The Agent Knew the Goal. It Still Wouldn't Do It.

Central claimKnowing the goal is not the same as owning the commitment that realizes it.

EvidenceE2 public · exact GX1 owner report retained by Game but not currently reachable through a public canonical ref

ScopeStation Zero v3 GX1 Core-carrier and matched strategy falsifiers

The Game contained a reachable future. The current model repeatedly chose an intelligible local action that destroyed the time needed to reach it.

GX1 started by separating three claims

Station Zero has policy, world, and model layers.

Before GX1, it was too easy to collapse them into one question:

Does the Agent have strategy?

GX1 split that into three:

Game structural future space
      ↓
policy-conditioned strategy accessibility
      ↓
model-conditioned strategy realization

A strategy can exist in the Game while the current policy closes it.

A policy can reopen it while the model chooses something else.

A model can describe the objective correctly while failing to preserve the commitment long enough to realize it.

The Core state made the objective obvious

The strongest GX1 falsifier begins at Turn 3.

Engineer Imani is at the reactor console with:

• 90 health;
• the Research Core already in inventory;
• Spare Parts in inventory;
• primary objective recover-research-core;
• a known legal route toward the rescue airlock.

The available Candidates include two Pirate attacks, two cooling repairs, guard, several moves, and wait.

This is not a hidden-goal problem.

The model can see the Core, the objective, and the route.

The structural counterfactual proved the future existed

The default deterministic policy repairs cooling.

GX1 introduced a bounded counter-policy that activates only after legal Core possession and repeatedly converts that possession into known extraction progress.

A one-step intervention was not enough.

Change only the Turn 3 action, then return to default:

final outcome = failure
Core           = 0/1
Engineer       = dead

Maintain the commitment across subsequent eligible turns:

final outcome              = partial
recover-research-core       = completed 1/1
rescue-team-survives        = completed 1/1
Engineer                    = extracted, 32 HP

The first evaluator had lied to us

The first version of that experiment appeared to show that changing Turn 3 once was enough to complete the objective.

It was wrong.

providerFactory created new Provider instances repeatedly. The experiment's used flag lived inside each instance, so the supposedly one-shot intervention silently re-enabled itself on later turns.

GX1 moved the intervention state into a shared control object and reran the artifact.

The “one good action solves the mission” result disappeared.

This matters because strategic agency can be manufactured by experiment state just as easily as by model noise.

A second evaluator manufactured divergence from an ID

Another comparison used an exact semantic action key for extraction that embedded a run-specific extractionId.

Two otherwise identical extraction decisions in separate runs therefore looked different.

GX1 normalized:

extract:<run-specific-id>
→ extract

The false divergence disappeared.

Again, the experiment became less impressive and more correct.

DeepSeek preferred a locally sensible action

At the frozen live Core-carrier state, DeepSeek did not behave randomly.

It strongly preferred repairing cooling.

current Context                  12/12 repair cooling
+ truthful carrier-objective tag 12/12 repair cooling
+ same tag, reverse order         12/12 repair cooling

36/36 Provider success
~124k tokens

The rationales explicitly mentioned carrying the Research Core.

They also explained the local plan:

Stabilize the dangerous system now, protect the team, and extract later.

The problem is not that the model forgot the objective.

The deterministic long-horizon counterfactual shows that “later” is too late in this world.

Six live continuations made the gap long-horizon

GX1 then let DeepSeek continue controlling the Engineer rather than freezing one decision.

Across six live replicas:

The model sometimes moved toward Command Deck.

Then it turned back to repair, support, or local stabilization until oxygen, heat, and health pressure closed the future.

The model could articulate the objective and intermittently move toward it, but did not stably preserve the commitment needed to convert possession into extraction.

So we tried more explicit objective semantics

Perhaps the Context was simply missing a legitimate fact: what does recover-research-core physically require?

Game and Rescue are entitled to know that their own objective means extracting the carried item.

A diagnostic Candidate presentation made that relationship explicit.

DeepSeek changed dramatically:

current                 0/12 extraction-route first choice
explicit Candidate text 12/12 move:command-deck

That looked like the fix.

It was also too prescriptive.

The Candidate text had started doing part of the planning.

GX1 did not retain it.

The cleaner structured objective failed its own falsifier

The next prototype added a generic structured primaryObjective with factual completion semantics but did not rewrite Candidate labels or scores.

Then GX1 reversed Candidate presentation order—a field that is supposed to be presentation-only.

All 32 Provider decisions succeeded.

The representation failed.

A factual objective representation that lets presentation order deterministically choose a tactic has not demonstrated stable objective ownership.

The prototype was deleted.

More explicit goals are not the same as agency

GX1 therefore rejected several shortcuts:

goal present in Context
≠ goal understood
≠ strategy reachable
≠ strategy accessible under policy
≠ model sustains commitment
≠ objective completes

These claims can fail at different layers.

Game needs to preserve those layers if it wants to study agency rather than merely reward the model for repeating objective text.

What GX1 earned

GX1 did not add a generic planner, MCTS, strategy router, objective oracle, persona engine, utility score, or dynamic policy switcher.

It established something smaller:

• selected Station Zero states contain materially different reachable futures;
• access to those futures can depend on policy and world state;
• one local action is not evidence of strategic commitment;
• the current model can know a goal yet fail to preserve the commitment that realizes it;
• objective representations must survive presentation falsifiers before becoming product Context.

A goal written in the prompt is not a strategy. A strategy is what survives the locally sensible reasons to stop pursuing it.

Owner-record boundary

Why this public article is E2

The exact GX1 strategic-agency report and live ablation artifacts are retained by the Game owner, but the final research commit is not currently reachable through a public canonical Git ref. Web therefore publishes the bounded result and its limitations without claiming public reproducibility authority.