Engineered Thinking

Can a language model build a world as it writes? In STLM, one neural network predicts a structured target instead of a word. The target can be an image, a description, or later a video. A second neural network must read the target to recover the next word. There is no direct text route around this interface. The research asks whether repeated, meaningful targets can help the model form a shared scene with reusable objects, relations, motion, and abstract structures. The first experiment uses one fixed image or description for each of 1,024 TinyStories tokens. The causal description and pixel readers reach 41.9% and 37.8% top-1 accuracy. A direct-token model reaches 44.7%, while stateless readers reach 35.0% and 27.8%. These results show that the compulsory interface can carry useful next-word information. They do not yet show that the model uses the meaning of the targets or builds a scene. The next tests will measure the value of earlier substrate states, compare aligned targets with permuted targets, and test composition across words, phrases, and sentences.

Sema creates verifiable words for reasoning and communication. Participants encode information in an agreed form, hash it, and use the content address as a term in messages. Sema can work with any representation that participants can encode, resolve, and decode in the same way. This paper develops one implementation with canonical JSON objects, Merkle commitments, and content-addressed references. It uses Pattern Cards as a worked application. A Pattern Card defines a reasoning procedure through its mechanism, conditions, invariants, dependencies, and failure modes. This structure makes deliberate reasoning moves explicit and checkable. The same references help agents verify that they resolved the same external definition. A strict handshake stops coordination when full identifiers do not match. The bootstrap library contains 457 Pattern Cards. Its measured contract fields used 26.0 times as many tokens as short wire references. This result measures transmission compression, not inference-token savings. The agent demonstrations are illustrative. They do not establish general performance gains.

Understanding — the movement from confusion to clarity — was always ephemeral. When AI reasons in tokens, it becomes directly storable in the medium where it occurs. Rather than indexing documents or extracting user facts — the two dominant paradigms in AI memory — the Understanding Graph captures the cognitive process itself: tensions, hypotheses, belief revisions, dead ends. Not what the AI concluded, but how it understood.

Most alignment methods intervene after capabilities have formed. Entangled Alignment belongs to a newer family of work that moves safety-relevant learning into pretraining, but proposes an unusually structural version of that idea. One continuous Reader processes arbitrary source works chronologically, carrying the same evolving, provenance-linked Understanding Graph from earlier works into later ones. At every selected Thought Moment, each generated thinking block opens with the full, verbatim Reader Core before continuing into source-specific evaluation and revision. The safety-relevant signal is therefore attached to capability-bearing material, supported by persistent memory, and anchored by the same recurrent orientation across the corpus—not confined to a separate collection of safety documents. The paper defines three hypotheses: whether reader traces improve capability and safety-relevant evaluation (H1), whether graph conditioning improves those traces (H2), and whether the recurrent Core creates a stable but revisable orientation (H3). Two case studies demonstrate the Teacher pipeline, but no Student has yet been trained. The distinctive empirical bet is the combination of persistent graph memory and exact Core recitation. Its long-run aim is inheritance rather than armor: making the orientation part of what capable systems learn and potentially pass to their successors.

Open-ended problems do not have a fixed list of solutions or a cheap score. This paper presents two search methods. World-diversity search creates target-independent worlds with different causal rules. A Solver works inside each world. A compiler then converts the result into a proposal for the target problem. Ontology-governed intervention search stores complete proposals in a typed graph. A Taxonomist compares each proposal with the graph. It can reject a repeated mechanism, change categories, and direct a fresh Explorer. The study produced 196 candidate records for one retirement-design problem. It includes 25 world-and-solve branches and seven complete graph-redirection chains. The artifacts show automated world branching and graph-based redirection. They do not show that the proposals are novel, useful, or better than matched baselines.

Existing frameworks decompose tasks. This paper decomposes concepts — the persistent structure of what a domain is made of. Each concept becomes a solver node behind one five-surface contract, so a leaf and a thousand-node subtree are indistinguishable to their caller: specialists all the way down. In a prototype of 100 problems across 20 domains, concept-based routing produced a shared graph of 456 nodes with 64% reuse — the structural precondition, not yet the improvement it predicts. If decomposed solving beats the conventional approach at matched compute, and independent attempts converge on the same concepts, the result is an internet of reasoning: a shared substrate where you post a problem rather than fetch a page.

Temporal Hindsight Learning (THL) uses resolved events to create training targets for forecasts at an earlier knowledge cutoff. After an event resolves, a future-aware Teacher sees the outcome. The Teacher then writes or revises a rationale that uses evidence designated as available at the cutoff. The Student does not receive the outcome as input. The hypothesis is that repeated training on these targets can make the computations for evidence, mechanisms, uncertainty, and prediction more reliable. The pilot trained Llama 3.3 70B on 505 Teacher-written traces from 106 events in 2024. It evaluated stored outputs for 75 prompts from 15 events in 2025. One model judge gave THL higher reasoning scores than the Base model. An event-level reanalysis supports this difference. It does not establish an accuracy improvement. The pilot also does not separate reasoning transfer from transferred facts, response format, decoding differences, or judge preference. The paper proposes Student-anchored refinement, chronological Council training, graph targets, checkpoint trajectories, and rolling updates. These extensions were not evaluated.