research-design-ecp7.md

Research Design ECP-7 — Weakening the Factor-Slot Bias

Status: development closed; confirmatory test remains sealed
Preregistered: July 18, 2026
Intervention configuration: config/ecp7-development.yaml
Positive-control configuration: config/ecp7-positive-control-development.yaml

The confirmatory ECP-7 test split has not been accessed. This document defines the first development batch only. No frozen confirmatory configuration exists yet.

Question

Does exact, compositional and transferable communication survive when sender and receiver architectures no longer encode a one-factor-per-slot binding, while the world, four-slot wire capacity, population and training budget remain fixed?

This is a narrower and more difficult question than whether a neural network can reproduce the ECP-6 code. The intervention must learn how the four meaning factors jointly determine a message, and a generic sequence receiver must learn how the entire message determines every factor.

Single intervention

ECP-7 changes one architectural assumption: factor-local processing.

The intervention sender:

1. embeds all four input factors into one joint context;
2. emits four symbols autoregressively with a GRU;
3. is constrained only by per-position alphabet sizes [16,16,8,8];
4. has no factor-specific output heads, factor codebook or predefined factor-to-slot binding.

The intervention receiver and universal translator are generic GRU sequence encoders with four classification heads. Slot-binding calibration and the factor/atom consensus losses are disabled.

The unchanged ECP-6 factorized sender and receiver form a positive control on the same new split. The positive control tests whether a failure comes from the intervention rather than the split or execution pipeline.

Capacity and remaining inductive bias

The world remains 16 × 16 × 8 × 8 = 16,384 = 2^14 equiprobable meanings. The message has four categorical positions with alphabet sizes [16,16,8,8], so the nominal widths [4,4,3,3] occupy exactly 14 bits.

ECP-7 therefore removes factor-slot binding, but it is not an unconstrained or natural language experiment. It retains:

• categorical, fully observed inputs;
• a synthetic product world;
• a fixed four-position message frame;
• position-specific alphabet capacities;
• supervised task rewards and a fixed training schedule.

Any conclusion must be limited to this setting.

Sealed data split

Validation uses the unused cyclic color-shape matching with shift +10; the confirmatory test uses shift +11. Both are disjoint from the ECP-6 validation and test matchings (+8 and +9).

Split	Meanings	Color-shape pairs	Permitted during development
Train	14,336	224	yes
Validation	1,024	16	yes
Confirmatory test	1,024	16	no

Split-SHA-256: 4947058c75ab07cb43a87eb82776b12cb2a7e2eeba7114de110d3b852cbc64cd.

First development batch

The first batch contains exactly two runs, both with seed 11 and at most 5,000 population-training steps:

1. ecp6_factorized_positive_control;
2. bounded_autoregressive_intervention.

Only train and validation data may be evaluated. Smoke runs may precede the full runs to verify mechanics, but may not be used to tune the intervention. No alternative architecture or hyperparameter variant is admitted into this batch.

The positive control validates the batch when both its mean and worst-link validation exactness are at least 99%. If it misses that gate, the batch is inconclusive rather than evidence against the intervention.

When the positive control is valid, the intervention passes the development gate only if all of the following hold:

1. known population exactness is at least 97% on average and 95% for the worst sender-receiver link;
2. mean validation exactness is at least 80%;
3. universal-translator validation exactness is at least 70%;
4. every logged symbol respects its position-specific alphabet boundary;
5. each sender assigns a unique message to every accessible train and validation meaning;
6. artifact and split integrity checks pass and no confirmatory test episode or metric is written.

If the positive control passes and the intervention misses any gate, the first ECP-7 architecture is recorded as a negative development result. The test stays sealed. A later variant, if justified, must be numbered and preregistered before training.

Passing the development gate is not confirmation. It only permits selection of a single frozen ECP-7 design and confirmatory criteria in a later commit.

Recorded configuration identities

Arm	Raw file SHA-256	Effective configuration SHA-256
Intervention	7863ffbb4fdfdf43d28fed897d75989153e72ed0c432eaadbddc835f27c77b8f	e5f6ea5042dc59b5b10c5860ce7a672274b6b3ba887656de6f56cb8192181561
Positive control	4d5904a894b5d854831049cb38331c2e42aabc5da844de3d95e665d941ca6c8f	db323732a1160859835c930f2ffa578df363ddf146537aa0d4ee2496e2b9744d

Batch 2 preregistration — Factor-agnostic code utilization

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b2-development.yaml
Raw configuration SHA-256: 95f402cc9b5f151b017d2c92ad4954d463aa73e0a13e9e1df84393f12a570cfc
Effective configuration SHA-256: fc419eff2065f6259553d7fb5bbbd7de396be055a66c860884fe1c193158ea0f

Batch 1 showed severe code collapse: the intervention used only 130–139 messages and approximately 6.45 bits of entropy. Batch 2 asks whether one factor-agnostic pressure for using the available code space can prevent that collapse without restoring a factor-to-slot binding.

For each sender, the new loss observes only its relaxed message distributions. It receives no factor labels, factor identity, target slot or desired code. For each slot it minimizes normalized conditional entropy and maximizes normalized marginal entropy. It also minimizes normalized pairwise mutual information between slots, discouraging four balanced slots from carrying duplicate information.

In compact form:

L_util = mean(H(slot | input) - H(slot)) + mean(pairwise normalized MI)

Each entropy is normalized by that slot's maximum entropy. The loss weight is fixed at 1.0, warms up linearly for 400 steps, uses relaxed temperature 1.0, and gives the independence term weight 1.0. An ideal deterministic, balanced, pairwise-independent code has loss -1; a deterministic collapsed code and a fully uncertain uniform code both have loss 0.

No architecture, task loss, population setting, data split, capacity, early stopping rule or translator setting changes from ECP7-B1-I. This objective adds a product-code utilization bias, but it does not specify which semantic factor belongs in any slot.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B2-I with factor-agnostic code utilization.

The positive-control validity gate and intervention development gate are unchanged from Batch 1. In particular, B2 must reach the registered population, validation and translator thresholds, produce an injective code over all 15,360 accessible meanings for every sender, respect every local slot boundary, and write no confirmatory test metric or episode. Failure is recorded without opening the test split. No weight, warmup, temperature or independence variant may be tried inside Batch 2.

Batch 3 preregistration — Straight-through hard utilization

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b3-development.yaml
Raw configuration SHA-256: 8a1fb70caeb16ee2e9f912fe37508d02c25d2b00e4ec887546d143fdc099ad39
Effective configuration SHA-256: 62cb6965f619a72905af2ba652e1feea3e718cad1201cdd4066063de4aef7756

Batch 2 improved its loss on relaxed distributions while the evaluated hard code collapsed further. Batch 3 tests whether that relaxation gap closes when the unchanged utilization formula receives the same straight-through one-hot messages already consumed by the receivers during task training.

This is the only Batch 3 change. The forward values used to calculate entropy and pairwise mutual information are discrete one-hot symbols. Gradients still flow through the straight-through Gumbel estimator. The architecture, factor inputs, loss weight 1.0, 400-step warmup, temperature 1.0, independence weight 1.0, task loss, population, split, capacity, training duration, early stopping and translator remain unchanged from Batch 2.

The utilization loss still receives no factor labels, factor identities or desired slot mapping. It does add an explicit pressure on the empirical hard symbol marginals and pairwise slot dependence of each training batch.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B3-I with straight-through hard code utilization.

The positive-control validity gate and intervention development gate remain unchanged. B3 must pass population, validation, translator, injectivity, channel and sealed-test checks together. A failure remains a development result and does not authorize confirmatory test access. No alternative estimator, weight or training schedule may be tried inside Batch 3.

Batch 4 preregistration — Direct joint-message collisions

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b4-development.yaml
Raw configuration SHA-256: 6229c5cd13c796a9ce297b9c38cc2ff1a2090a2a22a5678ec7afe38aa0bd6955
Effective configuration SHA-256: 4cffdb040ba358d7416e0643595c4e11fadc0fdffd5c0120761c4890a8c3f10e

Batch 3 improved marginal hard utilization but retained thousands of complete message collisions. Batch 4 asks whether directly penalizing those joint collisions produces an injective and compositional code without a semantic factor-to-slot binding.

For every sender and minibatch, the new loss is the average number of other distinct inputs with the same complete straight-through message:

L_joint = (1/B) × sum_i sum_j!=i 1[x_i != x_j] × product_s <m_i,s, m_j,s>

Because the forward messages are one-hot, the slot dot product is one exactly when both symbols match, and the product is one exactly when the full messages match. Identical sampled inputs are excluded; no factor label, factor identity or desired code is provided. Straight-through gradients flow back through the hard messages.

The joint-collision weight is fixed at 1.0 with a 400-step linear warmup. This is the only Batch 4 addition. The complete Batch 3 utilization objective, architecture, task loss, population, split, 14-bit channel, 5,000-step limit, early stopping and translator remain unchanged.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B4-I with the added direct joint-collision loss.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative collision weight, batch size, architecture or duration may be tried inside Batch 4.

Batch 5 preregistration — Straight-through sender consensus

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b5-development.yaml
Raw configuration SHA-256: 3333e3e150e242b0984e79a2923ffe94d68cee7d004ae51f4914097917503db7
Effective configuration SHA-256: 4c2648efa2719d1cd3e0bdf96a2cbea21ea697280b74bbce82ba45371bbc0ed9

Batch 4 regressed relative to Batch 3, so Batch 5 returns to the stronger ECP7-B3-I base and does not retain the joint-collision loss. Batch 3's four senders used substantially more of the hard code space than earlier batches, but their complete-message agreement on the accessible meanings was only approximately 2.47%. Batch 5 asks whether one shared factor-agnostic convention helps generic receivers learn the protocol.

For the same training minibatch, the new objective compares each unordered pair of senders. It averages their straight-through one-hot symbol disagreement over inputs and message slots, then averages over all sender pairs:

L_sender = mean_(a<b,x,s) [1 - <m_a(x,s), m_b(x,s)>]

The forward value is zero when two senders emit the same hard symbol and one when they differ. Gradients flow through the straight-through Gumbel estimator. The objective receives no factor labels, factor identities, target slots, desired symbols or privileged codebook. Per-sender hard utilization from Batch 3 remains active to oppose agreement through collapse.

The sender-consensus weight is fixed at 1.0 with a 400-step linear warmup. This is the only Batch 5 addition. Architecture, task loss, Batch 3 hard utilization, population, split, 14-bit channel, 5,000-step limit, early stopping and translator remain unchanged. The rejected Batch 4 collision loss is absent.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B5-I with straight-through sender consensus on the Batch 3 base.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative consensus weight, warmup, reduction, architecture or duration may be tried inside Batch 5.

Batch 6 preregistration — Normalized factor minimax

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b6-development.yaml
Raw configuration SHA-256: 29fffcbfbc81c96e5765594ba67eea8143ff6d45f77defc8ce3acaf1a628b6f9
Effective configuration SHA-256: 3e54202ce906426a2a576c54ba9a8a4d7d65711dd8bc9376c2d6551244db046a

Batch 5 increased sender agreement by creating a shared collapse, so Batch 6 returns to ECP7-B3-I and retains neither the Batch 4 collision loss nor the Batch 5 sender-consensus loss. Across B3–B5, color and especially shape remain the weak factors while size and texture are much easier. Batch 6 asks whether the mean task objective lets the system neglect its hardest semantic dimension.

For every sender-receiver pair and training minibatch, let CE_f be the cross-entropy for output factor f and K_f its number of values. The new objective is:

L_minimax = max_f [CE_f / log(K_f)]

Dividing by the uniform-guess cross-entropy makes difficulty comparable across 16-way and 8-way factors. The original mean factor-reconstruction loss remains active; the minimax term adds gradient pressure from the currently worst relative factor. It uses the same supervised output factors already present in the task, but receives no channel-slot identity, factor-to-slot assignment, desired symbol or privileged codebook.

The minimax weight is fixed at 1.0 with a 400-step linear warmup. This is the only Batch 6 addition. The Batch 3 straight-through utilization objective, architecture, population, split, 14-bit channel, 5,000-step limit, early stopping and translator remain unchanged.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B6-I with normalized factor minimax on the Batch 3 base.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative minimax weight, normalization, reduction, architecture or duration may be tried inside Batch 6.

Batch 7 preregistration — Bounded parallel sender

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b7-development.yaml
Raw configuration SHA-256: e885c6892b9da0b4af89dce96c00037a3813881499f20dd1b21e7a06ca4cf41d
Effective configuration SHA-256: 1ca8b5c618bc9193b60745040fa7f8a91e31a92e5786d2eab0597874938dee9e

Batches 4–6 added three different training pressures to the autoregressive Batch 3 sender, and every addition reduced its strongest result. Batch 7 therefore returns to the exact Batch 3 loss set and changes one architectural assumption: symbols no longer depend sequentially on earlier symbols.

The bounded parallel sender embeds all four meaning factors, concatenates them, and projects them into one shared joint context. Four independent output heads read that same context simultaneously and are limited only by the registered slot alphabet sizes [16,16,8,8]. Straight-through Gumbel sampling and the Batch 3 hard-utilization loss remain unchanged.

The heads identify channel positions, not semantic factors. There is no factor-specific projection, factor-local message head, recurrent state, factor-to-slot binding, binding matrix, desired symbol or semantic codebook. The generic sequence receiver and universal translator are unchanged. This tests whether sequential generation itself was an optimization bottleneck; it does not remove the fixed four-position channel frame.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B7-I with bounded parallel generation and the Batch 3 loss set.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No receiver change, hidden-size change, extra loss, longer duration or alternative parallel architecture may be tried inside Batch 7.

Batch 8 preregistration — Algebraic context invariance

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b8-development.yaml
Raw configuration SHA-256: 3eac0d25bf636d4a1aeb6161ad479a4a78f0a3e88788a0274c154c21ff192829
Effective configuration SHA-256: 6e8f9ea0e968693bd9f67a4ca1180ab447c6e2d6c2fbc6b9f9a66582aa02fa3c

Batch 7 greatly improved training exactness and code-space use, but those distinctions did not compose across the held-out color-shape matching. Batch 8 retains the entire B7 architecture and loss set, then activates one existing context-invariance objective that was previously disabled.

Every training quadruple contains meanings A, B, A', and B'. The transitions A→A' and B→B' change the same factor value in two different contexts, and all four meanings belong to the training split. For relaxed message distributions m, the objective is:

L_alg = mean ||(m(A') - m(A)) - (m(B') - m(B))||²

This encourages the same atomic semantic change to produce the same message displacement regardless of the other factor values. It uses factor identities only to construct matching transitions. It does not specify a target message, symbol, channel slot, factor-to-slot binding or codebook.

The loss weight is fixed at 1.0 with an 800-step linear warmup, relaxed temperature 1.0, 32 quadruples per update and a deterministic pool of 8,192 training-only quadruples. These values match the dormant inherited settings; only enabled and weight change from B7. The parallel sender, generic receiver, hard-utilization loss, channel, split, training duration, selection and translator remain unchanged.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B8-I with algebraic context invariance on the B7 parallel sender.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative quadruple sampling, weight, warmup, receiver, architecture or duration may be tried inside Batch 8.

Batch 9 preregistration — Position-aware MLP decoder

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b9-development.yaml
Raw configuration SHA-256: 1ceeb80a3657078167d6d2ebacc1d57546d41ce53860148b9dfe7941b855c3fc
Effective configuration SHA-256: d76fd03cfdffc97cdc4ded755868db93c6d77df62c14fb9fda3627fc2a3ad591

Batch 7 messages changed when shape changed, and a fresh translator extracted more shape information than the jointly trained receiver population. Batch 8's sender-side context loss suppressed the stronger Batch 7 code. Batch 9 returns to the complete B7 sender and loss set, then changes one shared decoder-family assumption: messages are no longer compressed into a GRU final state.

The position-aware MLP decoder embeds each of the four received symbols with one shared symbol embedding, concatenates the four embeddings in channel order, projects that complete vector into one hidden representation, and applies four factor-classification heads. Every factor head reads the same representation of all message positions.

The decoder preserves positional information but has no recurrent order bias, factor-specific input path, binding matrix, hard slot selection or factor-to-slot assignment. The post-freeze universal translator uses the same decoder family so transfer is evaluated without giving it a different architectural advantage. Its isolation metadata is unchanged.

This decoder-family replacement is the only Batch 9 change. The B7 parallel sender, Batch 3 hard-utilization loss, task loss, population, channel, split, duration, selection and transfer training remain unchanged. The rejected Batch 8 algebraic objective is disabled as in B7.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B9-I with the shared position-aware MLP decoder on the B7 sender.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No sender, loss, hidden-size, duration or alternative decoder variant may be tried inside Batch 9.

Batch 10 preregistration — Extended optimization horizon

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b10-development.yaml
Raw configuration SHA-256: ac1faed6538fcd7bcdd5dfe16a952ec317286629dcb3ecee7e9ec94e73939a2d
Effective configuration SHA-256: 415092b9215f7544f7a54f4b3d954a3a1e69ad90c43781c93dbcf8d3892cc76b

Batch 9 selected its final registered checkpoint at step 5,000 while task loss and exactness were still improving. It used about 11,000 high-entropy messages, its largest collision bucket contained only 8–9 meanings, and no factor remained near chance. Batch 10 therefore tests one explanation: the B9 design was underoptimized at the registered stopping horizon.

ECP7-B10-I keeps the complete B9 architecture, sender, position-aware receiver, translator, losses, optimizer, batch size, learning rate, data, evaluation interval, channel, thresholds and seed. It changes only the population optimization horizon and its associated checkpoint-selection window:

• maximum population steps increase from 5,000 to 15,000;
• minimum selection steps increase from 2,000 to 5,000;
• early-stopping patience increases from 1,600 to 3,000 steps;
• the original geometric temperature schedule remains exactly unchanged from steps 1 through 5,000, then holds its registered final value of 0.8.

Translator training remains unchanged. The schedule hold prevents the longer horizon from also becoming a slower-annealing intervention. Smoke runs are mechanical only and may not authorize tuning.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B10-I with the extended B9 population optimization horizon.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative horizon, patience, annealing schedule, architecture, loss, optimizer or seed may be tried inside Batch 10.

Batch 11 preregistration — Code-utilization weight decay

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b11-development.yaml
Raw configuration SHA-256: 6316b7361a7c4947089b7e33d189a80c4e37739650c75e961fcec7628b812c70
Effective configuration SHA-256: db05737365381f73afada0462de1f6ae111b92b96589364e29f05befac701519

Batch 10 passed the translator threshold and reached 72.98% validation, but validation plateaued after approximately step 12,000. At the selected final checkpoint, the unweighted utilization term had magnitude 0.7548 while task loss was 0.0765. The code nevertheless retained 12,358–12,906 unique messages, 13.48–13.57 bits of entropy, sender agreement above 92%, and collision buckets of at most four meanings.

Batch 11 tests one explanation: after the high-entropy code is established, full utilization pressure dominates the remaining task objective and prevents the last color and texture distinctions from resolving.

ECP7-B11-I keeps the complete B10 model, data, optimizer, duration, checkpoint selection, temperature schedule, translator, task loss and utilization formula. It changes only the coefficient applied to that existing utilization term:

• the original warmup and weight 1.0 are unchanged through step 5,000;
• the coefficient then decays linearly from 1.0 at step 5,000 to 0.1 at step 15,000;
• the final nonzero coefficient preserves an anti-collapse signal while making its expected weighted magnitude comparable to B10's remaining task loss.

This schedule reproduces every B10 optimization setting and weight through the original B9 horizon. Smoke runs are mechanical only and may not authorize tuning.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B11-I with utilization-weight decay on the B10 base.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative final weight, decay shape, decay boundary, duration, architecture, loss, optimizer or seed may be tried inside Batch 11.

Batch 12 preregistration — Deep position-aware decoder

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b12-development.yaml
Raw configuration SHA-256: 4cc37488453b756fb43b4782d0e3e93cf051815a367d1852047fcc34fa29d5ce
Effective configuration SHA-256: 798a6380ab144e5db63c8177d3dd70edf76e66b660b3cbb355f5beabe86b27b2

Batch 11 showed that reducing late utilization pressure worsened task accuracy, code use and sender agreement. Batch 10 therefore remains the base. Its shared one-hidden-layer decoder and independently trained translator both retained a texture weakness, while color was the other residual population error. Batch 12 tests whether the generic decoder is still too shallow for the entangled code.

ECP7-B12-I returns to the complete B10 training and loss schedule. It keeps the shared symbol embedding, flattened four-position input and first generic hidden projection, then adds exactly one second hidden_dim × hidden_dim projection with tanh. All four factor heads read that same second shared representation. The isolated translator uses the identical deeper family.

The decoder still has no factor-specific input path, factor-specific hidden projection, binding matrix, slot selector or hard factor-to-slot assignment. Hidden width, sender, losses, optimizer, data, 15,000-step duration, original 5,000-step temperature annealing, selection window and thresholds remain unchanged from B10. Smoke runs are mechanical only and may not authorize tuning.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B12-I with the deeper shared position-aware receiver and translator.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative depth, activation, width, residual connection, sender, loss, duration, optimizer or seed may be tried inside Batch 12.

Batch 13 preregistration — Deep bounded-parallel sender

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b13-development.yaml
Raw configuration SHA-256: 9b57b1ae0387483594a8171a1535c0e9a82fc988708d74d0300a775747ea7210
Effective configuration SHA-256: 661a0e286eb986a1b907df5e30dc416a1aefd5e5263056ccc12092c6beb3002d

Batch 12 rejected extra receiver and translator depth. Batch 10 remains the base, but its bounded parallel sender still constructs all four message slots from one shallow shared context. Batch 13 tests the symmetric remaining capacity hypothesis: the sender context is too shallow to construct the residual color and texture distinctions.

ECP7-B13-I returns to the complete B10 model, receiver, translator, training and loss schedule. It keeps the shared factor embeddings and first generic context projection, then adds exactly one second hidden_dim × hidden_dim projection with tanh. All four slot heads read that same second shared context.

The sender still has no factor-specific hidden path, binding matrix, hard slot selector or factor-to-slot assignment. Hidden width, receiver, translator, losses, optimizer, data, 15,000-step duration, original 5,000-step temperature annealing, selection window and thresholds remain unchanged from B10. Smoke runs are mechanical only and may not authorize tuning.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B13-I with the deeper shared bounded-parallel sender.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative sender depth, activation, width, residual connection, decoder, loss, duration, optimizer or seed may be tried inside Batch 13.

Batch 14 preregistration — Late learning-rate decay

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b14-development.yaml
Raw configuration SHA-256: 867e42a0a8c530a4bd1d42be68f2846b66ce368644e993be67f82b39c8caec04
Effective configuration SHA-256: f44a36695c59bd532d7372c6cb3fb5b6eeed06b5bc10f04e96603926f7a9f37e

Batches 12 and 13 rejected additional generic decoder and sender depth. Batch 11 rejected reducing late utilization pressure. Batch 10 remains the base: it continued to improve late, but its validation score fluctuated around a plateau while AdamW retained learning rate 0.001 through step 15,000.

Batch 14 tests one optimization-stability hypothesis: smaller late parameter updates can refine the established high-entropy code without changing its architecture, temperature or objective balance.

ECP7-B14-I keeps the complete B10 model, losses, optimizer type, data, temperature schedule, duration, selection window and translator. Learning rate remains exactly 0.001 through step 5,000, then decays linearly to 0.0001 at step 15,000. No other AdamW or training setting changes.

This reproduces the full B9 trajectory before the learning-rate schedules diverge. Smoke runs are mechanical only and may not authorize tuning.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B14-I with late learning-rate decay on the B10 base.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative final rate, decay shape, decay boundary, architecture, loss, duration, optimizer or seed may be tried inside Batch 14.

Batch 15 preregistration — 30,000-step optimization horizon

Status: preregistered for sealed development
Preregistered: July 18, 2026
Configuration: config/ecp7-b15-development.yaml
Raw configuration SHA-256: 2b4ed0979cb1f6a6aac0409e2624bf982fd020306492706f4aa0e6b3a8c9d15d
Effective configuration SHA-256: d60b8a9aebe75514560b35790a3de9674990724fff0a9c52283c91ff429a487b

Batch 10 selected its final 15,000-step checkpoint. Batch 14 showed that decaying late learning rate harms validation, while Batches 11–13 rejected loss removal and added capacity. Batch 15 therefore tests one remaining optimization hypothesis: the constant-rate B10 trajectory needs more time.

ECP7-B15-I keeps the complete B10 architecture, sender, receiver, translator, losses, AdamW settings, data, temperature schedule, evaluation cadence and thresholds. It changes only the population optimization horizon and associated selection window:

• maximum population steps increase from 15,000 to 30,000;
• minimum selection steps increase from 5,000 to 15,000;
• early-stopping patience increases from 3,000 to 5,000 steps;
• learning rate remains constant at 0.001;
• temperature still reaches 0.8 at step 5,000 and holds thereafter.

Translator training remains unchanged. Smoke runs are mechanical only and may not authorize tuning.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B15-I with the extended B10 population optimization horizon.

All existing validity and development gates remain unchanged. Failure is recorded without confirmatory test access. No alternative horizon, patience, learning rate, architecture, loss, temperature, optimizer or seed may be tried inside Batch 15.

Batch 16 preregistration — Late normalized factor minimax

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 14:07:11 UTC
Configuration: config/ecp7-b16-development.yaml
Raw configuration SHA-256: caef16ec93d0a9424fd3c4af46a987aad63118f24e57e77e9d51cc2eb77e073b
Effective configuration SHA-256: b40f35ab2f03fa8450805908713b4b9a59bbc060c38a4e68e4e3c5be2d2b4efb

Batch 15 established the first weak-structure protocol to pass validation and universal-translator thresholds together. Its remaining exact-match errors are concentrated in the weakest decoded factor after a high-entropy protocol is already established. Batch 16 tests whether late worst-factor emphasis can resolve those residual collisions without disrupting the successful early trajectory.

ECP7-B16-I keeps the complete Batch 15 architecture, base task and utilization losses, AdamW settings, data, temperature schedule, 30,000-step horizon, selection window, translator and thresholds. It adds the existing normalized factor-minimax term with one delayed coefficient schedule:

• weight is exactly 0 through step 15,000;
• weight increases linearly from 0 to 1.0 over steps 15,000–20,000;
• weight remains 1.0 from step 20,000 through step 30,000.

The minimax formula is unchanged from Batch 6: it selects the largest factor cross-entropy after normalization by the uniform-guess entropy for that factor. Unlike Batch 6, no minimax gradient is present while the base high-entropy code is forming. The original mean task loss remains active throughout.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B16-I with late normalized factor minimax on the complete B15 base.

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 start step, warmup, weight, normalization, duration, architecture, utilization, optimizer, data, translator or seed may be tried inside Batch 16. Failure is recorded without confirmatory test access.

Batch 17 preregistration — Late global collision-pair replay

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 14:35:00 UTC
Configuration: config/ecp7-b17-development.yaml
Raw configuration SHA-256: 81bc1c36f7fd19afdecd61eb917fe41925db8266d8ac8b2bc96e70178e39c412
Effective configuration SHA-256: 70921261774f38f5ca22f122d05a2b9f1f0bc7c9ca6ea4c327097b7619715d5b

Batch 16 showed that late semantic factor reweighting slightly improves texture but disrupts color generalization and does not make the code injective. Batch 17 therefore returns to the complete Batch 15 base and targets only the remaining complete-message collisions. It revisits the Batch 4 question after a mature high-entropy protocol exists, but replaces Batch 4's sparse random minibatch signal with deterministic training-codebook mining.

At step 15,000 and every 200-step registered evaluation boundary thereafter, each sender encodes all 14,336 training meanings with hard argmax symbols. For each complete message, every unordered pair of distinct training meanings that shares that message enters that sender's replay bank. Validation meanings, factor labels, factor identities and validation outcomes never enter mining.

For every subsequent population update, a separate seed-derived generator samples 32 mined pairs with replacement per sender. The sender produces relaxed message distributions at temperature 1.0. For pair (x,y), the new loss is the probability that all four relaxed symbols agree:

L_replay(x,y) = product_s <p_s(x), p_s(y)>

The loss averages over sampled pairs and senders. It specifies no desired symbol, target slot, semantic codebook or factor-to-slot binding. Reducing it requires at least one message position to distinguish each replayed pair.

ECP7-B17-I keeps every Batch 15 architecture, task and utilization loss, optimizer, data, temperature schedule, 30,000-step duration, selection window, translator and threshold. Collision-replay weight is exactly 0 through step 15,000, rises linearly to 1.0 over steps 15,000–20,000, and then holds. The separate replay sampler does not alter the registered task-batch sampler. The trajectory through step 15,000 must therefore match Batch 15 exactly.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B17-I with late global collision-pair replay on the complete B15 base.

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 mining source, refresh interval, pair batch size, temperature, weight, warmup, architecture, duration, optimizer, data, translator or seed may be tried inside Batch 17. Failure is recorded without confirmatory test access.

Batch 18 preregistration — Attenuated collision replay

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 14:53:31 UTC
Configuration: config/ecp7-b18-development.yaml
Raw configuration SHA-256: a9da48a2c76fd01a9f2889050c347ed9fcf616389297424f156e9f2d8489e59f
Effective configuration SHA-256: 21d76380640738ee98ddbdb7ae88c6a1d3185add860b96f1fe9c5c83a782d757

Batch 17 established that global training-codebook mining produces a usable collision signal and modestly improves hard code use. Its replay weight rose to 0.24 at the selected checkpoint, where replay loss 0.83844 contributed approximately 0.201 to the objective versus task loss 0.07453. Higher weights then caused collision churn, validation regression and early stopping.

Batch 18 tests only whether coefficient scale caused that failure. ECP7-B18-I inherits the complete B17 mining algorithm, training-only pair source, seed-derived sampler, 32-pair sender batches, 200-step refresh interval, relaxed temperature 1.0, architecture, base losses, optimizer, data, 30,000-step duration, selection window, translator and thresholds. The sole change is:

• replay weight remains 0 through step 15,000;
• replay weight rises linearly to 0.1 over steps 15,000–20,000;
• replay weight remains 0.1 through step 30,000.

For an initially near-one replay loss, full weight 0.1 is comparable to the remaining B15 task-loss scale instead of several times larger. The task-batch trajectory and all random sampling through step 15,000 must again match B15 exactly.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B18-I with attenuated late global collision-pair 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 coefficient, schedule, mining source, pair sampler, refresh, temperature, architecture, duration, optimizer, data, translator or seed may be tried inside Batch 18. Failure is recorded without confirmatory test access.

Batch 19 preregistration — Bounded collision-replay pulse

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 15:15:36 UTC
Configuration: config/ecp7-b19-development.yaml
Raw configuration SHA-256: dab2cd0dadbadb47a10e2c45a720611dc91d140d3783f0ca1a537e4e8a8549c0
Effective configuration SHA-256: cbd144a5fce451315a7d5fb322a11f2f29c32514fc22fbe06f46285aef677d13

Batch 18 confirmed that replay at task-loss scale avoids B17's early collapse, but validation peaked at step 22,400 and then regressed while weight 0.1 remained active. Batch 19 tests whether replay is useful as a bounded codebook diversification pulse rather than a permanent late objective.

ECP7-B19-I inherits the complete Batch 18 architecture, training-only mining, seed-derived pair sampler, 32-pair sender batches, 200-step refresh interval, relaxed temperature, base losses, optimizer, data, duration, selection, translator and thresholds. It is identical to B18 through step 20,000. The sole change is a coefficient decay:

• replay weight remains 0 through step 15,000;
• replay weight rises linearly to 0.1 at step 20,000;
• replay weight falls linearly to 0 over steps 20,000–25,000;
• replay weight remains 0 through step 30,000.

Training-codebook collision mining continues at evaluation boundaries after step 25,000 for diagnostics, but contributes no gradient and consumes no replay sampler draws. The final optimization phase therefore contains only the B15 task and utilization objectives. The complete trajectory through step 20,000 must match B18 exactly.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B19-I with one bounded global collision-replay pulse.

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 decay, endpoint, mining, pair sampler, architecture, base loss, optimizer, duration, data, translator or seed may be tried inside Batch 19. Failure is recorded without confirmatory test access.

Batch 20 preregistration — Global hard-meaning task replay

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 15:42:06 UTC
Configuration: config/ecp7-b20-development.yaml
Raw configuration SHA-256: 01afcad913222db6a6ef858919664cc50bfac9269f8405eb8cae0dd9832c7319
Effective configuration SHA-256: 44b1dbd8b767e4b3e58e88e7f1645e6dc1578b04fd30fa174777dffaf96c3b75

Batches 17–19 show that direct collision replay changes code use but does not stably repair the known-meaning error. A training-only diagnostic on the selected Batch 15 checkpoint found that 3,406 of 14,336 training meanings fail for at least one of the 16 sender-receiver links. Those meanings fail for 11.14 links on average and have mean ordinary reconstruction loss 0.20870, versus 0.06442 over the complete training set. Batch 20 tests whether this concentrated residual error is a hard-example sampling problem.

ECP7-B20-I inherits the complete Batch 15 architecture, random task batch, utilization objective, optimizer, data, temperature schedule, 30,000-step horizon, selection rule, translator and thresholds. It adds one mechanism:

• after the normal step-15,000 evaluation, and every 200 steps thereafter, evaluate hard predictions over training meanings only;
• include a meaning in the replay pool when any current sender-receiver link reconstructs at least one factor incorrectly;
• sample 64 pool meanings uniformly with replacement using a separate seed-derived generator;
• compute the ordinary all-senders/all-receivers reconstruction loss on that batch with the unchanged straight-through sender and current temperature;
• keep replay weight 0 through step 15,000, warm it linearly to 0.25 at step 20,000, and hold it through step 30,000.

At the Batch 15 diagnostic scale, weight 0.25 contributes approximately 0.0522 for the hard-example loss, comparable to rather than larger than the remaining ordinary task loss. The base random batch remains unchanged. Mining uses no validation meaning, factor-specific weight, slot assignment, code-distance target or collision pair. The complete trajectory through step 15,000 must match Batch 15 exactly.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B20-I with late global hard-meaning task 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 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
Preregistered: July 18, 2026 at 16:05:33 UTC
Configuration: config/ecp7-b21-development.yaml
Raw configuration SHA-256: d2d8d8bf3d87aef42b533bfd33eea4033af99444ec00b1cffe3d6f63c5970a5b
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.

Batch 22 preregistration — Sender-only shared-error replay

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 16:28:26 UTC
Configuration: config/ecp7-b22-development.yaml
Raw configuration SHA-256: 7d91bb825564681f059ba207766d0a392cc26bafaf2b113d5354c249abd1d2a0
Effective configuration SHA-256: 518ff24f4e1d625ff09366b9515dd5359a3a34fc7e8e801cca5dde40dd8f5acc

Batch 21's all-link replay pool grew from 1,745 to 1,997 meanings while validation and sender agreement improved. This is consistent with joint sender-receiver replay updates aligning the population around persistent ambiguous codes instead of forcing sender codes to move. Batch 22 isolates that gradient-routing explanation.

ECP7-B22-I inherits the complete Batch 21 predicate, Batch 20 replay mechanism, Batch 15 architecture, ordinary task batch, utilization objective, optimizer, data, temperature schedule, coefficient, replay batch, sampler, refresh, warmup, 30,000-step horizon, selection, translator and thresholds. The sole change is:

• during the additional all-link hard-replay forward pass, receiver parameters are treated as fixed constants, so that replay loss sends gradients through the receivers into sender messages but not into receiver parameters.

The ordinary base task remains unchanged and continues to update both senders and receivers every step. Receiver gradient flags are restored before backward, so ordinary-task receiver gradients remain active. Mining still uses hard training predictions only. The complete optimization trajectory through step 15,000 must match Batch 21, Batch 20 and Batch 15 exactly.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B22-I with sender-only population-shared hard 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 gradient route, predicate, coefficient, replay batch, refresh, schedule, architecture, base loss, optimizer, duration, data, translator or seed may be tried inside Batch 22. Failure is recorded without confirmatory test access.

Batch 23 preregistration — Phased shared-error replay routing

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 16:51:26 UTC
Configuration: config/ecp7-b23-development.yaml
Raw configuration SHA-256: 8e0f7ed9b4d9a487bf0700be85e922142aafa3f8ccadf28326227bace6b78015
Effective configuration SHA-256: 60f46c913a95077a9f40df695adf6bb9b8788f581c83a12deb35c1f9ca015fb9

Batch 21's joint replay preserved population alignment but grew the shared-error pool. Batch 22's sender-only replay reduced that pool but increased fragmented any-link errors and reduced sender agreement. Batch 23 isolates whether the two routing modes are useful in sequence: first move ambiguous sender codes against fixed receivers, then allow receivers to catch up.

ECP7-B23-I inherits the complete Batch 22 implementation, Batch 21 predicate, Batch 20 replay mechanism, Batch 15 architecture, ordinary task batch, utilization objective, optimizer, data, temperature schedule, coefficient, replay batch, sampler, refresh, warmup, 30,000-step horizon, selection, translator and thresholds. The sole change is:

• the additional hard-replay branch keeps receiver parameters fixed through step 20,000 and restores receiver-parameter gradients from step 20,001.

The replay loss remains zero through step 15,000 and warms linearly through step 20,000. The ordinary base task updates both senders and receivers at every step. Mining still uses hard training predictions only. The complete optimization trajectory through step 15,000 must match B22, B21, B20 and B15 exactly; through step 20,000 it must match B22.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B23-I with phased population-shared hard 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 switch step, gradient route, predicate, coefficient, replay batch, refresh, architecture, base loss, optimizer, duration, data, translator or seed may be tried inside Batch 23. Failure is recorded without confirmatory test access.

Batch 24 preregistration — Receiver-only replay catch-up

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 17:09:59 UTC
Configuration: config/ecp7-b24-development.yaml
Raw configuration SHA-256: 27e737c1eaa3c191253469fb863171c9875e1e9433c2b00b7a578e15609155dc
Effective configuration SHA-256: a01f184249b32859074909039ba86d495854a3c5f063c19f5519e2739135cb9e

Batch 22's sender-only replay separated population-shared failures but created link-specific fragmentation. Batch 23 restored joint replay after warmup and immediately recreated shared failures. Batch 24 isolates the remaining routing hypothesis: let receivers catch up to the separated codes without allowing the additional replay loss to move sender codes again.

ECP7-B24-I inherits the complete Batch 23 implementation, Batch 21 predicate, Batch 20 replay mechanism, Batch 15 architecture, ordinary task batch, utilization objective, optimizer, data, temperature schedule, coefficient, replay batch, sampler, refresh, warmup, 30,000-step horizon, selection, translator and thresholds. The sole change is:

• after step 20,000, the additional hard-replay branch detaches sender messages, making that branch receiver-only instead of joint.

Replay remains sender-only through step 20,000. The ordinary base task updates both senders and receivers at every step. Mining still uses hard training predictions only. The complete optimization trajectory through step 20,000 must match Batch 23 and Batch 22 exactly.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B24-I with sender-only then receiver-only hard 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 switch step, gradient route, predicate, coefficient, replay batch, refresh, architecture, base loss, optimizer, duration, data, translator or seed may be tried inside Batch 24. Failure is recorded without confirmatory test access.

Batch 25 preregistration — Extended receiver-only catch-up horizon

Status: preregistered for sealed development
Preregistered: July 18, 2026 at 17:26:20 UTC
Configuration: config/ecp7-b25-development.yaml
Raw configuration SHA-256: 3c125aa0e21b29327b332a77f0a367633267804523eb941d5dad9eb622255387
Effective configuration SHA-256: 00c9f13c5919f98f5b81c14c8ea2d852a1cd4be55ecbfcaf0d61abed6b9acc91

Batch 24 established the strongest mean ECP-7 performance, selected its best checkpoint at step 29,400, and reached the 30,000-step ceiling before the 5,000-step patience could assess a late plateau. Batch 25 isolates whether that receiver-only catch-up phase is still improving or has already converged to a non-injective solution.

ECP7-B25-I inherits the complete Batch 24 implementation, configuration, gradient routes, hard-meaning predicate, replay mechanism, architecture, ordinary task batch, utilization objective, optimizer, data, temperature schedule, coefficient, replay batch, sampler, refresh, warmup, selection, translator and thresholds. The sole change is:

• training.max_steps increases from 30,000 to 45,000.

The 5,000-step early-stopping patience and 15,000-step minimum remain unchanged. The run may therefore finish before step 45,000. The complete optimization trajectory through step 30,000 must match Batch 24 exactly.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B25-I with the extended receiver-only catch-up ceiling.

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 horizon, gradient route, predicate, coefficient, replay batch, refresh, architecture, base loss, optimizer, data, translator or seed may be tried inside Batch 25. Failure is recorded without confirmatory test access.

Batch 26 preregistration — Coordinate-descent replay cycle

Status: completed sealed development; valid negative
Preregistered: July 18, 2026 at 17:46:02 UTC
Configuration: config/ecp7-b26-development.yaml
Raw configuration SHA-256: 9721d3af452b72788568f7fc5262056aa770e443c0a3bc35112183a08c2badb7
Effective configuration SHA-256: be1310f79b4b7a72134393292480244eeba123d4828b3c94dbf32787dd6a4e3b

Batch 25 shows that receiver-only catch-up nearly saturates the mathematical capacity of the learned sender codebooks. The worst train link exactly matches its sender's 82.81% unique-code ceiling. Batch 26 tests whether one additional sender-only separation pulse against those saturated receivers can break collisions before a final decoder catch-up phase.

ECP7-B26-I inherits the complete Batch 25 implementation, 45,000-step ceiling, hard-meaning predicate, replay mechanism, architecture, ordinary task batch, utilization objective, optimizer, data, temperature schedule, coefficient, replay batch, sampler, refresh, warmup, selection, translator and thresholds. The sole change is the explicit one-sided replay schedule:

1. sender-only through step 20,000;
2. receiver-only through step 30,000;
3. sender-only through step 35,000;
4. receiver-only through step 45,000.

Each boundary is inclusive, so the next route begins on the following step. The ordinary base task updates both senders and receivers at every step. Mining still uses hard training predictions only. The complete optimization trajectory through step 30,000 must match Batch 25 and Batch 24 exactly.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B26-I with the coordinate-descent replay cycle.

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 phase boundary, gradient route, predicate, coefficient, replay batch, refresh, architecture, base loss, optimizer, horizon, data, translator or seed may be tried inside Batch 26. Failure is recorded without confirmatory test access.

Batch 26 outcome

The registered seed-11 intervention reached 85.37% mean train, 82.74% worst-link train, 85.38% mean validation, 82.71% worst-link validation and 85.38% universal-translator validation. Its sender codebooks used 13,438, 12,960, 12,720 and 13,440 unique messages, effectively unchanged from Batch 25. The second sender-only pulse reduced shared errors but increased meanings failed by at least one population link. It therefore failed the unchanged train and injectivity gates without accessing the confirmatory split.

This completes the registered replay-routing family. Further route cycling or horizon extension is disallowed as an unsupported repetition; any later development batch must preregister a direct sender-capacity intervention.

Batch 27 preregistration — Late bounded sender-collision pulse

Status: completed sealed development; valid negative
Preregistered: July 18, 2026 at 18:05:30 UTC
Configuration: config/ecp7-b27-development.yaml
Raw configuration SHA-256: d44df83436f5ca81a2a51229b8ffe1a6fd6832d28dff4c7ab2ead3ace8efc9bd
Effective configuration SHA-256: 649ddca4a285fbbb00b29c1cab82678e42ab8d6810f89f2131766cd82496c463

Batch 25 establishes the strongest stable receiver-saturated base. Batch 26 shows that ordinary-task route cycling cannot add sender code capacity. Batch 27 therefore changes the objective rather than the route: it directly penalizes each sender's globally mined hard-message collisions only after the complete Batch 25 trajectory through step 30,000.

ECP7-B27-I inherits the full Batch 25 configuration and implementation. The sole addition is the existing factor-agnostic global collision-pair mechanism with this fixed schedule:

1. weight zero through step 30,000;
2. linear warmup to 0.1 at step 35,000;
3. linear decay to zero at step 40,000;
4. weight zero through the 45,000-step ceiling.

Pairs are mined independently for every sender from hard messages over training meanings only, refreshed every 200 steps and sampled as 32 pairs per sender with replacement. Their relaxed full-message collision probability is measured at temperature 1.0. The coefficient is reused from Batch 18; the 5,000-step rise and decay durations are reused from Batch 19. The unchanged hard-meaning replay branch stays receiver-only after step 20,000, and the ordinary task continues to update both sides.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B27-I with the late bounded sender-collision pulse.

All existing validity and development gates remain unchanged. Additionally, the direct-capacity claim requires the selected minimum unique-message count to exceed Batch 25's 12,720 while worst-link validation is at least Batch 25's 82.7148%. No alternative coefficient, pulse boundary, pair sampler, replay route, hard-meaning predicate, architecture, base loss, optimizer, horizon, data, translator or seed may be tried inside Batch 27. Failure is recorded without confirmatory test access.

Batch 27 outcome

The registered seed-11 intervention reached 85.38% mean train and validation, 82.15% worst-link train, 82.13% worst-link validation and 85.35% universal- translator validation. Its global collision-pair banks became substantially smaller and message entropy increased, but sender codebooks still contained exactly 13,440, 12,960, 12,720 and 13,440 unique messages—the same capacities as Batch 25. It failed the unchanged train and injectivity gates and the added minimum-capacity/worst-link criterion without accessing the confirmatory split.

This closes coefficient and timing development for the existing relaxed collision-pair loss. Any later batch must preregister a distinct sender representation or occupancy-creation mechanism.

Batch 28 preregistration — Identity-preserving residual parallel sender

Status: completed sealed development; valid negative
Preregistered: July 18, 2026 at 18:30:20 UTC
Configuration: config/ecp7-b28-development.yaml
Raw configuration SHA-256: 1ce3094d0e372f4b052773a2c44d1abe36cd6255aea96f4218c175e5f93977d7
Effective configuration SHA-256: 7bb7c2c185ed46f580464fe4364b8070bf5fdb071d3d64b16ec66b7fbe7ec307

Batch 13 replaced the successful shallow sender context with an additional nonlinear layer and collapsed. Batches 25–27 now show a stable hard-code capacity ceiling that receiver training, replay routing and direct collision pressure do not cross. Batch 28 isolates a representation hypothesis without discarding the shallow path: add one identity-preserving residual interaction to the shared joint context.

ECP7-B28-I inherits the full Batch 25 configuration, including its 45,000-step horizon, receiver-only hard-meaning replay after step 20,000, architecture on the receiver side, losses, optimizer, data, temperature, selection, translator and thresholds. The sole configuration change is the sender family from bounded_parallel_sender to residual_bounded_parallel_sender.

The new sender first computes the unchanged shallow context. It then adds tanh(W context + b) before the unchanged slot heads. W and b are created as exact zero tensors without consuming random draws. Consequently the residual sender has exactly the same initial discrete messages and pre-existing parameters as the shallow sender under the same seed, while its residual parameters receive gradients immediately. The branch is shared by all slots and sees only the joint context; it has no factor-local projection, factor-slot binding, semantic symbol assignment or capacity change.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B28-I with the residual bounded-parallel sender.

All existing validity and development gates remain unchanged. Additionally, the representation claim requires the selected minimum unique-message count to exceed Batch 25's 12,720 while worst-link validation is at least Batch 25's 82.7148%. No Batch 27 collision loss, alternative residual width, activation, initialization, receiver, replay route, base loss, optimizer, horizon, data, translator or seed may be tried inside Batch 28. Failure is recorded without confirmatory test access.

Batch 28 outcome

The registered seed-11 residual sender reached 53.91% mean train, 53.12% worst-link train, 9.65% mean validation, 8.11% worst-link validation and 11.62% universal-translator validation. Its sender codebooks collapsed to 7,928–8,333 unique messages. Although tests prove exact functional and RNG-equivalent initialization, the trainable residual rapidly became dominant and destroyed the shallow sender's compositional trajectory. It failed every performance and injectivity gate without accessing the confirmatory split.

This closes generic sender-depth changes active from the start of training. A future residual experiment is justified only as a single late-activation test that preserves the complete established shallow trajectory before unfreezing.

Batch 29 preregistration — Late residual sender activation

Status: completed sealed development; valid negative; ECP-7 development closed
Preregistered: July 18, 2026 at 18:47:04 UTC
Configuration: config/ecp7-b29-development.yaml
Raw configuration SHA-256: 4c4e32dc054eb4af400e04fccc5c312edea0e5cbfa9c39fbee6e2615d56258b9
Effective configuration SHA-256: 6483d2cf84f2e7e1d0365e90aeefc99641fb9255c3617d8d3928623af5db3947

Batch 28 proves that a residual interaction trainable from the first update destroys compositional learning. It does not test whether the same capacity can split collisions after the successful shallow protocol is already established. Batch 29 is the single permitted temporal isolation of that hypothesis.

ECP7-B29-I inherits the complete Batch 28 representation and Batch 25 training setup. The sole change is a registered residual-parameter activation boundary:

1. residual weight and bias are frozen at exact zero through step 30,000;
2. residual weight and bias receive gradients from step 30,001 onward.

All original sender and receiver parameters remain trainable at every step. The residual branch consumes no random draws, and at zero it leaves context and messages unchanged. The complete trajectory through step 30,000 must therefore match Batch 25 across all shared history fields. After activation, no scaling, warmup or added loss is used. The existing hard-meaning replay remains receiver-only after step 20,000, so residual parameters are changed only by the ordinary joint task.

The batch contains exactly two seed-11 runs:

1. the unchanged ECP-6 positive control;
2. ECP7-B29-I with late residual activation.

All existing validity and development gates remain unchanged. Additionally, the late-capacity claim requires the selected minimum unique-message count to exceed Batch 25's 12,720 while worst-link validation is at least Batch 25's 82.7148%. No alternative boundary, gradual activation, residual architecture, collision replay, replay route, base loss, receiver, optimizer, horizon, data, translator or seed may be tried inside Batch 29. If it fails, ECP-7 development closes without confirmatory test access; no further residual schedule is admitted.

Batch 29 outcome and development closure

The registered seed-11 intervention exactly matched all 151 shared Batch 25 history records through step 30,000, with residual gradients disabled at every record. After activation it selected step 31,000 and reached 84.89% mean train, 82.28% worst-link train, 84.74% mean validation, 81.74% worst-link validation and 85.01% universal-translator validation. Its sender codebooks used 13,163, 12,960, 12,720 and 13,439 unique messages, a net loss relative to Batch 25. It failed the train, injectivity and added late-capacity/worst-link gates without accessing the confirmatory split.

No ECP-7 development variant passes the full registered gate. The development phase is closed after Batch 29; no confirmatory ECP-7 configuration is frozen and the confirmatory split remains sealed. Any future work must begin with a new ECP-8 research question, split and preregistration rather than another ECP-7 variant.