@@ -26,7 +26,7 @@ The protocol uses no words or alphabet, but four meaning-free local symbols. For
 See [`docs/results-ecp6.md`](docs/results-ecp6.md) for the conclusion, [`docs/protocol-specification-ecp6.md`](docs/protocol-specification-ecp6.md) for the wire format and [`evidence/ecp6/report.md`](evidence/ecp6/report.md) for the compact confirmatory evidence.
 
 ECP-7 development now tests how much of that result depends on the explicit
-one-factor-per-slot architecture. Its first twenty sealed batches are valid
+one-factor-per-slot architecture. Its first twenty-one sealed batches are valid
 development results but none passes the full gate. Batch 15 established the
 strong position-aware base: at 30,000 optimization steps it reaches 83.46% train
 exactness, 82.59% validation and 83.37% translator validation while using
@@ -41,7 +41,10 @@ reached 82.04% validation plus a new-best 80.57% worst-link validation, but
 again failed train exactness and injectivity. The ECP-6 positive controls remain
 perfect. Batch 20 replayed population-hard training meanings and established
 new-best 83.45% mean and 82.13% worst-link validation, but again failed train
-exactness and injectivity. The ECP-7 confirmatory test remains sealed. See
+exactness and injectivity. Batch 21 restricted replay to errors shared by all
+16 links and improved validation again to 83.63% mean and 82.62% worst-link,
+but its target error pool grew and train exactness still failed. The ECP-7
+confirmatory test remains sealed. See
 [`docs/development-log-ecp7.md`](docs/development-log-ecp7.md).
 
 ## Build on this work
@@ -102,6 +105,7 @@ Do not translate those artifacts in place. New experiment configurations and all
 - [`config/ecp7-b18-development.yaml`](config/ecp7-b18-development.yaml) — the Batch 17 replay mechanism with final weight reduced to 0.1.
 - [`config/ecp7-b19-development.yaml`](config/ecp7-b19-development.yaml) — the Batch 18 replay mechanism decayed back to zero after step 20,000.
 - [`config/ecp7-b20-development.yaml`](config/ecp7-b20-development.yaml) — late ordinary task replay on population-hard training meanings.
+- [`config/ecp7-b21-development.yaml`](config/ecp7-b21-development.yaml) — Batch 20 replay restricted to meanings failed by all 16 population links.
 - [`config/ecp7-positive-control-development.yaml`](config/ecp7-positive-control-development.yaml) — ECP-6 positive control on the ECP-7 split.
 - [`config/ecp0.yaml`](config/ecp0.yaml) — machine-readable configuration from step 1.
 - [`config/ecp1.yaml`](config/ecp1.yaml) — frozen configuration of the population trial.
 
@@ -0,0 +1,13 @@
+extends: ecp7-b20-development.yaml
+
+experiment:
+  title: ECP-7 population-shared hard-meaning task replay
+  arm: population_shared_hard_meaning_task_replay
+  objective: >-
+    Test whether keeping the complete Batch 20 replay mechanism but mining only
+    training meanings failed by all 16 current sender-receiver links can repair
+    the population-shared error concentration without changing replay scale.
+
+training:
+  global_hard_meaning_replay:
+    minimum_failed_links: 16
@@ -41,7 +41,7 @@ python3.12 -m venv .venv
 .venv/bin/ecp6 --config config/ecp6.yaml validate
 ```
 
-Expected baseline: 109 passing tests and split sizes `14336/1024/1024`.
+Expected baseline: 111 passing tests and split sizes `14336/1024/1024`.
 
 ## 4. Reproduce ECP-6
 
@@ -71,7 +71,7 @@ This reruns a known experiment. It must not be presented as a new independent co
 
 ## 5. Continue ECP-7 scientifically
 
-Do not modify ECP-0 through ECP-6 or the completed ECP-7 Batch 1 through Batch 20
+Do not modify ECP-0 through ECP-6 or the completed ECP-7 Batch 1 through Batch 21
 configs. ECP7-B1-I collapsed to 130–139 hard messages. ECP7-B2-I improved its
 soft objective but collapsed further to 85–104 hard messages. ECP7-B3-I applied
 that loss to straight-through hard messages and improved to 585–972 messages,
@@ -126,11 +126,15 @@ population link. It reached 83.77% train, a new-best 83.45% mean and 82.13%
 worst-link validation, and 83.96% translator validation. The pool shrank but its
 remaining errors became more population-wide, so train thresholds and
 injectivity still failed. The confirmatory ECP-7 test remains sealed.
+ECP7-B21-I restricted the same replay budget to meanings failed by all 16
+population links. It reached a new-best 83.63% mean and 82.62% worst-link
+validation and 92.96% sender agreement, but mean train remained 83.71% and the
+target shared-error pool grew. The confirmatory ECP-7 test remains sealed.
 
 Continue with this sequence:
 
 1. Read `docs/research-design-ecp7.md` and `docs/development-log-ecp7.md`.
-2. Define exactly one ECP7-B21 intervention and its failure criterion.
+2. Define exactly one ECP7-B22 intervention and its failure criterion.
 3. Register the variant and immutable configuration hashes before training.
 4. Add tests for every new invariant.
 5. Use only `smoke` and `develop`; keep the ECP-7 test split sealed.
@@ -143,25 +147,27 @@ Continue with this sequence:
 
 ## 6. Recommended next experiment
 
-Batch 20 shows that ordinary task replay on any-link errors improves
-generalization and cross-link balance, but mostly removes link-specific errors.
-At its selected checkpoint, 1,937 of 2,848 hard meanings fail for all 16 links,
-up from 1,469 of 3,406 at the Batch 15 checkpoint. The clean next question is
-whether the fixed replay budget should target only those population-shared
-errors. A clean ECP7-B21 progression is:
-
-- inherit the complete Batch 20 implementation and Batch 15 base;
-- change only the mining predicate from at least one failed link to all 16
-  failed links;
-- keep weight `0.25`, the 64-meaning uniform replay batch, 200-step refresh,
-  seed-derived sampler, warmup and horizon unchanged;
+Batch 21 directly targets population-shared errors, but its all-link pool grows
+from 1,745 to 1,997 at the selected checkpoint while validation and sender
+agreement improve. This suggests that joint replay updates align receivers and
+senders around persistent ambiguous codes rather than forcing senders to move.
+The clean next question is whether replay gradients must be sender-only. A clean
+ECP7-B22 progression is:
+
+- inherit the complete Batch 21 predicate, Batch 20 replay mechanism and Batch
+  15 base;
+- change only replay-loss gradient routing so receiver parameters are treated as
+  fixed during the additional replay forward pass;
+- keep ordinary base-task updates to both senders and receivers unchanged;
+- keep weight `0.25`, the 64-meaning replay batch, 200-step refresh,
+  seed-derived sampler, warmup, temperature and horizon unchanged;
 - continue to mine and replay training meanings only;
-- measure train exactness, injectivity, validation composition and new-reader
-  induction, and rerun the frozen ECP-6 positive control.
+- measure shared-error pool size, train exactness, injectivity, validation and
+  new-reader induction, and rerun the frozen ECP-6 positive control.
 
-Do not combine the predicate change with another coefficient, graded sampling,
-collision pressure, a new architecture, factor-specific weights, optimizer
-changes, longer training, variable-length messages or negotiation.
+Do not combine gradient routing with another predicate, coefficient, graded
+sampling, collision pressure, a new architecture, factor-specific weights,
+optimizer changes, longer training, variable-length messages or negotiation.
 
 ## 7. Definition of done for any contribution
 
 
@@ -28,6 +28,7 @@ this log.
 | ECP7-B18-I | Intervention | B17 replay with final weight 0.1 | Position-aware MLP | completed; gate failed |
 | ECP7-B19-I | Intervention | B18 with replay decayed to zero after step 20,000 | Position-aware MLP | valid negative; validation 82.04%, train and injectivity fail |
 | ECP7-B20-I | Intervention | Late global replay of population-hard training meanings | Position-aware MLP | valid negative; new-best validation 83.45%, train and injectivity fail |
+| ECP7-B21-I | Intervention | B20 replay restricted to meanings failed by all 16 links | Position-aware MLP | valid negative; new-best validation 83.63%, shared-error pool grows |
 
 ## Batch 1 preregistration
 
@@ -1566,3 +1567,106 @@ the same registered replay budget exclusively on population-shared errors. No
 coefficient, replay batch, refresh, schedule, architecture, base loss,
 optimizer, horizon, data, translator or threshold change may accompany that
 predicate test.
+
+## Batch 21 preregistration
+
+Preregistered: July 18, 2026 at 16:05:33 UTC<br>
+Seed: `11`<br>
+Maximum population steps: `30,000`<br>
+Minimum selection steps: `15,000`<br>
+Early-stopping patience: `5,000` steps<br>
+Hard-meaning predicate: all `16` current population links have an incorrect factor<br>
+Replay weight: `0` through step `15,000`, linear to `0.25` at step `20,000`, then hold<br>
+Replay batch size: `64` meanings, uniform with replacement<br>
+Training-only pool refresh interval: `200` steps<br>
+Replay sampler seed offset: `+1211`<br>
+Raw configuration SHA-256: `d2d8d8bf3d87aef42b533bfd33eea4033af99444ec00b1cffe3d6f63c5970a5b`<br>
+Effective configuration SHA-256: `913fa521edb8138b8e2319ef52b1ee1533ce13592f53303250d97ef217af0f5b`<br>
+Split-SHA-256: `4947058c75ab07cb43a87eb82776b12cb2a7e2eeba7114de110d3b852cbc64cd`<br>
+Test unsealed: **no**
+
+ECP7-B21-I inherits Batch 20 and changes only the failed-link threshold from the
+implicit default `1` to the complete population count `16`. The base random task
+sequence, replay sampler, replay scale and every other setting remain unchanged.
+This directly tests whether the B20 pool's population-wide error concentration
+requires a narrower curriculum. The positive control is rerun, and no
+alternative threshold or accompanying intervention is admitted into this batch.
+
+## Batch 21 results
+
+Positive-control run: `runs/ecp7-batch21-control-development/20260718T160742Z-ecp7-development`<br>
+Intervention run: `runs/ecp7-batch21-intervention-development/20260718T160742Z-ecp7-development`<br>
+Test unsealed: **no**
+
+| Metric | Positive control | ECP7-B21-I |
+|---|---:|---:|
+| Population train, mean | 100% | **83.7062%** |
+| Population train, worst link | 100% | **82.7916%** |
+| Population validation, mean | 100% | **83.6304%** |
+| Population validation, worst link | 100% | **82.6172%** |
+| Universal translator, validation | 100% | **83.6426%** |
+| Exact sender-message agreement | 100% | **92.96%** |
+| Unique messages per sender | 15,360 | 12,720–13,200 |
+| Collision meanings per sender | 0 | 2,160–2,640 |
+| Message entropy | 13.91 bits | 13.54–13.61 bits |
+| Development gate | pass | **fail** |
+
+The positive control again passed every gate at 100%. ECP7-B21-I establishes
+new-best ECP-7 mean validation, worst-link validation, worst-link train and
+sender agreement. It passes validation and translator thresholds, but mean and
+worst-link train remain far below their registered thresholds and every sender
+remains non-injective. The joint gate fails and confirmatory access remains
+unauthorized.
+
+All 76 optimization records through step 15,000 match Batch 20 exactly across
+the 30 shared non-pool fields. The registered predicate produces the expected
+different first pool at step 15,000: 1,745 meanings fail for all 16 links. That
+pool briefly falls to 1,504 at step 15,600, then grows as replay weight rises,
+reaching 2,031 at full weight on step 20,000.
+
+The selected checkpoint was step 29,400 with 1,997 all-link failures. Task loss
+was `0.07461`, hard-meaning replay loss `0.11821`, replay weight `0.25`, and
+utilization loss `-0.75199`. The replay contribution was approximately `0.0296`,
+well below the ordinary task loss. The run reached step 30,000, finishing with
+2,016 all-link failures and 83.18% validation.
+
+Compared with Batch 20, mean train fell by 0.07 percentage points and translator
+validation by 0.32 points. Mean validation improved by 0.18 points, worst-link
+validation by 0.49 points, worst-link train by 0.08 points and sender agreement
+by 0.57 points. The narrower curriculum therefore continues the cross-link
+alignment effect without improving mean known-meaning reconstruction.
+
+A selected-checkpoint training diagnostic confirms the tradeoff. Batch 21 has
+2,755 meanings failed by at least one link, fewer than Batch 20's 2,848, but
+1,997 fail for all links, more than Batch 20's 1,937. A hard meaning now fails
+for 13.57 links on average, up from 13.07. The all-link replay pool grows 14.4%
+from its step-15,000 value and its selected hard loss is `0.34190`. Directing
+more ordinary joint sender-receiver optimization at shared errors therefore
+reinforces population alignment around the remaining ambiguous mappings rather
+than separating them.
+
+Selected sender codebooks contain 12,840, 12,870, 12,720 and 13,200 unique
+messages, with 2,520, 2,490, 2,640 and 2,160 collision meanings. Population
+validation factor accuracies were `[99.95%, 99.95%, 98.44%, 85.29%]`;
+universal-translator factor accuracies were
+`[100%, 100%, 98.44%, 85.21%]`. Population color and shape are effectively
+complete, while the same size and texture bottlenecks remain.
+
+Both sealed analyses report 65 matching artifact hashes, 16,384 validation-only
+episodes, no confirmatory-test keys, valid local alphabets, and exactly 14
+declared bits for every logged message.
+
+## Batch 21 decision
+
+ECP7-B21-I is a valid negative result. All-link replay slightly improves the
+strongest validation result, but directly falsifies the hypothesis that a
+narrower shared-error curriculum will repair train exactness: its target pool
+grows and becomes more dominant. The confirmatory split remains sealed.
+
+A future ECP7-B22 may keep Batch 21's predicate, replay batch, weight, schedule
+and every Batch 15 setting while stopping replay-loss gradients into receiver
+parameters. Ordinary base-task updates would continue training both sides; only
+the additional all-link replay loss would update senders through temporarily
+fixed receivers. This isolates whether joint coadaptation prevents ambiguous
+sender codes from moving toward already learned decoder distinctions. No other
+parameter or mechanism may change in that test.
@@ -888,3 +888,45 @@ sealed-test requirement fails. No alternative predicate, score, replay weight,
 batch size, refresh, schedule, architecture, base loss, optimizer, duration,
 data, translator or seed may be tried inside Batch 20. Failure is recorded
 without confirmatory test access.
+
+## Batch 21 preregistration — Population-shared hard-meaning replay
+
+Status: **preregistered for sealed development**<br>
+Preregistered: July 18, 2026 at 16:05:33 UTC<br>
+Configuration: `config/ecp7-b21-development.yaml`<br>
+Raw configuration SHA-256: `d2d8d8bf3d87aef42b533bfd33eea4033af99444ec00b1cffe3d6f63c5970a5b`<br>
+Effective configuration SHA-256: `913fa521edb8138b8e2319ef52b1ee1533ce13592f53303250d97ef217af0f5b`
+
+Batch 20 reduced the any-link hard pool and established the strongest ECP-7
+validation result, but its remaining error became more population-wide. At the
+selected checkpoint, 1,937 of 2,848 hard meanings failed for all 16 links,
+compared with 1,469 of 3,406 at the selected Batch 15 checkpoint. Batch 21 tests
+whether replay should spend its fixed budget exclusively on those shared
+errors.
+
+ECP7-B21-I inherits the complete Batch 20 implementation, Batch 15 architecture,
+ordinary random task batch, utilization objective, optimizer, data, temperature
+schedule, hard-replay coefficient, batch size, sampler, refresh, warmup,
+30,000-step horizon, selection, translator and thresholds. The sole change is:
+
+- a training meaning enters the replay pool only when all 16 current
+  sender-receiver links reconstruct at least one factor incorrectly, instead of
+  when at least one link fails.
+
+Mining still uses hard predictions over training meanings only. Replay still
+samples 64 pool meanings uniformly with replacement, applies the ordinary
+all-link reconstruction loss, warms from weight zero at step 15,000 to `0.25`
+at step 20,000, and then holds. The complete trajectory through step 15,000
+must match Batch 20 and Batch 15 exactly.
+
+The batch contains exactly two seed-11 runs:
+
+1. the unchanged ECP-6 positive control;
+2. ECP7-B21-I with population-shared hard-meaning replay.
+
+All existing validity and development gates remain unchanged. The intervention
+fails if any train, validation, translator, injectivity, channel-integrity or
+sealed-test requirement fails. No alternative failed-link threshold, graded
+score, coefficient, replay batch, refresh, schedule, architecture, base loss,
+optimizer, duration, data, translator or seed may be tried inside Batch 21.
+Failure is recorded without confirmatory test access.
@@ -111,6 +111,12 @@ shrunk but concentrated into population-wide failures, so train thresholds and
 injectivity still failed. Exact identities and hashes are in
 `batch20-manifest.json`.
 
+Batch 21 restricted the unchanged replay budget to meanings failed by all 16
+population links. It reached a new-best 83.63% mean and 82.62% worst-link
+validation plus 92.96% sender agreement, but mean train remained 83.71% and the
+target shared-error pool grew. Exact identities and hashes are in
+`batch21-manifest.json`.
+
 Reproduce the development arms with:
 
 ```bash