prose: split two run-ons, drop the We-find openers, cut a restated clause and give Related Work its third paragraph

app/src/content/chapters/02-extraction.mdx

@@ -44,5 +44,5 @@ These alphabets are small by construction, against the tens of thousands of symb
 Even a 42-tool telecom customer service agent (tau2-bench telecom [@barres2025tau2]) needs only 43 states to capture its behavioral structure — the agent can only call the tools we gave it. So its capabilities stay bounded and so does its alphabet.
 
 <Note variant="info" title="How sensitive is the result to this choice?">
-$\phi$ is the one design decision in the pipeline — so we stress-test it. Fitness stays at or above 0.999 across all four granularities, from role-only (two to four symbols) to full tool-level — failure prediction stays within 0.03 [AUROC](https://texonom.com/c3fea8b9caa445768eb529fb629818c7) (area under the ROC curve) on any dataset — the rules above are one valid setting, not the only one.
+$\phi$ is the one design decision in the pipeline — so we stress-test it. Fitness stays at or above 0.999 across all four granularities, from role-only (two to four symbols) to full tool-level, and failure prediction stays within 0.03 [AUROC](https://texonom.com/c3fea8b9caa445768eb529fb629818c7) (area under the ROC curve) on any dataset. The rules above are one valid setting, not the only one.
 </Note>

app/src/content/chapters/03-construction.mdx

@@ -34,7 +34,7 @@ The SWE-agent prefix tree with 59,510 states collapses to just **25 states** —
 
 The FSM $\mathcal{M} = (Q, \mathcal{A}, \delta, q_0)$ encodes the agent's **behavioral topology**. Recurring patterns become loops. And the state count tracks the number of distinct behavioral modes.
 
-In the tau2-bench retail and telecom customer service agents, a tool-call loop (`assistant:tool_call` to `tool:text`) dominates execution, with the conversational path through `assistant:text` as a separate branch. We find that in a coding agent the `search`, `edit`, `execute` cycle accounts for most of the trace.
+In the tau2-bench retail and telecom customer service agents, a tool-call loop (`assistant:tool_call` to `tool:text`) dominates execution, with the conversational path through `assistant:text` as a separate branch. In a coding agent the `search`, `edit`, `execute` cycle accounts for most of the trace.
 
 <HtmlEmbed
   src="embeds/fsm-force-graph.html"

app/src/content/chapters/06-prediction.mdx

@@ -6,17 +6,17 @@ import Accordion from '../../components/Accordion.astro';
 
 # Prediction: Next Step and Failure
 
-We find the same FSM state answers both questions an operator asks: what the agent will do next, and whether this run is heading for failure. Both read off the per-state transition distribution, which compactness makes reliable.
+The same FSM state answers both questions an operator asks: what the agent will do next, and whether this run is heading for failure. Both read off the per-state transition distribution, which compactness makes reliable.
 
 ## Next-Step Prediction
 
-At each step the predictor estimates $P(a_t \mid \text{context})$, scored by [cross-entropy](https://texonom.com/fefc71fd930a41a7842f39bccb3abcc9) in bits via 5$\times$5-fold [cross-validation](https://texonom.com/184e7cfc80d6410188922d27ef4ef52c) (CV). Conditioning on FSM state alone, with no learning beyond an [order-1 Markov model](https://texonom.com/3bde15938b6b4c79981a3e0523b9877d), accounts for 83 to 99% of the total cross-entropy improvement on each dataset. We find the average drops to **0.93 bits**, a **62% cut** from the unigram baseline at 2.44 bits.
+At each step the predictor estimates $P(a_t \mid \text{context})$, scored by [cross-entropy](https://texonom.com/fefc71fd930a41a7842f39bccb3abcc9) in bits via 5$\times$5-fold [cross-validation](https://texonom.com/184e7cfc80d6410188922d27ef4ef52c) (CV). Conditioning on FSM state alone, with no learning beyond an [order-1 Markov model](https://texonom.com/3bde15938b6b4c79981a3e0523b9877d), accounts for 83 to 99% of the total cross-entropy improvement on each dataset. The average drops to **0.93 bits**, a **62% cut** from the unigram baseline at 2.44 bits.
 
 <Sidenote>
-We find RPNI goes the other way: 3.40 bits, worse than a unigram. Its state count, from hundreds to tens of thousands, spreads each transition estimate too thin, so the per-state distributions are noise.
+RPNI goes the other way: 3.40 bits, worse than a unigram. Its state count, from hundreds to tens of thousands, spreads each transition estimate too thin, so the per-state distributions are noise.
 </Sidenote>
 
-The cleanest test holds the predictor fixed and adds FSM state as a feature: under absolute discounting, FSM state conditioning adds **+0.155 bits on average** (0.580 vs 0.735) and helps on every one of six datasets, from +0.016 on SWE-agent to +0.364 on Mind2Web, whose branching web-action vocabulary gains most from knowing where in the workflow it is. We find that combining FSM state with a small learned model gives the best predictor in the study, at **0.73 bits**.
+The cleanest test holds the predictor fixed and adds FSM state as a feature: under absolute discounting, FSM state conditioning adds **+0.155 bits on average** (0.580 vs 0.735) and helps on every one of six datasets, from +0.016 on SWE-agent to +0.364 on Mind2Web, whose branching web-action vocabulary gains most from knowing where in the workflow it is. Combining FSM state with a small learned model gives the best predictor in the study, at **0.73 bits**.
 
 ## Workflow Memory
 
@@ -33,7 +33,7 @@ The gain generalizes to the agent's own LLM: feeding the current FSM state as co
 | ATBench | 600 | 47.8 | **62.5** | +14.7 |
 | OSWorld [@xie2024osworld] | 1,286 | 55.0 | **70.7** | +15.7 |
 
-The margin runs from +0.8 points on tau2-bench airline to +25.3 on SWE-smith — and it never reverses: on none of the eight does AWM come out ahead. Four of the eight gaps clear 15 points.
+The margin runs from +0.8 points on tau2-bench airline to +25.3 on SWE-smith, and four of the eight gaps clear 15 points.
 
 AWM extracts linear workflows from successful traces only, and it cannot represent cycles or branching.
 
@@ -62,7 +62,7 @@ Replay a trace through the FSM and read off **per-state behavioral features** (v
 Raw fitness is useless here (AUROC near 0.50): successful and failed traces both replay perfectly. But the signal is in the **per-state decomposition** and in *surprise*: failing traces take low-probability transitions under the FSM.
 </Note>
 
-Failure prediction scales with machine size: more states give a finer map of where a run can go wrong. The 43-state telecom agent tops out at **0.941**; WebArena [@zhou2024webarena] (0.903) and AgentNet [@wang2025opencua] (0.890) follow; SWE-agent [@yang2024sweagent] — with 25 states — reaches 0.799. ATBench, the only safety-labeled benchmark, reaches 0.894 (0.864 ± 0.024 under repeated CV). We see across all eight real-trace datasets the CV standard deviation stays in 0.012 to 0.031 — so these aren't single-split artifacts.
+Failure prediction scales with machine size: more states give a finer map of where a run can go wrong. The 43-state telecom agent tops out at **0.941**; WebArena [@zhou2024webarena] (0.903) and AgentNet [@wang2025opencua] (0.890) follow; SWE-agent [@yang2024sweagent] — with 25 states — reaches 0.799. ATBench, the only safety-labeled benchmark, reaches 0.894 (0.864 ± 0.024 under repeated CV). Across all eight real-trace datasets the CV standard deviation stays in 0.012 to 0.031 — so these aren't single-split artifacts.
 
 <Wide>
 <HtmlEmbed
@@ -74,7 +74,7 @@ Failure prediction scales with machine size: more states give a finer map of whe
 
 Inspect per-state feature importances and failure modes in the [failure view](https://seongland.com/article/asg/browser?tab=failure).
 
-And the predictors are interpretable. On SWE-agent the single strongest feature is whether the trace reaches the `submit` state: 94.8% of successes get there, only 55.7% of failures do. And we find this isn't a length proxy — structural features score 0.790 against 0.659 for trace length alone. Successful runs touch only 9 of 25 states along a focused `search`, `edit`, `submit` path — and failures spread across all 25 (Jaccard overlap 0.206).
+And the predictors are interpretable. On SWE-agent the single strongest feature is whether the trace reaches the `submit` state: 94.8% of successes get there, only 55.7% of failures do. And this isn't a length proxy — structural features score 0.790 against 0.659 for trace length alone. Successful runs touch only 9 of 25 states along a focused `search`, `edit`, `submit` path — and failures spread across all 25 (Jaccard overlap 0.206).
 
 ## Runtime Monitor
 

app/src/content/chapters/07-discussion.mdx

@@ -17,14 +17,15 @@ A 25-state machine can be read and checked by a person; the 59,510-state prefix
 
 Three lines of work touch this one.
 
-**State machines placed around agents.** StateFlow runs an agent through a state machine somebody wrote by hand [@wu2024stateflow], a definite finite automaton can be bolted onto a chatbot pipeline the same way [@sun2024dfa_llm], FlowMind builds a workflow from an API surface [@jin2024flowmind], and AFlow and ADAS search over agentic workflow designs [@zhang2025aflow; @hu2024adas], while MetaAgent assembles a whole multi-agent system out of one [@chen2025metaagent]. In every one of them the machine is an input — here it is an output, recovered from traces the agent had already produced. 
+**State machines placed around agents.** StateFlow runs an agent through a state machine somebody wrote by hand [@wu2024stateflow], a definite finite automaton can be bolted onto a chatbot pipeline the same way [@sun2024dfa_llm], FlowMind builds a workflow from an API surface [@jin2024flowmind], and AFlow and ADAS search over agentic workflow designs [@zhang2025aflow; @hu2024adas], while MetaAgent assembles a whole multi-agent system out of one [@chen2025metaagent]. In every one of them the machine is an input — here it is an output, recovered from traces the agent had already produced.
+
 **Experience kept as memory.** Agent Workflow Memory induces reusable routines from past trajectories [@wang2024agent_workflow_memory], Experience-to-Strategy trains a graph of them [@xia2025experience_to_strategy], ExpeL distils insights out of them [@zhao2024expel], and Voyager accumulates a skill library [@wang2023voyager]. Each of those keeps fragments of behavior. The automaton keeps the topology instead.
 
 **Automata learned from observation.** Model learning turns a black-box system into a state machine [@vaandrager2017model], Weiss and colleagues pull one out of a recurrent network with membership and equivalence queries [@weiss2018extracting], DeepDFA learns one by gradient [@umili2024deepdfa], prompt chaining extracts one from a flow description [@ali2024flowfsm], and AALpy packages the classical algorithms [@muskardin2022aalpy]. Process mining, which discovers process models from event logs, is the closest classical neighbour: van der Aalst's textbook treatment [@vanderaalst2016process], its re-thinking for the agent era [@berti2024processmining], an evaluation of large language models at the task [@grohs2024process_mining_llm], and skill learning from mined processes [@chen2024skill_process_mining]. This one works from a different input: positive traces only, no oracle to query, and an alphabet small enough that one merge suffices.
 
 ## Limitations
 
-The FSM accepts the observed prefix language, not the agent's true generating language: like any trace-replay method, it can't tell a trace that stays within the observed transition patterns from a legitimate one. The extraction function $\phi$ needs a small amount of per-domain knowledge — and fully automatic discovery of it is future work — failure prediction degrades on simpler machines: AUROC falls to 0.799 on SWE-agent and 0.70 on the 10-state SWE-smith — both are smaller task spaces, with less structure to exploit. Extending the workflow-memory comparison beyond AWM to other memory-injection methods is future work.
+The FSM accepts the observed prefix language, not the agent's true generating language: like any trace-replay method, it can't tell a trace that stays within the observed transition patterns from a legitimate one. The extraction function $\phi$ needs a small amount of per-domain knowledge, and fully automatic discovery of it is future work. Failure prediction degrades on simpler machines: AUROC falls to 0.799 on SWE-agent and 0.70 on the 10-state SWE-smith — both are smaller task spaces, with less structure to exploit. Extending the workflow-memory comparison beyond AWM to other memory-injection methods is future work.
 
 For agents with much larger action spaces or weaker sequential structure, the construction stays minimal but stops being compact — and the per-state observation density that drives every result above would degrade with it.