Golden Bench / story

Zero to sixty-five in a day

The first language-model runs scored nothing. A day later the same two models, with the same prompts and the same budget of decisions, averaged about 65 across the season's levels. Nothing about the models changed. What changed was what the observation says: the harness learned to tell the agent the things a player sees at a glance.

1. 2
   18 Oct, evening

   First baselines

   What the agent could now see

   • The player's position, facing and health.
   • The one objective that counts at this difficulty, in one line.
   • The nearest objects and characters, with ids.
   • A short list of actions: walk to a point, use, attack, wait.

   What happened

   Both models scored 0 on the first level with 60 decisions. One probed east past a locked gate instead of shooting its lock; the other got through and roamed about 4 km of road without finding the way on. Walking actions gave up well short on long legs. The bottleneck was the observation, not the model.

   first level only 0 and 0

2. 2b
   19 Oct

   Ways out, longer legs, a brief

   What the agent could now see

   • The ways out of the current area: which door, where it leads, open or locked.
   • A walk's time limit scaled to its length; a point with no floor route answers "unreachable" at once, costing no game time.
   • A one-line level brief in our words: where the mission starts, roughly how far and in which direction it ends.

   What happened

   Both models reached the first gate within 20 decisions and then spent the rest there. Still 0: the gate had no handle.

   first level only 0 and 0

3. 2c
   19 Oct, evening

   Gates with no handle

   What the agent could now see

   • A door with no handle names the panel that opens it, and panels are listed.
   • Vehicles appear as obstacles.
   • Walking up to a panel presses it once; a walk that stalls where it stalled before says so.

   What happened

   One model died fighting in a tunnel; the other hopped between doorway centres inside a single stretch for 150 decisions, "arriving" 106 times. Still 0.

   first level only 0 and 0

4. 2d
   19 Oct, late

   Moving between areas

   What the agent could now see

   • An action that walks through a chosen way out and stops just inside the next area.
   • Map memory a player would have: ways out marked when already visited, and a count of areas seen.
   • Object ids that stay the same for the whole episode.

   What happened

   Both models passed the first two gates and an airlock and reached the inner road, then spent about a hundred decisions trying to "enter" an opening that was a drop to the level below.

   first level only 0 and 0

5. 2e
   19 Oct, night

   Locks and storeys

   What the agent could now see

   • Padlocks are listed, and a locked door near one says "shoot it".
   • Ways out carry the height of the far side ("down 190", "up 300"); reaching a doorway without entering is reported as such, not as arrival.

   What happened

   The first completion: one model finished the first level in 64 decisions (score 82). The other passed both gates and died fighting on the inner road. The models now scored above 0 and ordered, so the whole season was run at one seed.

   season 1, seed 0 41 and 25.9

6. 2f
   20 Oct

   Indoors: keys, inventory, one-way doors

   What the agent could now see

   • What the player carries and holds, with an action to draw an item and use it on something.
   • Doors say "locked from this side, needs key K", and where that key is: held, carried by a character, or not seen.
   • Windows are no longer offered as ways out; objective targets and the level exit are marked; runs of identical objects collapse into one line.

   What happened

   The bunker level was completed by both models (100 and 97.9). The chemical plant stayed at 0: each model knew the plan but stood a storey below the person holding the key, pressing doors above it, because the text gave horizontal distances only.

7. 2g
   20 Oct

   Height in the text

   What the agent could now see

   • Distances to objects carry a storey offset.
   • Walk targets carry the object's height; a walk that ends above or below answers "wrong storey".

   What happened

   One model got the key at decision 112, then lost thirty decisions to a door that shut again while the walker backed off and re-pressed, and reached the console room with three decisions left. The 150-decision cap was binding on a four-objective level.

8. 2h
   20 Oct

   A better walker and per-level budgets

   What the agent could now see

   • Nothing new in the text. The walker underneath (ours) routes around walls, presses once, and backs off properly.
   • Each level gets a decision budget scaled to its anchor time (150 to 300).

   What happened

   Season 1 at one seed: 65.5 and 64.6, up from 41.0 and 25.9 in the morning. Three of five levels were being completed by both models.

   season 1, seed 0 65.5 and 64.6

9. 2i
   20 Oct

   Allies, hazards, places

   What the agent could now see

   • Friendly characters are marked and never targeted by the harness's return fire.
   • A hazard line when poison gas is active, with a rough time left; a standing rule about what starts it.
   • Objectives completed by entering a room name the area and how far it is; a panel press reports how long its door stays open against the route to it.

   What happened

   The chemical plant reached 3 of 4 objectives (score 45). Every season-1 level had now been completed by at least one model run except that one. Single runs were noisy, so the next measurement was the three-seed sweep on the leaderboard.