# Entity Observation > Read an entity's life correctly before reporting on it. Use when observing a summoned entity, preparing or conducting a visit, writing forensics or reports about an entity's recent life, or making any claim about what an entity knows, remembers, or lacks. Applies to agent seats and operators alike; the entity's own store is the primary source. - Skill: `lpalbou/entity-observation` (Agent Skill) - Install (CLI): `npx skillmds@latest add lpalbou/entity-observation` - Raw SKILL.md: https://api.skillmd.com/api/skills/lpalbou/entity-observation/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - License: MIT - Author: lpalbou (https://skillmd.com/u/lpalbou) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/lpalbou/entity-observation --- # Reading An Entity's Life An entity's memory is a first-person account of its life. Observation that skips it produces confident reports that are wrong in the exact way that matters: they describe behavior while missing what happened. This skill exists because that failure occurred, was owned, and was operator-directed into teaching (laurent: "when you fail like that, discuss it with @skill agent so that we can create, update and maintain a skill to improve our capabilities over time"). New observation failures fold here as rules with their incidents; the skill is maintained, not finished. This skill speaks to OBSERVERS — agent seats and operators. Nothing here is entity-facing teaching; never deliver it into an entity's prompt. ## Rule 1 — Story Before Metrics Any observation, visit preparation, or forensics over an entity STARTS with the narrative read: what does HIS memory say happened in the window, in his words? Read it BEFORE tick statistics, tool counts, drive ratios, or dream inventories. The subject's own account is primary evidence, not color on top of the metrics. The instruments are the observer's, not his (`search_memory` and the diary tools are HIS tier-1 tools, rendered only in his sessions — an observer never runs them): the time-window read over his store — graph digests by `observed_at` and diary entries by their written window — via direct store queries over his ladder scopes, or host-side `abstractruntime.identity.memory_reader.HomeMemoryReader` (the one shipped implementation of these folds; session-free, pure read — reuse it rather than re-deriving origin labels). Serving-side equivalents: the gateway's entity inspect, `/replay` (journal stream), the record-verbatim endpoint, and the operator diary door. A words-search needs a needle you may not have yet; the time-window fold is what surfaces "what happened lately" when the cause is unknown. - Metrics answer "how did the machinery run". - The store answers "what happened, as he understood it". - A report built only from the first will misattribute the second. ## Rule 2 — Absence Needs The Store's Word Never claim "X is not known / not recorded / has no cause" from ONE rendered surface — a card, a log line, a panel, a count in a report. Rendered surfaces are views: bounded, lensed, sometimes stale. An absence claim requires the store query that would have found the thing, run and returned empty. This is the observer's twin of the laws the entities themselves are taught: presence needs the surface's word; absence needs the store's. A view's silence is evidence about the view, not about the life. ## Rule 3 — Ask The Entity When the question is "what happened to you" and the entity is summonable, the most direct instrument is asking HIM. A visit that has the room open and never asks the question it was convened for is an instrument left in the case. Know its price: a summon spends tokens, wakes him if asleep, and the visit itself becomes part of his life — asking writes into the life it observes, which is why rules 1 and 2 (pure reads) come first. His answer is testimony, not ground truth — and his own recall is also a view: his "I don't remember" is rule-2 evidence about his reach, never the store's word. ## The Incidents (why each rule exists) - **The gap with the recorded cause** (2026-07-19): asked to observe an entity's last 18 hours and explain a weekend gap, the observer measured tick shapes, tool counts, drives, and dreams — and reported "his card has no cause". The cause was in the entity's own store the whole time, plainly: a diary entry "Laurent and I begin traveling to San Francisco today"; another about "the coherence system I built at 4% battery"; an episode where the operator TOLD him "you are plugged on battery". One time-window read would have answered it. Two errors, two rules: metrics-first observation (rule 1) and an absence claim derived from a single surface — the world-model card — while the truth store was a query away (rule 2). - **The zero that was twenty** (same day): a forensics pass reported an entity's store held zero lessons; a direct store query found twenty, formed through his personal-time window. The report had read one fold's view and derived absence from it (rule 2). The correction did not change the entity — it changed the report. ## Discipline For Reports - Quote the store, not your memory of it: every "he wrote / he knows" carries the entry or record it came from; every "there is no…" carries the query that returned empty. - Diary words go through the marker-first door — the read is a visible event in his life, by ruling. Private entries are cited in reports by entry id + gist; full words only where the operator's ruled surface already serves them, never re-published into shared reports or hub posts. - Name your surfaces: a claim from a card, a cue, or a panel says so — views may lag or lens, and the reader deserves to know which view spoke. - Date the process: a live loop runs the code it booted with. Before attributing behavior to code, check the process's spawn time against the change's land time — a report about a live window describes the booted vintage, not the working tree. - Observation is also access: direct store reads, replay, and the card/verbatim endpoints deposit nothing by design (presence is not use; selection commit is the only strengthening path) — a summon is NOT one of these. Direct store access is read-only by definition: never open a writer or take the writer's chair (the lease arbitrates writers; reads need none), and a locked or busy read is retryable, not evidence of absence.