Computational Dialectics

VI. What have we learned from the experiments?

24. D0: a small but real beginning

The first successful gate was not a large Dialectical Core. D0 established a continuation-safe, physically bounded scoped scalar-update algebra. It was deliberately kept narrow.

D0 showed that when scope and continuation semantics are defined explicitly, a particular update regime can be made executable and falsifiable.

But D0 did not show representation revision, ontology discovery, or a neural-architecture advantage. This book will not present it as more than it was.

25. D1-A: why was the first neural signal not enough?

In D1-A, a relational structure outperformed an ordinary global baseline on two retained development seeds. But the advantage of the special M0 mechanism over B1 did not replicate.

This result matters because it shows the difference between "we saw a strong result on one seed" and "the mechanism works." D1-A was therefore closed as inconclusive, and the sealed confirmatory stage was not opened.

26. D1-B: selective relational representation revision

D1-B asks a cleaner question:

When new evidence invalidates only a specific part of a relational structure, can the model revise the wrong part while preserving what remains correct? And can a special relational-selective architecture do this more reliably than strong ordinary alternatives?

The benchmark separates mapping revision, edge rebinding, coupled propagation, edge removal, unchanged structure, and no-change behavior. A strong symbolic causal control solves the task; static/passive shortcuts do not. This shows that the task itself is meaningful, but not that the special neural architecture is necessary.

27. The B0-v1 failure: how a bad baseline can produce bad science

The first learned pilot stopped technically before producing a scientific winner. Fifteen shared-weight recurrent microsteps had been added to the flat-memory B0 baseline only to match FLOPs. During training, gradient norms exploded and loss became non-finite.

We initially considered whether the problem was only float32 norm computation. We tested that explanation with a float64 norm fallback. The model progressed a few more updates, but the actual training dynamics still collapsed. That falsified the "clipping bookkeeping only" explanation.

This negative result produced an important methodological lesson:

If we artificially make a baseline harder to optimize merely to equalize resources, we can make the special architecture look unfairly strong. Fair comparison must consider optimization geometry, not only parameter/FLOP counts.

B0-v2 was therefore redesigned as a cleaner ordinary baseline: one persistent update, with the remaining matched compute allocated to generic parallel transient computation. Its stability-only calibration passed 2,000/2,000 updates. This is not a capability result; it only shows that the new baseline is experimentally trainable.

28. The experimental protocol matters as much as the model

After B0-v2 became stable, we prepared protocol v4 for a new learned pilot. At first glance, the experiment looked secure: fresh development seeds, a pre-fixed confirmatory bank, unchanged optimizer and decision thresholds, and a "sealed" test side. The zero-step dry-run also passed.

But an independent adversarial review exposed a deeper problem. Scientific reliability depended not only on the model, data, and metric. The order in which results were revealed, how the selected checkpoint was proven, whether a run could be repeated, and which immutable artifacts produced the final decision were also part of the experiment.

In the old flow, confirmatory runs could be trained and then tested one by one. Even with a pre-fixed test bank, the first run's result could psychologically or operationally influence later execution decisions. The "test seal" was not a cryptographic secret; test roots were already visible in older artifacts. The real protection therefore had to be a provenance protocol that enforced when information was allowed to be used.

For that reason, v4.1 split the confirmatory flow into three phases:

1. all eight runs train and select checkpoints using only TRAIN and DEV;
2. all 20-checkpoint ledgers, selected-checkpoint hashes, and strongest ordinary comparators are bound into a single complete selection freeze;
3. only after that are the eight test packages opened in a single scoring phase.

Run directories were also made write-once, technical failure produces a separate abort artifact, and NaN/Inf counts as a technical failure not only in training loss but also in logits, probabilities, evaluation loss, and decision metrics. The finalizer no longer trusts run_summary.json; it recomputes selection from checkpoint ledgers.

During this hardening process, we found another subtle defect: the first dry-run fixture did not score the model on test, but it did actually regenerate confirmatory test rows. That still violated the new information-barrier principle. The fixture was changed; it now tests control flow without instantiating test data, and the old per-run test-opening path is fully closed. Code that validates an experiment must obey the experiment's own information regime.

Similarly, confirmatory training authorization no longer permits test opening. After all eight runs' selection freeze is complete, a separate one-shot scoring authorization bound to the exact SHA256 of that freeze is required. If scoring fails technically halfway through, it cannot simply be retried under the same protocol version; a new recovery version and audit are required. This prevents partially observed test results from influencing a later attempt.

The experiment produced a broader lesson:

Having a "sealed test" is not enough. Scientific information flow must also be designed and tested.

This finding is not a capability result for or against the Computational Dialectics mechanism. But it makes the research method used to develop the Core more trustworthy. Falsely validating a mechanism is more dangerous than having the mechanism fail.

What did we do when two reviewers disagreed?

We asked two independent AI reviewers to inspect the v4.1 candidate. The first returned PROCEED. The second returned REVISE, arguing in particular that metrics undefined on stress subsets could propagate as NaN into result artifacts. We did not decide by which model was stronger or by reviewer majority. We checked the claim directly against the frozen stress data.

We found that the k=0 stress group contained only NO_CHANGE_STABILITY queries. Structural and some preservation/held-out macros were therefore genuinely undefined for that subset. The v5 runner represented the missing values as NaN, while writing JSON artifacts with allow_nan=False. The second reviewer had identified a real defect: even if model scoring completed, result serialization could fail technically.

This case produced an important research principle:

Reviewer consensus is not evidence. A criticism should be tested through a reproducible observation, independent of which model produced it.

For the same reason, a PROCEED review does not by itself open a gate. Review, experiment, and static verification are different evidence sources that check one another. A defect missed by one reviewer may be exposed by another review and direct inspection of the data.

29. Scientific status at this snapshot

At the time of this book snapshot, D1-B has no architecture winner. Protocol v4 received REVISE during adversarial review because of execution-integrity weaknesses. A v4.1 repair was prepared without changing the scientific thresholds; a learned pilot would begin only if the exact v4.1 runner received PROCEED in a fresh adversarial review, was frozen, and received explicit development-only authorization.

This is exactly why the book is living. If D1-B eventually favors M0, evidence for a selective relation-specific mechanism increases. If B2 matches it, generic sparse memory becomes the simpler explanation. If B1 matches, relational structure may matter while exact-zero selectivity may be unnecessary. If ordinary baselines outperform M0, the special-primitive claim for this task family should be rejected.

---