From Semantic State to Formal Analysis: Graphs, Provenance, and Auditable LLM Workflows

Beyond the Chat Window · Article 5 of 8

The method developed across this series began with a practical observation: a chat transcript is not an ideal state model for a long-running project. From there, the architecture acquired a schema, mutation rules, commit behavior, a manifest, adaptive types, and a composition pipeline capable of surviving both late requirements and failed prose generation.

At first glance, these may look like separate engineering conveniences. They become more interesting when considered together. Stable objects have identities. Dependencies, support relations, revisions, and derivations connect those objects. Change events affect specific downstream artifacts. Once those relations are explicit, the workspace can be treated as a graph.

That move creates two capabilities that are difficult to obtain from prose history alone. First, the topology of the work can be analyzed using ordinary graph methods. Second, the lineage of an output can be represented as a traversable provenance structure. The LLM remains probabilistic, but parts of the workflow become formally inspectable.

From semantic relations to computable structure

Natural language is very good at saying that one thing depends on another. “Article 3 assumes the persistent-state model from Article 1 and the mutation protocol from Article 2” is immediately intelligible. What prose does not give us automatically is an addressable relation that another process can traverse.

If Article 1, Article 2, and Article 3 have stable IDs, the same information can be represented as typed edges. Claims can support other claims. A decision can supersede an earlier decision. A schema change can affect several articles. A paragraph can derive from a set of argument objects. Different edge types preserve different semantics, but they all share a useful property: they turn relationships from things that must be inferred repeatedly into things the system can inspect directly.

This matters because many project questions are graph questions in disguise. If one foundational assumption changes, which artifacts need review? That is reachability. Are there circular dependencies in a structure that is supposed to be acyclic? That is cycle detection. Which object has unusually many downstream dependents? That is a centrality or degree question. Are there isolated objects that were created but never used? That is a connectivity question.

The graph does not make the semantic judgments objective. The model may have been wrong when it decided that one claim supports another. But once a particular relation has been committed to state, the computation over that relation can be deterministic. This separation is important: semantic inference can remain probabilistic while structural analysis becomes exact relative to the represented graph.

The series itself has such a graph. One major belief node establishes that persistent structured state is distinct from conversational history. Another establishes that state representation needs transition semantics. A third depends on both of those results to justify resumable execution. The adaptive-schema argument depends on the existence of governed mutation, and the final provenance argument depends on stable identities and explicit relations established throughout the earlier articles. The series is sequential, but not merely linear. Later conclusions inherit whatever earlier results they actually need.

That point became visible when the original orchestration tried to force each article to inherit exactly one result from its predecessor. The rule was rejected because it imposed an artificial topology on the argument. The corrected protocol allows dependency structure to follow inferential necessity rather than a fixed one-edge chain. In graph terms, the series is a directed network with a readable publication order, not a linked list.

Provenance is lineage, not just citation

Graph structure becomes especially valuable when some edges represent derivation. A citation tells us which source is associated with a statement. Provenance can tell us how an artifact came to exist.

Imagine a claim object that records three upstream inputs, a transformation type, and the pass in which it was created.

A paragraph in the final article can point to that claim. The claim can point to its source observations and earlier conclusions. Those objects can point further upstream. The resulting chain is not a transcript of the model's hidden reasoning; it is an external record of what the workflow treated as inputs, what transformation it requested, and what output it accepted.

That distinction matters because “auditability” is often confused with access to internal chain-of-thought. The method described here does not require, and should not pretend to provide, a transparent view into the model's private inference process. Instead, it aims at workflow auditability. We can inspect which state was available, which dependencies were declared, what operation was performed, what artifact was produced, which editorial decisions were accepted, and which later artifacts depend on the result.

The current series provides a concrete provenance chain. Late in the project, I introduced the idea that the series itself should become the central worked example. That instruction became a structured change event. The event recorded its origin and affected scope. Each article then acquired a requirement to demonstrate its own methodological layer using the live production history. The composition maps incorporated those requirements. Fresh prose was generated from those maps. Revision records then captured competing editorial proposals and referee decisions. The final PDF is downstream of that entire sequence.

The lineage can therefore be described without inventing a story about hidden reasoning: my requirement became a change event; the change event produced a schema extension; the extension created article-level obligations; those obligations shaped the composition maps, drafts, refereed revisions, and final artifact. That is a demonstrable process claim because a reviewer can inspect the external objects and verify the sequence.

Formal analysis can guide revision

Once provenance and dependency are represented together, formal analysis can do more than document history. It can guide future work. Suppose a claim near the beginning of the series changes. A dependency traversal can identify every article, section, or paragraph that derives from it. The system can classify those artifacts as unaffected, review- required, or recomposition-required. The LLM can then inspect only the affected region semantically. This is far more controlled than regenerating the whole project because something upstream moved.

The same logic applies to editorial work. In the revised protocol used for this version of the series, each paragraph is reviewed by two independent revision agents. One is biased toward compression and precision; the other is biased toward development and rhetorical continuity. A strict editor referees the proposals and can accept one, synthesize them, keep the original, or remand the problem to the composition orchestrator if the defect is structural. Those rulings become provenance objects too.

Why bother storing that history? Because editorial decisions create dependencies. A later paragraph may rely on a qualification introduced during revision. A structural remand may change the function of an entire section. If the final artifact is ever revised again, the workflow can inspect not just the latest prose but the reasons certain forms were chosen.

This does not mean every writing project needs a massive graph database. The representation should remain proportional to the task. The principle is that once relations matter operationally, representing them explicitly makes both analysis and revision more reliable.

A hybrid architecture for long-running LLM work The broader architecture is now visible. The LLM is used for what it does well: interpreting human intent, recognizing semantic similarity, proposing distinctions, constructing arguments, generating prose, and judging whether formal results matter in context. Persistent state is used for what it does well: preserving identity, authority, current values, versions, and relations. Protocols govern mutation. Manifests represent execution progress. Graph algorithms analyze topology. Provenance records lineage. Editorial orchestration turns structured argument back into coherent human- readable artifacts.

These components can operate in a loop. The model infers that a new dependency exists. The relation is persisted. A deterministic traversal discovers which artifacts depend on it. The model interprets those results and proposes revisions. Competing editorial agents challenge the local prose. A referee accepts or rejects changes. The state is updated, and the next operation proceeds from the new checkpoint.

This is not an attempt to make an LLM behave like a conventional program by removing ambiguity. It is an attempt to put probabilistic semantic work inside an environment where the consequences of that work can be represented, inspected, resumed, and sometimes analyzed exactly.

The five articles themselves have served as a small demonstration. The project began as a conversational idea I developed with the model. It acquired a schema. The schema acquired a protocol. The argument was progressively formalized. Execution was interrupted and resumed. A late idea propagated without restart. The first prose pipeline failed, and the system rolled back to a trustworthy structural layer rather than discarding everything. A new composition protocol introduced independent revision agents and a strict editorial referee. The final artifact emerged from that accumulated state.

No single part of this architecture is novel in isolation. State machines, manifests, dependency graphs, schema migrations, provenance records, and editorial workflows all have long histories. What is interesting is how naturally they fit around an LLM once the model is given read-write access to persistent files. The model supplies semantic flexibility; the external system supplies continuity and form.

That combination suggests a different way to think about serious LLM use. The chat window is not the whole application. It is the place where a person talks to a system whose real working state can be larger, more structured, more durable, and more analyzable than the conversation itself.