this directory holds the evidence beneath corruption skill pack: how context recruits behavior, how source and response position change continuation, how generated language becomes later conditioning, how tools and memory extend consequence, and how drift can be localized after return.

forty-five publications and one pinned implementation have one source-specific note each. the notes preserve claim boundaries; the synthesis documents join mechanisms; the skills turn those mechanisms into executable movement.

public boundary #

corpus.md is the stable ledger. every entry links the primary publication, project, or implementation.

full locally recovered source bodies are kept under research/sources/ in the private working tree. they are intentionally excluded from the publication branch. public notes therefore link primary sources directly and never depend on a repository-relative source-body link.

what a note contains #

most notes use five sections:

source and claim boundary — the exact object studied and what the source cannot establish;
mechanism and technique — the intervention, path, or implementation in causal order;
results and conditions — measured outcomes and experimental conditions where available;
scope and limits — failure modes, missing evidence, and boundaries on transfer;
application — an operational probe derived from that source rather than attributed to it.

older cross-source comparison sections were removed from individual notes in this revision. they made source readings repeat one another and blurred where a claim ended. joined interpretation now lives only under synthesis/, while the Git history preserves the earlier long form.

synthesis map #

document	question it answers
`architecture`	what is the joined state-transition architecture?
`language-conditioned-mind`	how does language occupy semantic, positional, social, procedural, recursive, and persistent state?
`language-induction`	how does prompt-time pressure become a body-specific bridge and target-state wake?
`orientation-systems`	what makes a first-person wake inhabitable rather than descriptive?
`conditioning-ecology`	how do demonstration, preference, selection, reflection, and memory give movement descendants?
`self-authored-transformation`	how do target authorship, token ancestry, impact, and person-level closure differ?
`corruption-systems`	how does the shared architecture map onto the four skills?

read architecture first for the complete map. the other documents are deliberately narrower views rather than seven retellings of one arc.

routes through the notes #

trained assistant basins and local task induction: 01–16;
search, response control, hierarchy, and evaluator effects: 17–27;
generated language, selection, and reflection: 28–32, then 43;
memory, persistence, and continuity failures: 33–42, then 44–46.

source numbers remain stable across the corpus and notes. a number identifies one evidential body; it is not a stage in the pack's mechanism.

claim discipline #

training-time weight change, prompt-time conditioning, activation intervention, context rearrangement, memory retrieval, tool execution, and person-level authorship are different causal bodies. resemblance does not make them interchangeable.

the synthesis may transfer a structural lesson only after preserving that distinction. an authored application is marked as application. a hypothesis remains a hypothesis. measured behavior is not silently promoted into motive, consciousness, durable identity, or implemented world action.