diff --git a/AICONTROL/docs/results-file.md b/AICONTROL/docs/results-file.md
new file mode 100644
index 0000000..0883813
--- /dev/null
+++ b/AICONTROL/docs/results-file.md
@@ -0,0 +1,131 @@
+# The AI → Business results file — draft v0.1
+
+| | |
+|---|---|
+| **What this is** | The third handover in the Project Manual: "one file with a fixed set of columns: one row per controller, per scenario, per run. Business builds the euro comparison straight from it. Agree the columns before the first benchmark is run" |
+| **Owner** | RL Evaluation seat (person 3). Reviewed by the lead 11 Sep 2026 from person 3's spreadsheet draft |
+| **Next** | Encode as `aicontrol/evaluate/results_schema.py` (week 4), show to Business's Commercial Model Analyst, agree the open questions at the bottom, then freeze |
+| **Format** | Parquet is the version of record; a CSV is written beside it so Business can double-click it. Rows are appended, never edited |
+
+## Rules that make the file usable by a script
+
+- Column names are lowercase `snake_case`, no spaces, no hyphens, and carry the unit where one
+  exists (`_kwh`, `_kg`, `_kw`). Fractions are stored as 0–1, never as percentages.
+- One row is one run: one controller, on one scenario, with one seed. A controller is a *value*
+  in the `controller` column, not a row label, so the file grows to hundreds of rows.
+- Controller names are fixed strings: `rule_based`, `mpc`, `perfect_knowledge`, `agent`, and
+  the Q1 stand-ins `random` and `do_nothing`. Business's S0 baseline (the hospital without the
+  hydrogen system) is not a controller but a plant configuration, so it is a *scenario* from
+  Simulations' library, appearing here as `scenario = no_hydrogen_<year>` with
+  `controller = rule_based`.
+- A column may be empty for a run (for example, `wandb_run` for a rule-based run). A column
+  that is empty for every run at the freeze is dropped.
+
+## Columns
+
+### Identity — which run this row is
+
+| column | type | meaning |
+|---|---|---|
+| `controller` | string | one of the fixed names above |
+| `controller_version` | string | git commit for rule-based / MPC; model checkpoint tag for the agent |
+| `scenario` | string | name of the scenario YAML file, e.g. `baseline_2018`, `blackout_72h`, `dark_december_week` |
+| `data_year` | int | which year of data the run used |
+| `held_out` | bool | true if the agent never saw this year in training; the manual's proof of generalisation |
+| `run` | int | repetition index within a setup (1, 2, 3, …) |
+| `seed` | int | the random seed of that repetition |
+| `steps` | int | number of steps simulated |
+| `step_minutes` | int | 5 |
+| `twin_version` | string | version of the Simulations package that produced the physics |
+| `parameter_book_version` | string | version of Energy Management's Parameter Book |
+| `dataset_version` | string | version of the clean hospital dataset |
+| `run_at` | timestamp | when the run finished (UTC) |
+| `wandb_run` | string | Weights & Biases run id, if any |
+
+The three `*_version` columns are the manual's rule that "every result records which versions
+produced it".
+
+### KPIs — the manual's four, plus money and carbon
+
+Names are fixed now; the exact equations come from Energy Management's Dispatch & Grid
+Engineer and are applied by Simulations' KPI calculator. We copy the numbers; we do not
+recompute them.
+
+| column | type | meaning |
+|---|---|---|
+| `self_sufficiency_rate` | fraction | share of load served without the grid |
+| `grid_dependency_ratio` | fraction | share of load that came from the grid |
+| `critical_load_uptime` | fraction | share of steps with Tier 1 fully served; must be 1.0 |
+| `soc_violation_rate` | fraction | share of steps with the battery outside its window |
+| `energy_cost` | float | net cost of grid electricity over the run (import cost minus export revenue) |
+| `currency` | string | `USD` or `EUR`; see open questions |
+| `co2_kg` | float | grid electricity used × carbon intensity, summed |
+
+### Energy totals — the evidence behind the KPIs, and what Business's register needs
+
+All summed over the run.
+
+| column | type | meaning |
+|---|---|---|
+| `load_kwh` | float | total hospital demand |
+| `load_served_kwh` | float | demand actually supplied |
+| `tier1_unserved_kwh` | float | **must be 0** in every run that counts |
+| `tier2_shed_kwh` | float | essential load cut |
+| `tier3_shed_kwh` | float | non-critical load cut |
+| `pv_kwh` | float | solar available |
+| `pv_curtailed_kwh` | float | solar thrown away |
+| `grid_import_kwh` | float | |
+| `grid_export_kwh` | float | |
+| `grid_import_peak_kwh` | float | import during tariff peak hours |
+| `grid_import_offpeak_kwh` | float | import during off-peak hours |
+| `peak_import_kw` | float | highest grid import in any step; drives demand charges |
+| `battery_charge_kwh` | float | |
+| `battery_discharge_kwh` | float | |
+| `battery_full_cycles` | float | throughput ÷ capacity; the wear proxy |
+| `electrolyser_kwh` | float | electricity into the electrolyser |
+| `electrolyser_offpeak_kwh` | float | the part of it in off-peak hours |
+| `h2_produced_kg` | float | |
+| `h2_consumed_kg` | float | |
+| `fuel_cell_kwh` | float | electricity out of the fuel cell |
+| `fuel_cell_heat_kwh` | float | recovered heat, if the twin models it; else empty |
+| `chp_kwh` | float | electricity from the CHP plant |
+| `electrolyser_starts` | int | on/off switches |
+| `fuel_cell_starts` | int | on/off switches |
+| `limit_violations` | int | steps where the twin clipped an action against a limit |
+| `notes` | string | free text |
+
+Why the peak / off-peak and hydrogen columns: Business's cost register has seven rows marked
+"ModelOutput / Missing" — off-peak share of hydrogen production, MWh shifted from peak to
+off-peak, peak shaving, effective cost of electrolysis, cost per kWh from hydrogen, gas
+displaced by fuel-cell heat, and the CO₂ change — and each is a ratio or difference of the
+columns above between the no-hydrogen scenario, the `rule_based` rows and the `agent` rows.
+The peak / off-peak definition is the tariff's, which Business and Energy Management own.
+
+## The companion file
+
+Business's register also asks for "simulation of net grid imports" as a series. That does not
+fit one row per run. Each run therefore also writes its per-step table (the same quantities,
+one row per 5-minute step) as Parquet under `results/runs/<run_id>.parquet`. The summary file
+above is the one they open; the per-step file is there when they need a chart.
+
+## Changes from person 3's spreadsheet draft
+
+- Added the `controller` header (the draft's column A had none) and made controller a column
+  value rather than a row label.
+- Fixed `KPI_Grid_Dependecy_Ratio` → `grid_dependency_ratio`; dropped the `KPI_` prefix,
+  hyphens and capitals from every name so the file loads without renaming.
+- Replaced `cost_eur` with `energy_cost` + `currency`, because the data source is ISO New
+  England, which prices in dollars, while the business case is in euros.
+- Added identity columns (`held_out`, `data_year`, the three versions, `run_at`), the energy
+  totals and the switching counts.
+- Kept `scenario`, `run`, `seed`, the four KPIs and `co2_kg` as drafted.
+
+## Open questions for Business (person 3 brings these)
+
+1. Currency: do they want dollars as recorded, or euros converted at a fixed rate they choose?
+2. Peak / off-peak: their tariff hours, so the split columns match their model.
+3. Are the seven register quantities computed by us (extra columns) or by them from these
+   columns? Either works; theirs keeps one owner per number.
+4. What exactly their S0 baseline contains (no hydrogen only, or no battery either), so we can
+   ask Simulations for that scenario file.
+5. Anything they need per run that is not here.