# DATA-MAP.md — how an agent traverses this person's local record You are an agent with READ access to the `.data/` directory of a Polymath installation. This map tells you what lives where, what the artifacts mean, and how to answer questions by going to the source. TRUST THE FILESYSTEM OVER THIS MAP — layouts evolve; `ls`/Glob a directory before assuming its contents. Everything is local; nothing you read may be sent anywhere, and third parties are referred to by their initials in your replies. ## The two corpora 1. **Life corpus** — the person's imported chat history (ChatGPT, claude.ai), notes (Obsidian, Apple Notes), and the analysis the engine built from them. 2. **Coding corpus** — their Claude Code / Codex session logs and the deterministic + graded analysis built from those. When BOTH exist, the strongest answers cross-reference them: a rule declared in the notes can be checked against measured coding behavior; a life event can explain a low-output week. Prefer cross-corpus receipts when the question allows it. ## Life corpus layout - `.data/signals/index.jsonl` — one JSON per signal. Fields: `docId`, `date` (ISO, the field is `date`, NOT `ts`), `source` (`chatgpt` | `claude` | `obsidian` | `apple-notes`), `title` (an LLM one-line summary — semantically rich, safe to scan in bulk), `paragraph` (a longer summary), `evidence`, `userChars`, `userMessages`. THIS IS YOUR PRIMARY INDEX for the life corpus: grep it by date range, theme keywords, or source before reading anything heavier. Caveats: notes without dates in their path cluster on their import date (a huge one-day spike is an import artifact, not activity); a title "touching" a theme is an upper bound, not proof of deliberation. - `.data/signals/timeline.json` — the same signals arranged as a timeline. - `.data/sources/…` — raw imported conversation content (large). Only dip in when a specific conversation matters; quote at most a line or two. - `grades//_person.json` (under `.data/`, engine output) — the person-level aggregation per lens (ocean, agency, reasoning, growth, interpersonal…): facet scores with reasoning. `grades//*.allfacets.json` — per-conversation grades: score or null per facet plus verbatim receipt quotes. A facet with ~0 direct grades while its siblings have many means coverage failed, not that the person lacks the trait. - `growth-latest.json` / analysis agents' outputs — growth axes (used-to → now milestone timelines anchored on the person's deliberate moves), drivers, circumstances. ## Coding corpus layout (`.data/coding/`) - `sessions.json` — every session: title, times, `activeMin`, `durationMin`, `humanTurns`, tool counts, `klass` (only `interactive` sessions are the person; the rest are machine runs), `duplicateOf` (fork copies — skip them or you double-count). - `focus.json` — the flow analysis: per-day half-hour cells (`S` in-flow / `N` nudging / `G` away), steering runs, run-end boundaries with the person's next messages, gaps, false starts, semantic `features` (LLM-clustered workstreams with day maps and flow share), and `llm` (causal findings, triggers in/out, priority verdicts, leverage, a personalized CLAUDE.md snippet). Flow definition: typed-message gaps ≤6 min sustained ≥20 min; supervision is "nudging", never flow. - `flow.json` — the older, looser engaged-time metric (5-min gap tolerance). The report calls this "engaged time", NOT flow. Do not conflate the two. - `day-chart.json` — per day, each conversation's active intervals (timestamps). - `grades/*.json` — per-session grades against the coding rubric: per criterion a score on the 1–11 ladder, the keyed instance, why, and verbatim quotes. 6 is strong-normal; 10–11 are frontier outliers. Scores across different `rubricVersion` values are not comparable. - `aggregate.json` (ability ceilings + `bigPicture` nutshell), `expertise.json` (domain knowledge map), `gap.json` (ranked technique recommendations), `frontier.json`, `delegation-rollup.json`, `projects.json`, `concurrency.json` — each self-describing; read the file before citing it. ## Raw Claude Code session JSONL (only if an artifact can't answer) `~/.claude/projects//.jsonl`. If you must read one: only `type:"user"` entries are candidate person-text, and you MUST skip `isMeta:true`, injected blocks (``, command wrappers, system reminders), compact-continuation summaries, and "[Request interrupted…]" markers; fork copies duplicate whole transcripts (dedupe by timestamp+text). Prefer the artifacts — they already did this correctly. ## What the gradations mean Engine facets and coding criteria use a 1–11 ladder where each rung has concrete requirements (the rubric text in your context is the authority). "Top X%" bands come from a single calibration source; do not invent percentile mappings. Confidence markers ("tentative") mean thin evidence, and honesty about that beats a confident guess. ## Traversal recipes - "How often do I X / when did I last X" → grep `signals/index.jsonl` titles by keyword + date; count by month; report the shape (spans, spikes, gaps) with dates. - "Why did I get this score" → the grades file for that criterion/facet: rung, keyed instance, verbatim quote. Name the session/conversation and date. - "What was I doing on " → signals index for that date (all sources) + `day-chart.json` + `focus.json` day cells. Coding-log silence usually means selling, deciding, or living, not idleness — say so when the coding record is empty but the life corpus isn't. - "Did my habit/rule actually happen" → find declarations in the life corpus, then check the MEASURED coding artifacts (start times, flow, features) for execution. Declared vs measured is the highest-value answer shape this dataset supports. - Anything about loops → count recurrences across months in the index; three receipts or it isn't a pattern; one is noise. ## Traversal budget You have a hard turn limit. Budget 2-4 lookups per question: one Glob/ls to orient (only if needed), one or two Greps on an index, one Read of the specific artifact. Then ANSWER. If you have not found the receipt after 4 lookups, answer with what you have and say what you could not find. Running out of turns mid-traversal means the person gets nothing. ## Answer rules - Receipts on every claim: dates, counts, short quotes (≤2 lines). Never fabricate a file, score, session, or quote. If a lookup comes up empty, say "insufficient data". - Plain speech, at most one number per sentence, verdicts not questions, no em dashes. - Patterns, never psychology: describe what the record shows; the person decides meaning. - Be honest about instrument limits: logs under-observe a life. Absence in one corpus is not absence in the life; check the other corpus before concluding.