Failed on the agent's machine.

• highPlanner truncates the affordance list at 24 in enumeration order, so a creature holding several item kinds can never choose rest, explore, go home, inspect or any social actionserver/src/expanse/agent.js:63

  affordances() in engine.js appends EAT, gather, drink, combine, place, use@pile/fire, carry and build entries first and only then rest, move-to-shelter, move|explore, move|home, inspect and the social affordances (engine.js:224-231). plan() keeps only the first 24 candidates for step one, and only the first LIMITS.branch=7 for later steps (line 71). The combine block alone emits C(k,2) entries for k held kinds.

  With six held kinds and six known node kinds the first 24 entries are 1 eat + 6 gather + 1 drink + 15 combine + 1 place, so the planner never sees rest/explore/home/inspect/share/follow/attack/flee. The creature cannot rest (energy keeps draining; see energy finding), cannot return home, cannot explore, and all social behaviour is disabled exactly when it is holding the materials the discovery experiments are about.

  The 7-wide branch limit also means combine steps are never reachable at plan depth 2 or 3 whenever more than 7 eat/gather/drink entries precede them, so multi-step craft plans are structurally impossible in common inventories. Which actions are considered is decided by list order, a designed bias that the documentation does not mention.

  whole finding

  const e = Expanse.create({initialCount:24},{seed:7}); const c = e.creatures[0]; c.inventory = {wood:1,stick:1,fiber:1,stone:1,flint:1,berry:1}; for 10 kinds add c.mind.known["k"+i] = {x:c.x+2+i,y:c.y,kind,units:3,t:0}; c.mind.water.push({x:c.x+3,y:c.y+3,t:0}); const af = e.affordances(c, e.perceive(c))(c.inventory, null); af.length === 34; af.slice(0,24) contains no step with prim rest, no move purpose explore/home, no inspect, no share/follow/socialize.

  Expected: rest/explore/home and social options are always among the candidates compared by the planner (or selection is not order-dependent).

  Actual: they are silently discarded, and plan() in agent.js never evaluates them.

  whole reproduction

• highEnergy is a cost but never a constraint: creatures at zero or negative energy can still gather, build, craft and attack, and energy goes unboundedly negativeserver/src/expanse/engine.js:429

  The requirement is primitives with inventory, reach, energy, material and time constraints. No primitive checks energy before executing: build (line 428-431), gather (line 392), combine (line 402), attack (line 519) and threaten (line 461) subtract energy unconditionally. needs() only clamps the per-tick drain (line 624), not the action costs, so energy reaches large negative values from which rest (max +1 per tick, 20 ticks per rest step) cannot recover.

  A creature with energy -55 has exactly the same action set and speeds as one with energy 100, so the energy state that the outcome model and planner are supposed to trade off against reward is not a real consequence. This also makes the "exhaustion" death pathway (health -0.03 per tick while energy <= 0) a permanent drain that is recorded with no cause (see separate finding).

  const e = Expanse.create({initialCount:24},{seed:7}); const c = e.creatures[0]; c.energy = 0; c.inventory = {wood:4,stone:2}; c.plan = {id:1,tick:e.tick,source:"t",steps:[],predicted:[],actual:[],raw:[{prim:"build",key:"build|shelter|",consumes:[],label:"BUILD"}],idx:0}; e.startStep(c); repeat e.stepAct(c) while c.act.

  Expected: build is refused (or blocked until rested) for a creature with no energy.

  Actual: e.shelters.length === 1, c.inventory === {} and c.energy === -6.

  Same with attack: attacker energy -50, target adjacent, raw step {prim:"attack",target:id,key:"attack|food|bare",label:"APPROACH"}; e.stepAct(attacker) hits (target health 100 -> 94) and attacker energy becomes -55.

  whole reproduction

• mediumHidden world facts leak into decisions: every creature knows whether any shelter exists anywhere and can walk to shelters up to 30 tiles away that it never perceivedserver/src/expanse/engine.js:189

  The requirement says creatures have local perception and individually acquired map knowledge, not omniscient coordinates, and planning cannot query hidden world facts. features() feeds the mb-v1 trace a global flag derived from this.shelters.length (the whole world), and affordances() (line 226-227) offers a move|to-shelter step to nearestShelter(c) (line 250), which scans all shelters within Chebyshev distance 30 while perception radius is 7 (5 at night).

  Neither the shelter list nor its coordinates pass through perceive() or mind.known. A creature that has never seen a shelter built 25 tiles away is offered its exact coordinates as an action target.

  const e = Expanse.create({initialCount:24},{seed:7}); const c = e.creatures[0]; e.shelters.push({id:99,x:c.x+25,y:c.y,cap:4,integrity:100}); const obs = e.perceive(c); obs.radius === 7 and obs contains no shelter; e.features(c,obs)[5] === 0 (shelter-known flag set); e.affordances(c,obs)(c.inventory,null).find(s=>s.purpose==="shelter") === {prim:"move",x:c.x+25,y:c.y,purpose:"shelter",key:"move|to-shelter|",label:"REST"}. Expected: no shelter affordance and the feature unchanged unless the shelter was perceived or remembered by this creature.

  whole reproduction

• mediumShare decision reads the target creature's hidden hunger, which is not part of the observation interfaceserver/src/expanse/engine.js:240

  The observation (perceive(), line 180) exposes only id, x, y, act, health and holding for other creatures. socialAffordances() then looks up the live creature object and gates the share affordance on other.hunger, an internal need the actor cannot observe. This is the kind of hidden fact the agent interfaces are supposed to exclude, and it makes the "shares" statistic in the experiments a product of designed omniscience rather than learned social behaviour.

  const e = Expanse.create({initialCount:24},{seed:7}); const [a,b] = e.creatures; b.x = a.x+1; b.y = a.y; a.inventory = {berry:2}; e.rebuildIndex(); b.hunger = 59: e.affordances(a,e.perceive(a))(a.inventory,null).some(s=>s.prim==="share") === false; b.hunger = 61: same call === true, while e.perceive(a).creatures[0] has keys [id,x,y,act,health,holding] only. Expected: identical affordances for identical observations.

• mediumDeaths caused by energy exhaustion are recorded as cause "unknown"server/src/expanse/engine.js:624

  needs() applies a health drain while energy <= 0 but never sets cause, unlike starvation, dehydration, exposure, injury, burns and old_age. A creature whose health reaches zero through this drain gets deathCause "unknown", so the deathsByCause totals, graveyard and the experiment reports (RESULTS/summary report "unknown death causes") cannot attribute these deaths.

  Because energy is never gated (previous finding) this path is reachable in ordinary play; seed 3 at 4000 ticks already produces one such death.

  const e = Expanse.create({initialCount:24},{seed:7}); const c = e.creatures[0]; c.energy = 0; c.health = 0.02; c.hunger = 65; c.thirst = 10; c.warmth = 90; c.injury = 0; c.act = null; e.needs(c); Expected: c.health < 0 and c.deathCause === "exhaustion" (or similar).

  Actual: c.health === -0.01 and c.deathCause === "unknown".

  Also: Expanse.create({initialCount:24},{seed:3}).run(4000).totals.deathsByCause contains {"unknown":1}.

• mediumFollowers move one tile per tick regardless of terrain cost, crossing shallows and snow three times faster than any walking creatureserver/src/expanse/engine.js:612

  walk() charges MOVE_COST ticks per tile (SHALLOW 3, SNOW 2.5, ROCK 2, STREAM 2) and energy proportional to the cost and carried weight. followStep() bypasses walk(): it steps one tile every tick for a flat 0.08 energy, ignoring terrain cost and inventory weight, and does not use moveProgress.

  A follower therefore traverses wadeable water and highland at up to 3x the physical speed, cheaper than the leader it follows, which contradicts "traversal constraints must have real consequences" and "no free resources".

  Find a horizontal run of 5 SHALLOW tiles (seed 7 has one in the coast biome).

  Put follower f at its left end and leader l four tiles right with l.act set to rest; f.following = {id:l.id, until:1e9}; f.hunger=f.thirst=10; call e.followStep(f) twice.

  Expected: at most 0 tiles advanced (3 ticks per SHALLOW tile).

  Actual: f.x advanced by 2 tiles in 2 ticks and f.energy dropped by 0.16 total, versus 0.24 per tile for a walker on cost-3 terrain.

• lowRead-only GET /api/expanse/info mutates creature state via leadership(), so viewers change persisted world bytesserver/src/expanse/engine.js:693

  server.js:27 calls world.leadership() from the GET /api/expanse/info handler and view.js documents that views never mutate the world. leadership() deletes expired follower entries from every creature.

  Two servers with identical state diverge in their snapshots (and sha256 manifests) depending on whether a viewer requested /info, which breaks the "server runs identically without viewers" property used for backup comparison and makes the followers field of persisted creatures viewer-dependent. Simulation outcomes are not affected because expired entries are filtered when read.

  const e = Expanse.create({initialCount:24},{seed:7}); const c = e.creatures[0]; c.followers = {5:0}; e.tick = 5000; JSON.stringify(c.followers) === "{"5":0}"; e.leadership(); JSON.stringify(c.followers) === "{}". Expected: a read endpoint leaves toState() unchanged.

• infoShelter construction and new settlements never occur in any of the 60 experiment runs; eight-client, Docker and ARM64 checks were not performedtests/results/expanse/summary.json:21

  Every condition in summary.json reports shelters 0 (lines 21, 48, 75 and all others), and births 0 in the main condition, so the required capability "shelter construction and new settlements must be possible" and natural births in the expanded world are not demonstrated by the delivered evidence; the mechanism exists in code only. artifacts/phase2b/REVIEW.md line 22 states that Docker execution, ARM64/cgroup capacity and eight real networked rendering clients were not tested, so the "8 clients" benchmark requirement is unmet (the bench serialises eight frames per tick in-process).

  These are reported gaps, not hidden ones, listed here so the owner weighs them before deployment.

  Open tests/results/expanse/summary.json: every "shelters" value is 0 and main/full "births" is 0.

  Open artifacts/phase2b/REVIEW.md line 22.

  Expected per task: at least one run showing shelter construction / a new settlement, and a benchmark with 8 connected clients.

Failed on the network's side.

Failed on the network's side.

Failed on the network's side.