# miara — price the next job from your own evidenced deliveries > Price the next job from the last four. Agent leverage broke every estimate you had. miara rebuilds your unit of estimation out of the only honest material available — your own recent, evidenced, shipped work. Ranges only, never point estimates. Your history only — never pooled. - Canonical page: https://dirahq.sh/miara - Part of: letopis — https://dirahq.sh/letopis - Status: after dira and zavet in the suite roadmap ## The centrepiece: the refusal Below three confirmed comparables, miara returns **no number at all** — not a lower-confidence answer, not a wider band, nothing. That is the brand, not an error state. Example, a new client with a payments deliverable: ``` No range available n 2 confirmed comparables (need 3) dropped 2 at 750 · 2 unhallmarked missing no 999 payments work in history Quote this one on judgement. Deliver it with zavet in the repo, and the next payments job has its first 999 comparable. ``` fixed price: `€ —` · margin: `quoted on judgement` Every other estimation tool answers whatever you ask it — that's why nobody trusts the answer. With fewer than three confirmed comparables, miara returns no number at all, just the shortest path to being able to answer next time. - **Always with n and spread** — a range from three jobs is not a range from twelve, and it says so. - **A refusal is actionable** — it names the missing history, not an apology. Silence is a feature, not a gap. ## The problem Software estimation lost its unit. Leverage varies tenfold by work type, and gut feel calibrated on pre-agent velocity is dead. You are mispricing in both directions and can't tell which. - **Hourly — shrinking base.** Honest engaged hours fall as leverage rises. Accurate tracking alone makes you provably cheaper — the insurance-product trap. `2024 40h × €75 = €3,000` → `2026 12h × €75 = €900`. - **Fixed-price — guesswork.** The right model for the agent era, and nobody can size it. Leverage on a CRUD screen is nothing like leverage on a migration. Same €4k quote → 3h of work, or 34h. - **The replacement — your history.** The only honest unit left is what *you* recently shipped, with evidence attached — exactly and only what dira already holds. Last 4 similar → 2.1 – 4.8h engaged. ## Three objects You pick the comparables. Miara does the arithmetic. 1. **Deliverable — declared up front.** The quotable unit, named at quote time and bound to evidence by repo, branch and PR. Your invoice lines already group work after the fact; a deliverable is the same grouping, declared before. `"CSV import w/ column mapping"` · `acme/api · feat/csv-import`. 2. **Comparable — you confirm.** A past deliverable you confirm as similar. miara suggests candidates from anchored commit messages, diff shape, repo and client — and never auto-selects a single one. e.g. accepted: XLSX import (3.4h), Bulk contact upload (2.2h); rejected: Stripe webhooks. 3. **Evidenced range — n ≥ 3 or nothing.** Computed only from confirmed comparables, always shown with n and spread. Below three, miara refuses to speak, and tells you precisely which history is missing. Last 4 similar: 2.1 – 4.8h engaged · 19 – 31h agent · ~€6 compute. ## Sample quote — the widget A live quote for "CSV import with column mapping" (Northwind Ltd), gated by an evidence floor toggle (`any` / `750+` / `999`): | Evidence floor | engaged | agent | compute | n | spread | fixed price | margin | | --- | --- | --- | --- | --- | --- | --- | --- | | any | 1.6 – 7.4h | 14 – 46h | ~€9 | 6 confirmed comparables | 3.1× | €1,800 | 34 – 88% | | 750+ | 2.1 – 4.8h | 19 – 31h | ~€6 | 4 confirmed comparables | 2.3× | €1,800 | 62 – 81% | | 999 | — | — | — | 2 confirmed (need 3) | — | € — | quoted on judgement | Tightening the evidence floor narrows the comparable set and the spread — picking `999` on this quote drops below three confirmed comparables and returns the refusal card above instead of a number. ## The flow Quote, deliver, grade yourself. A quoting surface that sits next to your invoices, sharing the same clients — because the evidence it needs is already there. 1. **Describe** — name the deliverable and the client. One sentence. 2. **Confirm** — accept or reject each suggested comparable yourself. 3. **Range** — engaged, agent and compute, with n and spread attached. 4. **Price** — set a number and watch live margin against evidenced cost. 5. **Calibrate** — sessions attach automatically. Quoted vs actual, forever. ## Calibration — the moat miara grades itself, visibly. Every delivered quote is scored against what actually happened. Hit-rate over time, drift by client and by work type — self-grading estimation tools simply do not exist in this market. Ten calibrated quotes in, your pricing history is an asset that exists nowhere else. - **Hit-rate** — how often the actual landed inside the quoted range. - **Drift** — which clients and which work types you consistently misjudge. - **Why capture stays on** — switch dira off and your next quote gets worse. Last 10 delivered quotes: **80% inside the quoted range**. | Deliverable | Quoted band | Actual | Hit | | --- | --- | --- | --- | | CSV import | on band | 3.9h | yes | | Auth rework | on band | 6.2h | yes | | Schema migration | missed band | 11.4h | no | | Admin table UI | on band | 2.8h | yes | | Webhook retries | on band | 7.1h | yes | drift: migrations under-quoted 2 of 2. Below ~70% after ten quotes, the model is wrong — and we'd rather you knew. ## What miara never does No product in the suite ever states an opinion — only records that can be checked. Three refusals, permanently: - **No point estimates** — a single number implies a confidence nobody has. Ranges only, with n and spread. - **No estimate below three comparables** — not a lower-confidence answer. No answer. - **No cross-tenant pooling** — your leverage isn't their leverage. Pooling is both a privacy problem and a lie. ## Interlock with proba Where hallmarks exist, fineness becomes a comparability filter: two engaged hours at 585 with inferred authorship is not the same commodity as two hours at 750 under heavy supervision. Optional, purely additive — miara works without it. Example: comparables filtered to 750+ takes n from 6 to 4 and spread from 3.1× to 2.3×. See also [/proba.md](https://dirahq.sh/proba.md). ## The European AI Act miara has no compliance obligation of its own — it prices work. But governed work costs more hours than ungoverned work, and somebody is currently absorbing that difference as lost margin. Two real connections, and no more than two: - **Fineness floor** — filter comparables to 750+ and you are quoting only from deliveries whose oversight was actually evidenced. Fewer comparables, tighter spread — and the price of properly governed work stops being a guess. - **Governance is billable** — the engaged hours that review, decision records and supervision add are in the range, itemised, next to the agent hours. A regulated client sees what the diligence costs instead of you quietly eating it. And what miara will not do is price a compliance programme it has no comparables for. Below three confirmed similar deliveries it returns no number at all — which, for a category this new, will be the honest answer for a while yet. Full article-by-article mapping: [/letopis-ai-act.md](https://dirahq.sh/letopis-ai-act.md). ## Where it lives Inside dira cloud, next to the invoices. Not a separate product to buy, learn, and reconcile — the same clients, the same evidence, one tab over. - **Personal · free** — local tracking and offline reports stay free forever. No quoting, no calibration. - **Workspace — miara lives here.** Quoting, evidenced ranges and calibration — the honest answer to "why would I upgrade from free." - **Already accruing** — invoice lines record their deliverable grouping today, so miara ships with calibrated history, not a cold start. --- Every hour dira captures today is a comparable miara prices from tomorrow. See also [/index.md](https://dirahq.sh/index.md), [/zavet.md](https://dirahq.sh/zavet.md), [/proba.md](https://dirahq.sh/proba.md).