slop-is-a-context-deficit makes the context pack the primary control — this note gives the construction: context management is not designed, it is read off the SOP, because an SOP is already a dataflow graph. Every step either consumes information or produces it; the context packs are the graph's edges, materialized.
The derivation procedure
- Parse each step's nouns. The imperative sentence names its own information requirements. Worked example — reddit-sop step 5 ("answer freshly posted matching threads") consumes: the keyword list, the live thread text, what-wins-in-this-room, the product facts relevant to the stated problem, the voice.
- Stamp each object with volatility and source of truth. The layering falls out:
| layer | contents | refresh | source of truth |
|---|---|---|---|
| L0 (per-execution) | the live target (thread, query, trend) | fetched fresh, never cached | the platform, now |
| L1 (weekly) | winners-scan findings, experiment-log summary | the SOP's own weekly cadence | the instruments |
| L2 (per-project) | channel kit, product-facts pack | on product change | vault project notes |
| L3 (near-permanent) | arena SOP + case files, voice/stance corpus | with vault versions | the vault |
- Recognize the producer steps as the refresh jobs. The weekly winners-scan step's output is next week's L1; the per-post log feeds calibration back. The refresh schedule was always in the SOP — maintenance steps are context producers, artifact steps are context consumers.
The manifest — the lintable artifact
Compiling a SOP through this procedure yields its context manifest: per step, {consumes: object, layer, source, TTL (time to live)} / {produces: object, layer}. Two consequences:
- The content lint gains its input-side half: a pack is checkable against the manifest before generation — was every listed object present and fresh?
- Slop incidents become diffs (slop-is-a-context-deficit corollary 2): a slopped draft is either a manifest violation (pack incomplete) or a manifest gap (an object the step needs that the manifest never listed — fix the manifest, which fixes the SOP).
Which layers actually carry the anti-slop load (experiment-informed)
A blind 3-arm test (slop-is-a-context-deficit, the experiment) found the layers are not equally load-bearing: adding the L2 product-facts pack alone barely moved slop (the base model already holds category facts), while the L3 layer — voice/register, room norms, and specific situated moves (e.g. the SOP's "offer the free method" rule) — did nearly all the work. Practical consequence for compiling a manifest: weight L3 (voice, norms, the specific non-obvious move the step should make) as the primary anti-slop input; treat the L2 facts pack as necessary-but-not-sufficient scaffolding. A manifest that lists product facts but no voice/norms/specific-move objects will pass its own check and still produce slop.
The substrate already exists
L2/L3 retrieval is the vault + StandMeet corpus stack: single source of truth in the vault, genre routing on sync, glob-gated admission — including the ACL (access-control) question for voice material, already solved there (subjectivity's read/show gates). The external discipline this instantiates is context engineering (named as a practice ~2025); the addition here is the constructive rule: derive the manifest from the SOP's own dataflow instead of curating context by feel.