | +| ECP8-B1-P | `config/ecp8-positive-control-development.yaml` | `e996bf66c30b3e699b27046e9824c13c7ecf3f2cb9831c4b9f0528d17b7b5f05` | `3a8e55ee262418f1005405afec3eeee7ebaaad164cb83274bd95c95e2c0939c3` | |
| 16 | +| ECP8-B1-C | `config/ecp8-control-development.yaml` | `9e3c865fcec9698bb95fcfc3ad429a354e26ffd39f2533dd7dc836dfdd3ab34a` | `fcaf5e8f39df2349cb3dedefcc233e5a08d813763a5b644d6a2ed2c050748f06` | |
| 17 | +| ECP8-B1-I | `config/ecp8-development.yaml` | `f6219110ebfe8f26248d9dbd646d80e1033f63d618a2524f7b96668ebd7fdaea` | `0572857aa786e0a41995aedd1cee8c2bb6dbdbc47f46025411bb8c06615b9693` | |
| 18 | + |
| 19 | +ECP8-B1-C reruns the strongest ECP-7 weak-structure setup on the fresh ECP-8 |
| 20 | +split. ECP8-B1-I changes only the local channel capacities: `[16,16,8,8]` |
| 21 | +becomes `[16,16,16,16]`, increasing fixed payload size from 14 to 16 bits. |
| 22 | +The sender still has one generic joint context and four position heads; neither |
| 23 | +sender nor receiver receives a factor-slot binding. |
| 24 | + |
| 25 | +The full decision gate is fixed in `docs/research-design-ecp8.md`. In addition |
| 26 | +to the existing train, validation and translator thresholds, the intervention |
| 27 | +must be injective over all 15,360 accessible meanings, improve the paired |
| 28 | +control's minimum codebook, and preserve both mean and worst-link validation. |
| 29 | +Only these three seed-11 runs are admitted. The confirmatory test remains |
| 30 | +unauthorized regardless of a single development result. |
| 31 | + |
| 32 | +## Batch 1 results |
| 33 | + |
| 34 | +Positive-control run: `runs/ecp8-batch1-positive-control-development/20260718T200136Z-ecp8-development`<br> |
| 35 | +Paired-control run: `runs/ecp8-batch1-control-development/20260718T200256Z-ecp8-development`<br> |
| 36 | +Intervention run: `runs/ecp8-batch1-intervention-development/20260718T200637Z-ecp8-development`<br> |
| 37 | +Test unsealed: **no** |
| 38 | + |
| 39 | +| Metric | Positive control | 14-bit control | 16-bit intervention | |
| 40 | +|---|---:|---:|---:| |
| 41 | +| Population train, mean | 100% | 79.5476% | 98.7566% | |
| 42 | +| Population train, worst link | 100% | 78.6551% | 98.6886% | |
| 43 | +| Population validation, mean | 100% | 70.2698% | 74.4629% | |
| 44 | +| Population validation, worst link | 100% | 64.8438% | 73.3398% | |
| 45 | +| Universal translator, validation | 100% | 75.1953% | 81.0547% | |
| 46 | +| Exact sender-message agreement | 100% | 93.74% | 77.56% | |
| 47 | +| Unique messages per sender | 15,360 | 12,048–12,255 | 15,095–15,135 | |
| 48 | +| Collision meanings per sender | 0 | 3,105–3,312 | 225–265 | |
| 49 | +| Message entropy | 13.91 bits | 13.45–13.48 bits | 13.87–13.88 bits | |
| 50 | +| Declared payload | 14 bits | 14 bits | 16 bits | |
| 51 | +| Registered development gate | pass | fail | **fail** | |
| 52 | + |
| 53 | +The positive control passes every validity requirement at 100%, proving that |
| 54 | +the fresh split remains fully solvable under the factorized 14-bit protocol. |
| 55 | +The paired weak-structure control selects step 7,400 and stops at the registered |
| 56 | +15,000-step minimum. The 16-bit intervention keeps improving, selects step |
| 57 | +16,200 and stops at step 21,200. |
| 58 | + |
| 59 | +The extra two bits have a large isolated effect. Mean train exactness improves |
| 60 | +by 19.21 percentage points and worst-link train by 20.03 points. Mean |
| 61 | +validation improves by 4.19 points, worst-link validation by 8.50 points and |
| 62 | +translator validation by 5.86 points. The minimum sender codebook gains 3,047 |
| 63 | +occupied messages and the final all-link hard-error replay pool falls from |
| 64 | +1,953 to 119. |
| 65 | + |
| 66 | +Capacity is nevertheless insufficient. The intervention retains 225–265 hard |
| 67 | +collisions per sender, so none is injective. Mean and worst-link validation are |
| 68 | +74.46% and 73.34%, below their registered 80% thresholds. The intervention |
| 69 | +passes both train thresholds, the translator threshold and both paired |
| 70 | +non-regression checks, but fails injectivity and both validation thresholds. |
| 71 | + |
| 72 | +The error pattern also distinguishes memorization from learned compositional |
| 73 | +structure. The 16-bit arm reaches near-perfect training and perfect texture |
| 74 | +validation, while held-out color and shape remain at 94.66% and 79.77% factor |
| 75 | +accuracy. Sender agreement falls from 93.74% to 77.56%, and sender topographic |
| 76 | +correlations fall from 0.761–0.771 to 0.693–0.700. The surplus code space lets |
| 77 | +independent senders separate most training meanings, but it does not induce a |
| 78 | +shared compositional organization for the withheld relation. |
| 79 | + |
| 80 | +All three post-hoc analyses verify 65 artifact hashes, 16,384 validation-only |
| 81 | +episodes, absent confirmatory-test metrics and valid hard messages. The global |
| 82 | +channel audit checks every intervention message for four symbols in `0..15` |
| 83 | +and exactly 16 declared bits. |
| 84 | + |
| 85 | +## Batch 1 decision |
| 86 | + |
| 87 | +ECP8-B1-I is a valid negative development result. Two surplus unallocated bits |
| 88 | +substantially reduce the hard-code bottleneck and improve every joint accuracy |
| 89 | +metric, but they do not produce an injective or sufficiently compositional |
| 90 | +protocol. The pure-capacity hypothesis is rejected, and the ECP-8 confirmatory |
| 91 | +split remains sealed. |
| 92 | + |
| 93 | +Do not continue with an 18-bit, 20-bit or longer fixed message search. A Batch |
| 94 | +2, if attempted, must keep the 16-bit channel and Batch 1 training fixed while |
| 95 | +preregistering exactly one structure-induction mechanism. The most direct |
| 96 | +candidate is the previously defined context-invariant transition objective, |
| 97 | +because Batch 1 now supplies the code capacity that was missing when that |
| 98 | +objective collapsed in ECP7-B8. It must be tested without a slot binding or |
| 99 | +additional simultaneous change. |
0 commit comments