Independent ML research

Structured State for Evidence-Bound Conversational Reasoning

This program asks whether an explicit, bounded relation-state can help a conversational system preserve the right entity, source, status, and correction—and whether that state adds causal value beyond the language model and surrounding symbolic machinery.

View résuméDiscuss the work

4 seeds

Replicated finite transfer

Exact held-out performance in the strongest finite selector study, paired with a decisive disabled-state control.

3,840

Validated finite traces

Persisted trajectories passed the declared finite-observation schema; this validates representation, not open-domain accuracy.

0 selected

Bounded conversational cores

The architecture arena stopped at validation instead of promoting a weak candidate or opening sealed evaluation.

Evidence progression

Positive result, then increasingly difficult boundaries

Study 1

Finite transfer

Positive within scope

A learned relation/status selector reached exact held-out transfer across four seeds. The state-disabled path scored 0%, and 3,840 finite traces passed the declared schema checks.

Boundary: The task is finite, synthetic, and moderately templated. Entity copying, query binding, recent-record retention, and closure remain symbolic.

Study 2

Natural-language intake

No front end promoted

The strongest validation candidate reached 0.921 answer exactness and 0.869 stale-record F1, but missed the answer, stale-state, minimum-family, and minimum-shape gates.

Boundary: Evaluation and stress labels stayed sealed. The result rejects the tested front ends, not natural-language structured state in general.

Study 3

Conversational architecture arena

No architecture selected

Five matched bounded cores were compared. The best two-scale mean interactive success was 0.257 against a 0.85 gate; every candidate had a zero-success family.

Boundary: Replication was not triggered and sealed evaluation stayed closed. This rejects the tested implementations under the declared data and compute budget.

Study 4

Pretrained decoder admission

Language substrate admitted

A frozen 0.5B decoder failed admission at 19/30. A first 7B audit scored 25/30 and exposed a scorer false negative. A fresh evaluator-owned bank then admitted the frozen 7B decoder at 30/30.

Boundary: The 30/30 result establishes only a competent shared language substrate. No attachment was trained or compared, and no formal benchmark row was generated.

What the evidence supports

A finite mechanism worth testing at a more credible language boundary

The finite result demonstrates task-congruent learned selection and a causally necessary state pathway on its declared benchmark. The follow-ups show that this does not automatically survive ordinary language intake or produce conversational intelligence in small learned systems.

Claims deliberately not made

  • No open-domain language understanding result.
  • No general conversational intelligence result.
  • No universal learned-reasoning architecture.
  • No complexity-theory separation or physical-law conclusion.
  • No claim that machine checks establish novelty or significance.

Next decisive experiment

Compare structured attachments only after language competence is fixed

  1. 1.Freeze the admitted decoder, evaluator, data root, and compute accounting before training.
  2. 2.Compare decoder-only, workspace, fixed-mesh, and adaptive structured-state attachments under matched budgets.
  3. 3.Require a smaller evidence-binding ladder to pass before opening the full conversational arena.
  4. 4.Use evaluator-owned paired-state swaps and separately trained checkpoint ablations to test causal dependence.
  5. 5.Treat a clean negative result as evidence against these implementations rather than against intelligence or emergence in general.

Authorship, AI assistance, and disclosure boundary

I used AI systems extensively for drafting code and prose, generating candidate formulations, repository organization, and mechanical checks. I selected the research questions, experimental gates, interpretations, and follow-on decisions. I can explain and defend the public methods and conclusions independently.

This page is a manually authored public summary. The complete research archive, raw predictions, training data, checkpoints, prompts, logs, and unpublished proof chains remain private. Formal checks establish only their stated propositions—not novelty, significance, or broader scientific conclusions.