Engineered Thinking

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 language. Git and IPFS use content addresses as infrastructure identifiers; Sema makes them readable words in the medium models already think and speak in, so a reference can anchor a step in a model's own reasoning as much as a message between agents. 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, so a procedure can be cited, checked, and composed rather than paraphrased. A change to any hashed field produces a different address, and a strict handshake halts on mismatch. The bootstrap library contains 457 Pattern Cards.

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 instead moves safety-relevant learning into pretraining: one continuous Reader processes arbitrary works chronologically, carrying the same evolving, provenance-linked Understanding Graph across the corpus. At every selected Thought Moment, each thinking block opens with the full, verbatim Reader Core before source-specific evaluation, binding a recurrent orientation to capability-bearing material rather than isolating it in safety documents. The 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 have no fixed solution list or cheap score. This paper presents two methods. World-diversity search creates target-independent worlds with altered causal rules. A Solver works inside each world, and a compiler converts the result into a target proposal. Ontology-governed intervention search stores proposals in a typed graph. A Taxonomist judges structural equivalence and can turn rejection into targeted commissioning. Its diagnosis states what the next proposal must change. Accepted proposals and revised categories guide later Explorers. The study produced 196 records, including 25 world-and-solve branches and seven complete rejection-to-mechanism-change chains.

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) turns resolved events into forecasting targets for an earlier cutoff. A future-aware Teacher uses the outcome to write or revise a rationale from evidence designated as available at that cutoff. The Student does not receive the outcome. A specificity frontier limits the target to claims supported by the earlier evidence, even if the forecast later proves wrong. THL tests whether this training strengthens causal reasoning and prediction. The pilot trained Llama 3.3 70B on 505 Teacher-written traces from 106 events in 2024 and evaluated 75 prompts across 15 events in 2025. One model judge scored THL higher than Base for reasoning, but the design did not isolate transferable reasoning.

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%.