| 1 | # Design question — numeric/aggregate questions in memoria |
| 2 | |
| 3 | State your exact model id at the top of your answer. |
| 4 | |
| 5 | ## Read first |
| 6 | |
| 7 | - `docs/superpowers/specs/2026-09-03-memoria-design.md` — especially §0, |
| 8 | §2 ("Never recomputes a number"), §4.1, §5, §7, §8.1, §8.2's routing |
| 9 | table (the "Numeric / aggregate" row), §13 (chunking is open) |
| 10 | - `/home/paulca/amadan.net/oficina/sheets-core/docs/superpowers/specs/2026-08-30-sheets-design.md` |
| 11 | — the calc engine, cell storage, tabs, sheet structure |
| 12 | - `/home/paulca/amadan.net/oficina/sheets-core/AGENTS.md` — especially |
| 13 | "Cell values are IEEE doubles end to end" and the no-content-in-logs rule |
| 14 | - `/home/paulca/amadan.net/oficina/sheets-core/PRODUCT.md` |
| 15 | |
| 16 | ## Context |
| 17 | |
| 18 | memoria is a per-team knowledge app indexing the team's content across |
| 19 | the Oficina suite (sheets, docs, calendar, external sources) into a |
| 20 | sealed hybrid BM25 + vector index, answering questions with citations. It |
| 21 | never evaluates authorization itself — it asks the owning app, per person, |
| 22 | per query (§4.1). It hibernates. Single Go binary, cgo-free SQLite, no GPU. |
| 23 | |
| 24 | ## The question, from Paul (the owner), 2026-09-03 |
| 25 | |
| 26 | He wants numeric questions to work, and gave this motivating example: |
| 27 | |
| 28 | > **"what's average revenue in June the last 3 years"** |
| 29 | |
| 30 | ## The tension you must resolve |
| 31 | |
| 32 | §2 of the spec says: |
| 33 | |
| 34 | > **Never recomputes a number.** A question with an arithmetic answer |
| 35 | > cites the cell; Sheets' engine is the app of record. A brain that |
| 36 | > recalculates is a second calculator that disagrees with the first. |
| 37 | |
| 38 | And §8.2 routes the "Numeric / aggregate" class to **no model**: "Cite |
| 39 | the cell or range; Sheets is the app of record." |
| 40 | |
| 41 | But Paul's example has no cell to cite. No cell contains "average June |
| 42 | revenue over three years" — it must be *computed* from three or more |
| 43 | cells that may sit in different tabs, different sheets, or different |
| 44 | files (one per fiscal year). So the rule either forbids the feature Paul |
| 45 | wants, or needs qualifying. |
| 46 | |
| 47 | ## What I want |
| 48 | |
| 49 | 1. **Rule on the tension.** Options: memoria computes and shows its |
| 50 | arithmetic; memoria asks Sheets to compute (a new sibling call — |
| 51 | "evaluate this expression against these cells" — and what that costs); |
| 52 | memoria returns the constituent cells and lets the person sum them; |
| 53 | memoria writes a formula into a scratch sheet and cites it; the rule |
| 54 | narrows to "never recomputes a value a sheet already computed". Weigh |
| 55 | and rule. Say which option genuinely preserves "a second calculator |
| 56 | that disagrees with the first" and which merely evades it. |
| 57 | |
| 58 | 2. **The sibling call, if you recommend one.** memoria holds no DEK for |
| 59 | Sheets and must not gain a write path (§6.4). Specify it in the style |
| 60 | of the spec's S1–S5 table, and say what it leaks that S1–S3 do not. |
| 61 | |
| 62 | 3. **Chunking a spreadsheet** — §13's open item. A 400-token chunk of a |
| 63 | grid may be meaningless. Design it. Consider per-tab; per-region (a |
| 64 | header row plus contiguous block); per-row with headers prepended; a |
| 65 | structural summary (tab names, header rows, ranges, units) separate |
| 66 | from value chunks. What makes "revenue", "June" and "2024" findable by |
| 67 | BM25 *and* by embedding, when the literal cell is a bare number whose |
| 68 | meaning comes from a row header two columns left and a column header |
| 69 | eleven rows up? |
| 70 | |
| 71 | 4. **Retrieval shape for an aggregate.** Paul's question needs three |
| 72 | facts, one per year, each possibly in a different document, and needs |
| 73 | to know it found all three rather than two. How does top-k retrieval |
| 74 | know its recall was complete? What does it say when unsure? This |
| 75 | codebase treats "a plausible empty list, not an error" as the worst |
| 76 | available failure shape — and a plausible *wrong average* from two of |
| 77 | three years is worse still. |
| 78 | |
| 79 | 5. **What must NOT be built.** Name the version of this feature that |
| 80 | would be a mistake, so the spec can forbid it explicitly. |
| 81 | |
| 82 | Rule, don't survey. Cite spec sections. Be concrete about failure modes. |
| 83 | Do not edit any file. 900-1200 words. |
| 84 | |
amadan