diff --git a/docs/env_reference.md b/docs/env_reference.md
new file mode 100644
index 0000000..f1b5fb0
--- /dev/null
+++ b/docs/env_reference.md
@@ -0,0 +1,196 @@
+# Environment variable reference
+
+Every knob the bot reads. **All are read ONCE at process start.**
+
+> **The bot is spawned by the server (or by the GUI if it starts its own server).**
+> So the variable must be set on **that** process, not in an unrelated terminal.
+> Export it in the shell that launches the server/GUI.
+
+An unset (or empty, or unparseable) value means **the shipped default** — a typo
+can never silently change behaviour, it warns on stderr and falls back.
+
+---
+
+## The ones you'll actually use live
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_MOVEMENT` | `tfil` | movement engine: `tfil` (shipped) or `tfil_ring` (range-weighted variant) |
+| `TR_RACK_<GUN>` | see below | move a gun in/out of the rack per mode |
+| `TR_POWER_POLICY` | `1` | `0` = uncapped control arm (today's behaviour without the energy policy) |
+| `TR_POWER_LOG` | off | `1` = log each power decision with its reason |
+| `TR_RAM_LOG` | off | `1` = log ram on/off with the reason |
+| `TR_MOVEMENT_LOG` | off | `1` = log movement band/class changes |
+| `TR_TMHORIZON_LOG` | off | `1` = let the horizon TM gun log its thinking per shot |
+
+---
+
+## 1. Gun rack — which gun(s) may be chosen
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_RACK_<GUN>` | `PATTERN=both`, all others `off` | membership: `both` \| `1v1` \| `melee` \| `off` |
+| `GUN_RACK_DISABLE` | empty | comma-separated gun **ids** to remove entirely (older mechanism; the gun never even spawns virtual bullets) |
+| `GUN_STATS_PATH` | `/tmp/gun_stats.jsonl` | where the per-round per-gun stats are written |
+
+`<GUN>` names: `HEADON LINEAR TSETLIN CIRCULAR GUESSFACTOR PATTERN WALLBOUNCE ACCEL
+STOPSHOT DISPLACE AVGLEAD DECAYGF KNN TMSELECT TMPATTERN TMHORIZON`.
+
+**The shipped default is Pattern only.** To restore the full old rack:
+
+```sh
+TR_RACK_PATTERN=both TR_RACK_HEADON=both TR_RACK_LINEAR=both TR_RACK_TSETLIN=both \
+TR_RACK_CIRCULAR=both TR_RACK_GUESSFACTOR=both TR_RACK_WALLBOUNCE=both \
+TR_RACK_ACCEL=both TR_RACK_STOPSHOT=both TR_RACK_DISPLACE=both TR_RACK_AVGLEAD=both \
+TR_RACK_DECAYGF=both TR_RACK_KNN=both TR_RACK_TMSELECT=both ./out/ModularBot
+```
+
+To force one gun alone: `TR_RACK_PATTERN=off TR_RACK_TMHORIZON=both`.
+If **every** gun is `off`, the rack falls back to admitting all of them.
+
+## 2. The selector (virtual-bullet fitness)
+
+| variable | default | what it does |
+|---|---|---|
+| `GUN_VBULLET_METRIC` | `path` | how a virtual bullet is scored: `path` (swept ray — the better coarse signal, 7.4% vs 4.7% real) or `point` (arrival accuracy) |
+| `GUN_SELECTOR_MODE` | `relative` | tie-band model: `relative` (scale-aware) or `absolute` (legacy fixed bars) |
+| `GUN_SELECTOR_WINDOW` | `100` | rolling window (ticks) for the fitness rates |
+| `GUN_SELECTOR_TIE` | `0.20` | tie band width: tied if `rate >= bestRate*(1-this)` |
+| `GUN_SELECTOR_DWELL` | `10` | ticks the incumbent gun is held before it may be displaced |
+| `GUN_SELECTOR_MARGIN` | `0.05` | a challenger must beat the incumbent by this fraction to displace it |
+| `GUN_SELECTOR_TIEBREAK` | `off` | `off` (shipped) / `point` / `pointCommit`. Measured **negative** on real hit rate; opt-in only |
+| `GUN_SELECTOR_POINT_TIE` | `0.5` | margin used by the `point` tie-break |
+
+**Note:** the per-tick random draw inside the tie band is **load-bearing** —
+replacing it with commitment to the virtual-best measurably *lowered* real hit
+rate (7.02% → 5.10%, p=0.002). Three attempts to "smarten" the band all failed.
+
+## 3. Power policy (energy economy)
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_POWER_POLICY` | `1` | `0` = uncapped control arm. **Measured: turning it off drops real hit rate 10.6%→7.9%** |
+| `TR_POWER_FAR_DIST` | `200` px | beyond this = "bad chances zone" |
+| `TR_POWER_FAR_CAP` | `1.0` | power cap beyond that distance |
+| `TR_POWER_MID_CAP` | `2.0` | cap when close + healthy but chances are not above average |
+| `TR_POWER_REF` | `0.0` | reference probability: `0` = the gun's own mean, `>0` = a fixed threshold |
+| `TR_POWER_ENERGY_HI` | `80` | at/above this energy, no energy-slope cap |
+| `TR_POWER_ENERGY_LO` | `20` | at/below this energy, cap = `ENERGY_MIN` |
+| `TR_POWER_ENERGY_MIN` | `0.5` | the cap at/below `ENERGY_LO` |
+| `TR_POWER_ENERGY_MAX` | `3.0` | the cap at/above `ENERGY_HI` (3.0 = effectively uncapped) |
+| `TR_POWER_FINISH_KILL` | `1` | cap power at the **smallest bullet that can still kill** the enemy |
+| `TR_POWER_LOG` | off | `1` = log each decision: `[power] p=… cap=… reason=far\|lowEnergy\|belowAvg\|full\|ram` |
+
+WHY the slope: bullet speed is `20-3p`, so **lower power = faster bullet** (less
+lead error, higher hit chance), fires more often (`10+2p` ticks) and drains energy
+more slowly (`p`/shot). `E[dE] = p(3P-1)`, so break-even is **1/3 regardless of
+power** — and our measured hit rates are 5–27%, far below it.
+
+WHY the finishing cap: server 1.3.1 caps the damage **score** at the energy
+actually removed, so **overkill scores nothing**. Damage is `4p` (p≤1) / `6p-2`
+(p>1), so the smallest killing power is `E/4` for E≤4 and `(E+2)/6` for 4<E≤16.
+
+## 4. Movement
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_MOVEMENT` | `tfil` | `tfil` (shipped) or `tfil_ring` (the range-weighted variant) |
+| `TR_MOVEMENT_LOG` | off | `1` = log band/range-class changes for the ring mover |
+| `TR_TFIL_RANGE_LO` / `_HI` | `100` / `200` px | the target-range band the ring mover prefers |
+| `TR_TFIL_RANGE_TEMP` | `0.4` | peakiness of the range weighting. `0` = plain uniform draw = exact control |
+| `TR_TFIL_RANGE_K` | `60` | softness of the falloff outside the band |
+| `TR_TFIL_CORRIDOR_HEAT` | `10.0` | heat added along a bullet's corridor to the wall |
+| `TR_TFIL_WALL_HOTNESS` | `15.0` | wall radiance |
+
+**Ring mover is NOT the default and should not be shipped as-is:** it had the best
+offline hit rate of anything measured, but it halves engagement range and **halves
+survival** (round wins 16/49 → 6/49, p=0.012) — a glass cannon.
+
+## 5. Ramming
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_RAM_OPPORTUNITY` | off | enable the proactive ram. **Measured: converts 0/6 times** — a straight-line pursuit cannot catch an equal-speed enemy |
+| `TR_RAM_OPP_DIST` | `200` px | opportunity trigger distance |
+| `TR_RAM_OPP_MARGIN` | `15` | how much MORE energy we must have than the enemy |
+| `TR_RAM_ABORT_DMG` | `2.0` | abort an in-progress ram above this incoming **energy per turn** (the "bullet rain" abort) |
+| `TR_RAM_PLAN` | off | the speculative "change of plan" trigger |
+| `TR_RAM_PLAN_DIST` / `_MARGIN` / `_HITRATE` | `250` / `20` / `0.05` | its thresholds |
+| `TR_RAM_LOG` | off | `1` = log `[ram] ON/OFF` with reason |
+
+The **finisher** ram (enemy <20 energy, we are healthier) is always on and is the
+only path that converts.
+
+## 6. The horizon TM gun (`TMHORIZON`)
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_TMHORIZON_LOG` | off | `1` = per-shot thinking log + per-round summary |
+| `TR_TMHORIZON_SHIFT` | `2.0` deg | how far it may move the aim off Pattern's answer. **`0` = predict but never move the aim** (isolates prediction from application) |
+| `TR_TMHORIZON_BIG_MULT` | `1.5` | multiplier when the predicted correction is BIG |
+| `TR_TMHORIZON_NSTATES` | `64` | **automata inertia** — the "mood". Lower = adapts faster, higher = more stubborn (2…4096) |
+| `TR_TMHORIZON_RESET_ON_TARGET` | on | wipe the model when the target changes (no-op in 1v1) |
+| `TR_TMHORIZON_WINDOW` | `0` | sliding window: retrain on only the most recent N samples. `0` = keep everything |
+| `TR_TMHORIZON_RESET_DROP` | `0` | change detection: if rolling accuracy falls this many points below its own peak, treat it as "the enemy changed" and re-learn |
+| `TR_TMHORIZON_ACCURVE` | off | `1` = log the rolling accuracy periodically |
+| `TR_TMHORIZON_RETRAIN_EVERY` | `50` | samples between windowed retrains |
+| `TR_TMHORIZON_EPOCHS` | `1` | epochs per windowed retrain |
+
+Learning **persists across rounds** of the same battle (so it can overfit the
+current enemy) and wipes only on a new battle or a target change. **No learning is
+written to disk — nothing carries between battles.**
+
+## 7. Other gun knobs
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_PATTERN_RAD_OFFSET` | `0.0` px | shift Pattern's aim along its own bearing (negative = aim short) |
+| `TR_PATTERN_RAD_SCALE` | `1.0` | multiplier on Pattern's aim distance |
+
+Both are **structurally incapable of changing the shot** — the live aim is
+bearing-only, so a purely radial offset is invisible. Measured byte-identical on
+`bmPath`. Kept for experiments; leave at defaults.
+
+## 8. Measurement / instrumentation (all off by default, all zero-cost when off)
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_VBULLET_ADMIT_ONLY` | `1` | only admitted guns spawn virtual bullets. **Turning it on gave +68% tick rate** (87→146 ticks/s); `0` restores the old behaviour |
+| `TR_RECORD_WORLDSTATE` | off | append every tick's WorldState to `/tmp/worldstate_record.jsonl` for offline replay |
+| `TR_RADAR_FORCE_SPIN` | off | force the old stateless full-spin melee radar (for A/B on one binary) |
+| `TR_RADAR_SCANLOG` | off | append per-round radar scan/coverage JSON |
+| `TR_RADAR_SCAN_LOG_PATH` | `/tmp/radar_scan_log.jsonl` | where that goes |
+| `TR_TRACKER_PROBE` | off | per-tick enemy tracker vs server enemy count |
+| `TR_TRACKER_PROBE_PATH` | `/tmp/tracker_probe.jsonl` | where that goes |
+
+## 9. Test harness only (not read by the bot itself)
+
+| variable | default | what it does |
+|---|---|---|
+| `TR_SERVER_JAR` | the 1.3.1 jar in `~/Downloads/robocode-tankroyale/` | which server the tests launch. Set to the 0.35.5 jar to reproduce old numbers |
+| `TR_BATTLE_RUNNER` | the 1.0.2 runner jar | which runner the tests launch |
+| `TR_BATTLE_RUNNER_DIR` | derived | directory form of the runner |
+
+## 10. Compile-time knobs (`-d:` flags, need a rebuild)
+
+| flag | default | file |
+|---|---|---|
+| `-d:TMH_NSTATES=N` | 64 | `guns/tm_horizon.nim` — automata inertia (the runtime env overrides it) |
+| `-d:TMH_NCLAUSES=N` | 40 | `guns/tm_horizon.nim` |
+| `-d:TMH_MIN_OBS=N` | 24 | cold gate: below this many samples the TM emits no correction |
+| `-d:TMH_STALE_MAX=N` | 8 | a sample is dropped if the enemy was not seen this recently |
+| `-d:TM_S_DEF=x` | `"3.0"` | `guns/tm_pattern.nim` specificity (`s`). **Higher = LONGER clauses**, measured. `s=1.0` is degenerate |
+| `-d:TM_NSTATES=N` | 64 | `guns/tm_pattern.nim` |
+
+## Adding / finding knobs
+
+```sh
+# everything the code reads
+grep -rn 'TR_[A-Z_]*"' --include=*.nim ModularBot_garage/src common_libs | sort -u
+# what a specific knob defaults to
+grep -rn 'TR_POWER_ENERGY_MIN' --include=*.nim .
+```
+
+Convention followed throughout: read once at module init, into a `let` with an
+`EnvVar` name constant beside it, so one compiled binary can A/B every arm by
+environment alone.