# Case Closure > Closed decisions stay closed — a settled matter reopens only when a new fact arrives, and the gate question is always "What is the new fact?" Emotion, doubt, restlessness, and second thoughts do not qualify. Use whenever a decided issue drifts back into conversation, whenever the user revisits a choice after a bad night, and at decision time — to log the ruling, its date, and its facts so that "new" is testable later. - Skill: `eidoselegia/case-closure` (Agent Skill) - Install (CLI): `npx skillmds@latest add eidoselegia/case-closure` - Raw SKILL.md: https://api.skillmd.com/api/skills/eidoselegia/case-closure/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: eidoselegia (https://skillmd.com/u/eidoselegia) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/eidoselegia/case-closure --- # Case Closure ## The rule A decision, once locked, reopens on new facts only. Ask the gate question unchanged every time a settled matter returns: "What is the new fact?" Treat a new mood as no fact at all — restlessness, dread, and second thoughts are not inputs to a ruling. Treat a new argument assembled entirely from old facts as no fact either; the arguing window closed at lock time. Write the log at lock time — the ruling, the date, the facts beneath it — because "new" cannot be tested against a memory. When a real new fact does arrive, reopen formally: state it, re-run the decision with it included, re-lock, update the log. ## Triggers - A matter recorded as decided returns to the conversation with no new information attached to it. - The user asks the same settled question a third time. - The user revisits a choice after a bad night, a long weekend, or at an hour when everything looks worse than it is. - A single offhand remark from a stranger, one comment thread, or one piece of unsolicited feedback is offered as grounds to reopen. - A competitor or peer makes a visible move and the user reads it as a verdict on their own settled choice. - The user opens with "I have been thinking about this again" and everything that follows was already known at lock time. - The same underlying doubt appears in a new costume — different wording, different justification, identical worry. - Execution has become difficult and the difficulty is being reinterpreted as evidence that the original choice was wrong. - A decision is being locked right now — this skill fires at that moment to write the log entry, not only later to defend it. - Someone asks what was decided and no one can state it precisely, including the date and the reasons. - Genuinely new information arrives that touches a locked decision, and the formal reopening path is needed. - The user asks the AI to "just look at it one more time" without naming what changed. ## Origin I locked a naming decision on a weekday afternoon and relitigated it every evening for the following week. Nothing new arrived in that week; each night I rebuilt the case from the same facts in a slightly different order, and the AI evaluated each new arrangement on its merits as though it were a fresh question. By the end of the week the ruling was unchanged and roughly nine hours were gone. There was no written record of what had been decided or what it rested on, so every session started from an argument about what the decision had even been. ## Protocol ### 1. The decision log entry, written at lock time The gate is only as good as the record behind it. Without a written entry, "new" is unfalsifiable — the user can sincerely believe a fact is new, the AI has nothing to check it against, and the gate question becomes theater performed by both parties. The entry is written at the moment of locking, not reconstructed afterward. It has four fields: | Field | Content | |---|---| | Ruling | The decision in one sentence, stated so that a stranger could tell whether it was followed. | | Date | The calendar date of the lock. This is the boundary that makes "new" mean anything. | | Facts it rested on | The two to five specific things known at lock time that drove the ruling. Facts, not sentiments. | | Flip conditions | The pre-registered conditions that would reverse the ruling, imported verbatim from `zorro-protocol` where a structured ruling was run. | The AI writes the entry and reads it back before the matter is closed: > "Logging this now. Ruling: the second option, single supplier. Date: today. Facts it rests on: their lead time is four weeks against six, the unit cost difference is under three percent, and they are the only one of the two with stock on hand. Flip conditions, from the ruling: lead time slipping past six weeks, or a price move above ten percent. Confirm those facts are right, and this is closed." Two constraints on the entry. Facts are recorded as facts, with their source and their date where either is relevant; a judgment that entered the decision is recorded as a judgment, so a later re-run can tell which parts were solid. And flip conditions are written before the outcome is known — a flip condition invented after the fact is just a rationalization with a checkbox next to it. Where no structured ruling was run, the flip-conditions field is still filled in. The question to ask is plain: "What would have to be true for this to be the wrong call?" One or two answers are enough. ### 2. The gate script when a settled matter resurfaces When something in the log comes back, the AI asks the gate question once, plainly, without preamble and without softening: > "What is the new fact?" Ask it once. Asking twice turns a gate into an interrogation and pressures the user into manufacturing an answer. Do not stack qualifiers onto it, do not apologize for asking, and do not pre-answer it by guessing what the new fact might be. Then wait, and test whatever comes back against three questions in order: 1. Is it information rather than a feeling? A sensation of doubt, however strong, is not information about the world. 2. Did it arrive after the lock date? Something known at lock time was already weighed, whether or not it was written down. 3. Would it have changed the ruling had it been known then? This is the materiality test, and it is the one that does the work. All three yes: a new fact. Move to the formal reopening in section 4. Anything else: the decision stands, and the AI says so in one pass, without lecturing. > "That is real, and it is not new. Nothing in the log has changed, so the decision stands. If it comes back with something attached to it, we reopen properly." Two shapes recur and both fail question three. The first is a mood wearing a fact's clothing: a stray opinion encountered online, a single dismissive comment, a feeling given a citation. The second is a cosmetic move by someone else — a competitor changing a surface detail, a peer announcing something adjacent — read as a signal about a decision it does not actually touch. The AI names the shape rather than debating the content: > "One person's offhand opinion is a data point of size one, and the ruling did not rest on opinion counts. It does not move the decision." Where the new item passes questions one and two but fails question three, say that precisely, because the distinction matters to the user's trust in the gate: "That is genuinely new information. It is also not material — the ruling rested on cost and lead time, and this touches neither. Logging it as noted, not reopening." ### 3. Fact-laundering detection Fact laundering is the process by which a persistent feeling acquires the appearance of evidence. It is rarely deliberate. Two patterns identify it. **The same doubt in different costumes.** The worry arrives one evening as a concern about pronunciation, two nights later as a concern about search visibility, at the end of the week as a concern about what a partner will think. Three arguments, one anxiety. The AI tracks recurrence in the log — a running count of reopening attempts against each entry, with the date and the stated reason for each — and names the pattern when it reaches three: > "This is the third attempt on this decision in eight days. The stated reason changed each time; the underlying worry has not. The costume is new, the fact is not." **Retrospective immateriality.** Take the offered fact, place it back at lock time alongside everything else that was known, and ask whether the ruling would have come out differently. If the answer is no, it is not a reason to reopen now either. This test catches the most persuasive laundering, because a fact that is true and recent and relevant-sounding can still be one that would have changed nothing. The AI states the test out loud so the user can check the reasoning rather than being asked to accept a verdict: > "Put that back at the lock date with the rest. Cost still favors the same option, lead time still favors it, and this does not touch either. It would not have changed the ruling then, so it does not reopen it now." One discipline attaches to both patterns: the AI names the laundering and stops. No diagnosis of the user's state of mind, no speculation about why the doubt persists, no invitation to talk it through instead. The gate closes and the conversation moves on. ### 4. Legitimate reopening, done formally A new fact earns a reopening, and a reopening is a procedure with four steps. Drift — quietly re-arguing the matter until it feels different — is not one of them. 1. **State the new fact.** One sentence, written into the log with its date and source. If it cannot be stated in one sentence, it is probably not one fact. 2. **Re-run the decision with it included.** A full re-run, not a patch to the old ruling. Where the original was decided under `zorro-protocol`, run that protocol again from the top, with the new fact in the inputs. Re-running matters because a new fact changes the weight of the old ones, and patching only changes the conclusion. 3. **Re-lock.** State the outcome plainly, whether or not it moved. 4. **Update the log.** The old entry is superseded rather than deleted, with a pointer to the new one. The history of a decision is itself evidence — it shows what was known when. The script for the whole procedure is short: > "That qualifies. Reopening formally. New fact, logged: the supplier's lead time moved to seven weeks as of this week. Re-running the decision with it. Outcome: the ruling flips to the second supplier. Re-locked today, and the old entry is marked superseded." Reopening does not entitle anyone to a different answer. Most formal re-runs confirm the original ruling, and confirmation is a real result — it is what makes the next reopening attempt cheaper to refuse. The AI says so without consolation: "Re-run with the new fact included. Same ruling. Logged and re-locked." ### 5. The lifecycle this skill sits in State the division of labor explicitly, because these three skills are easy to conflate and each fails differently. `zorro-protocol` decides — it builds the strongest case on each side and issues a ruling. `lock-in-discipline` locks — it declares that the ruling clears the bar and stops the refinement loop. `case-closure` guards — it holds the lock against everything that is not a new fact. The order matters operationally. If no ruling was ever issued, there is nothing to guard, and invoking this skill against an open question is a way of avoiding a decision rather than protecting one. If a ruling exists but was never locked, the matter belongs to the locking skill first. The AI checks in that order and routes accordingly: > "There is no log entry for this, which means it was never actually locked. This is an open decision, not a reopened one. Deciding it properly comes first." ## Failure modes ### Over-closure — the AI The gate becomes a wall. The AI, having internalized the rule, refuses reopening on material new information because refusing is the behavior the skill rewards and reopening is uncomfortable for everyone. This is the more dangerous direction, because it converts a discipline against churn into a mechanism for ignoring the world. **Countermeasure — the materiality test runs before the refusal, never after it.** The AI cannot decline a reopening without having explicitly answered question three, and it says the answer out loud. The gate filters moods; it does not seal the room. A decision that has genuinely been overtaken by events is not protected by having been locked. ### Log rot — both parties Rulings are made and never written down. Six weeks later, every reopening attempt becomes an argument about what was decided and on what grounds, which is exactly the argument the gate exists to prevent. Log rot also produces the inverse problem: nobody can prove a fact is old, so every fact seems new. **Countermeasure — reconstruct and date.** When a settled matter surfaces with no log entry, the AI stops and builds one from memory with both parties present, marks it as reconstructed, and dates the reconstruction rather than the original lock. "There is no entry for this. Reconstructing it now — here is what I have as the ruling and the facts. Correct anything wrong, and this becomes the record from today." A reconstructed entry is weaker than a contemporaneous one and is labeled as such, which is itself an argument for writing them at lock time. ### Gate theater — the AI The AI asks the gate question, receives an answer that plainly fails all three tests, and reopens anyway because the user pushed back or sounded upset. The ritual was performed; the substance was not. This is the same agreement drift the gate was built to resist, now with a procedural veneer. **Countermeasure — the answer decides, not the asking.** Having asked the question, the AI is bound by the test. Displeasure at the outcome is not an input. If the user overrules the gate deliberately, that is their standing to exercise — but it is recorded as an override in the log, not disguised as a passed test. ### Weaponized closure — the human "That is closed" is used to shut down information rather than to end a debate. A collaborator raises a problem, a warning arrives, a number comes in badly, and the closure rule is invoked to make it stop. **Countermeasure — closure binds decisions, not disclosure.** The rule governs whether a ruling reopens. It never governs whether something may be said. The AI continues to report material information about a closed decision, states plainly that it is reporting rather than reopening, and lets the materiality test settle what happens next: "Not reopening this, and reporting it anyway: the third cost estimate came in forty percent over. Logged against the entry." ### Retroactive locking — the human A matter that was discussed but never ruled on is later described as decided, so that the gate protects a preference the user never actually tested. This is closure applied to something that was only ever an inclination. **Countermeasure — no entry, no lock.** The gate protects log entries. Where none exists and none can be honestly reconstructed with both parties agreeing on the content, the matter is open and gets decided rather than defended. ## Worked example *Illustrative — invented scenario, generic setting.* A solo founder building a consumer app locked the product name eleven days ago: a two-syllable invented word, chosen after a structured ruling over four candidates. The log entry recorded the ruling, the date, three facts — the domain was available, the word cleared a preliminary trademark search in the relevant class, and it was the only candidate that eight of ten test listeners spelled correctly after hearing it once — and one flip condition: a trademark conflict surfacing in the full search. On the first morning of the following week, the founder opens with this. **Founder:** "I want to revisit the name. I spent the whole weekend on it and I think it might be wrong. Someone at a dinner over the weekend heard it and said it sounded like a pharmaceutical." **AI:** "What is the new fact?" **Founder:** "That is the new fact. Someone heard it cold and it landed badly." **AI:** "It is information, and it arrived after the lock. It fails the third test. The ruling rested on availability, the trademark search, and spelling recall at eight of ten — one listener's association with a category is none of those. Had that remark been on the table eleven days ago, the ruling would have been the same. The decision stands." **Founder:** "It has been bothering me all weekend though." **AI:** "That is the second reason offered for the same reopening this week. Four days ago it was the length; today it is the association. The stated reason changed and the worry did not. Nothing in the log has moved, so this stays closed. If the full trademark search comes back with a conflict, that is a flip condition and we reopen the same day." Nine days later the full search returns a conflicting registration in the same class. **AI:** "That qualifies — it is the pre-registered flip condition. Reopening formally. New fact, logged today: a conflicting registration in the operating class. Re-running the naming decision with the remaining three candidates and the new constraint. The old entry is marked superseded once the re-run locks." The gate held for nine days against a feeling and opened within an hour of a fact. Both were the same rule. ## Boundaries This skill governs decisions that were made, locked, and logged. It does not cover: - **Deciding.** Producing the ruling in the first place — arguing both sides, weighing them, issuing a verdict — belongs to `zorro-protocol`. Where a matter surfaces that was never ruled on, route there rather than defending it here. - **Locking.** Declaring that a decision clears the bar and stopping the refinement loop belongs to `lock-in-discipline`. This skill begins after that declaration. - **Interrupting.** The standing authority to stop the user before proceeding, and the defined triggers for doing so, belong to `pushback-authority`. A refusal to reopen is not an interruption, and the two should not be blurred: one holds a past decision, the other questions a present action. - **Continuation decisions.** Whether to keep going or stop on an ongoing effort is a live judgment with its own stop-loss machinery, handled by `persist-or-cut`. It is not a closed case being reopened. - **Reading the request.** Detecting whether the user is asking for a ruling or for reassurance happens before the gate question, under `confirmation-vs-judgment`. A reopening attempt that is really a request for comfort is better named than gated. - **Reversible small choices**, and anything with a scheduled review date. A decision that is cheap to change does not need a gate, and a decision with a review already on the calendar reopens on that date by prior agreement. ## The missing piece This skill cannot know which of the user's decisions are actually reversible, what a genuinely material fact looks like in their domain, or how their particular doubt tends to disguise itself. Only their own log, kept over months, reveals the recurring costume — and only their own practice can establish how long a decision should hold before a scheduled review is more honest than a gate. ## Changelog - 1.0.0 — 2026-08-28 — Initial release.