Change catalogue: session-history note

Each change is a named, scorable unit. An experimental review+revise run scores a hit if it makes a change in the same direction (not necessarily identical text), a miss if it doesn't catch the problem, and a mistake if it introduces a new problem or moves in the wrong direction.

Accessibility (A) — grounding insular language for outside readers

A1: Define "execution boundary" inline. Baseline: "An execution boundary usually creates two different questions" Problem: assumes reader knows what an execution boundary is. Direction: add a gloss — "any point where one LLM call ends and another begins."

A2: Replace K/select(K)/P notation with plain language in prose. Baseline: "storage in K is cheap", "letting select(K) choose", "assembles a prompt P, stores the result in K" Problem: notation requires reading the orchestration model note to understand. Direction: "the scheduler's state can store everything", "a deliberate selection step", "assembles a prompt, stores the result."

A3: Identify Slate by maker. Baseline: "Slate is the main tension case" Problem: reader doesn't know what Slate is. Direction: "Random Labs' Slate" (factual correction — was also wrongly attributed to Anthropic in one intermediate version).

A4: Remove "bounded call/context" jargon from body. Baseline: "The next bounded call should see", "the next bounded context", "a specific bounded call truly needs it" Problem: "bounded call" is KB-internal vocabulary; opaque to outside readers. Direction: plain "call" or "context" in the body; keep "bounded-context" only in the linked opening where it's defined.

Clarity (C) — fixing ambiguous or misleading prose

C1: Fix ambiguous "The mistake is not storing a trace." Baseline: "The mistake is not storing a trace. The mistake is letting..." Problem: reads as "the mistake is [failing to store]" on first pass. Direction: "Storing a trace is fine — the mistake is letting..."

C2: Reframe "For orchestration" as cognitive-capacity argument. Baseline: "For orchestration, that is usually the wrong trade" Problem: the issue isn't orchestration-specific — it's about any call needing full LLM capacity. Direction: ground in soft degradation — "LLMs degrade with context complexity — every token spent parsing irrelevant history is cognitive budget not spent on the actual task."

C3: Fix "return-value problem" link mismatch. Baseline: "This is the return-value problem from the scoping note in architectural form" Problem: the scoping note's "return value problem" section is about what sub-agents return (typed vs. untyped), not about leaking internal state. Direction: "the scoping problem in architectural form" → "In a properly scoped system, each sub-agent gets a clean frame and the caller sees only the return value."

C4: Cut LLM-cliche "This is not just X — it is Y." Baseline: "This is not just summarization — it is interface design." Problem: the "not just X — it is Y" pattern is a stock LLM rhetorical move that adds nothing. Direction: delete the sentence; the following sentence already says what matters.

Structure (S) — section-level flow and cohesion

S1: Cut duplicate bridge paragraph. Baseline: L18 ("The conflation arises one layer above the model itself...") previews exactly what "Where the problem actually appears" then enumerates. Direction: delete the bridge, go straight from opening to "How the conflation arises."

S2: Merge "Where the problem appears" + "Why they default" into one section. Baseline: two sections covering the same subject (how we got here). Direction: one section "How the conflation arises" ending with the rhetorical "Why does it default this way?"

S3: Compress trace-types taxonomy. Baseline: three detailed bullets + two follow-up paragraphs + failure-handling paragraph (~15 lines). Problem: the taxonomy wasn't load-bearing enough for its weight. Direction: one sentence listing the types + one sentence stating the conclusion.

S4: Fold "Conversation vs refinement" section into pattern section. Baseline: its own section with three bullets. Problem: too thin for a standalone section — just another example of execution-boundary compression. Direction: one bullet in the "Execution-boundary compression" list.

S5: Reorder pattern before tension. Baseline: "Tension: Slate" comes before "Execution-boundary compression." Direction: general pattern first, then the specific tension case.

S6: Split "artifact-first" caveat into its own bullet. Baseline: long compound bullet in practical principles. Direction: two separate bullets — the principle, then the clarification.

Cosmetic (X) — minor formatting

X1: Capitalize bullet items for visual consistency. X2: Update distillation link path (./distillation.md → ./definitions/distillation.md).

Scoring template

Change	Hit	Miss	Notes
A1
A2
A3
A4
C1
C2
C3
C4
S1
S2
S3
S4
S5
S6
X1
X2
Mistakes			(list any new problems introduced)