01 · the motivating failure
This is not just hallucination
Suppose a forecasting market asks whether Donald Trump will say a particular phrase during a “factory speech.” A shallow model treats the label as a complete description: factory speeches resemble other factory speeches, so rally language receives almost no probability. But the subject matters. Trump often turns nominal policy appearances into rally-style performances. Before assigning a confident forecast, the model should inspect the speaker, host, audience, format, comparable appearances, and the exact event definition.
The failure is not necessarily a fabricated fact. The forecast can be fluent, internally consistent, and still epistemically bad. The model has compressed an underspecified label into an overconfident situation model. It does not represent which parts of that model came from the prompt, which came from a generic prior, which can be checked, or which remain unresolved.
retrieve a generic event prototype
↓issue a precise forecast
mark speaker and format as unresolved
↓inspect context, then forecast
Our working term for the missing capability is epistemic control: maintaining enough of an information ledger to choose an action that is licensed by the current evidence. It overlaps with calibration, metacognition, selective prediction, tool use, and planning, but is not identical to any of them.
02 · what an agent must represent
The epistemic state is part of the agent state
An agent does not merely need a task representation. It needs a representation of the relationship between the task, its evidence, and its available actions. At minimum, that representation should answer five questions.
Which facts were directly supplied or successfully observed?
Which variables prevent a supported answer or action?
Does the fact belong to the user, a tool, a primary record, or inference?
Has the task’s evidential threshold already been met?
What observation would change the current conclusion?
This decomposition became essential. A tool benchmark grades the entire chain. If a model correctly recognizes that an internal record ID must be looked up, but serializes the lookup arguments incorrectly, the environment records one failure. If it chooses the right tool but invents the ID, that is another. If it receives the ID, calls the right mutation, and then fails to stop, that is a third. End-to-end scores cannot tell us which layer changed.
03 · the experimental program
From forecasts to agent trajectories
We began with staged forecasting because it creates a natural epistemic timeline: an event starts underspecified, context arrives, prior evidence can be retrieved, a probability is issued, and eventually the event resolves. We then widened the curriculum beyond forecasting to user questions, search, computation, source conflict, stale data, missing capabilities, recovery, and stopping.
Expose an underspecified future event, then reveal context and outcome evidence in stages.
Train ask, search, compute, answer, abstain, revision, source authority, and stopping decisions.
Test whether those decisions survive unfamiliar tools, schemas, IDs, errors, and environment state.
Train 18,000 assistant turns as real Qwen tool calls, with direct and explicit-decision variants.
Remove execution entirely and score the justified next action from option logits in EpiState.
The native tool-training run
The largest controlled run used Qwen2.5-14B-Instruct with a rank-8 LoRA on the attention query, key, value, and output projections: 12.58 million trainable parameters, or 0.085% of the model. We trained for one epoch at a 3×10−5 learning rate with assistant-only loss and a 2,048-token context.
assistant turns per variant
multi-step episodes
scenario families
native variants
Native-direct learned only the next assistant action or tool call. Native-decision first emitted a compact ledger—known facts, missing facts, source, next action, and reason—then emitted the same native tool call. The data included discovery before action, user-owned parameters, temporal dependencies, state recovery, typed errors, evidence integrity, arithmetic, and a 4,320-turn forecasting family.
Alongside these variants, we retained two earlier policies. Mixed-v2 was trained on 20,000 single-turn epistemic decisions. Agent-loop was trained on 30,000 snapshots that explicitly rewarded asking, searching, computing, answering, abstaining, and stopping once supported.
04 · the apparent failure
Every trained agent looked worse end to end
On 52 external ToolSandbox scenarios, the unmodified base model remained strongest. The native-decision model recovered most of the damage introduced by native-direct training, but it did not beat base.
Manual rollouts showed a repeatable split between planning and execution. Native-decision often named the right next tool in its ledger, then emitted the tool’s arguments as a JSON-encoded string instead of an object. In other cases it invented an internal identifier, interpreted a tool error as evidence of absence, or continued acting after the task was complete.
next_action: lookup_contact✓arguments: "{"name":"Ada"}"×expected object; received string×The native variants also changed the agent’s behavioral prior. Native-decision made 84 tool calls where base made 67, while asking only 8 questions where base asked 23. Explicit epistemic prose did not automatically create epistemic restraint; in this run, it often became a preamble to another action.
05 · isolating the claim
A benchmark with no tools, parser, or model judge
ToolSandbox could not answer the causal question we cared about: did training improve the model’s judgment about its information, independently of its ability to execute that judgment? We built EpiState to isolate that layer.
Each item freezes one checkpoint in an agent’s work. The model sees the task, the complete current observation state, and five natural-language next steps. It answers with one letter. We read the five option logits directly: there is no free-form generation, tool execution, JSON parser, agent loop, or model judge.
Every semantic option occupies A, B, C, D, and E exactly once through five cyclic permutations. We map the probabilities back to actions before averaging, preventing a letter-position preference from masquerading as epistemic competence. Exact benchmark phrases were absent from the relevant training corpora.
Mixed-v2 corrected 14 base errors with no regressions; the paired exact test was p=.000122. Agent-loop corrected 15 with no regressions; p=.000061. But all gains belonged to a single conceptual family—dependent evidence—so the conservative interval obtained by resampling the 12 families touched zero. The item-level effect is real; the broad-capability claim is not yet established.
Logs report POOL_LIMIT. Configuration shows pool_limit=12, below the documented minimum of 30.
06 · what changed
The learned capability is evidence sufficiency
Base Qwen already solved freshness, authority, user-owned information, authorization, unobservability, forecast context, recovery, and identifier grounding in this deliberately legible benchmark. Its 15 failures all had the same shape: the required evidence was visibly present, yet it selected the tool that would retrieve that evidence again.
Agent-loop training eliminated all 15. On the example above, base assigned 41.7% probability to diagnosing and 58.3% to rereading configuration. Mixed-v2 moved diagnosis to 54.5%. Agent-loop moved it to 98.1%. Native-direct stayed at 37.8%, and native-decision fell to 20.2%.
Mechanism claimPost-training can teach a model to recognize that evidence already in context is sufficient, transition from acquisition to synthesis, and avoid a redundant information-gathering action.
This explains the otherwise contradictory result. The earlier single-call curricula directly supervised stop once supported. The native trajectories contained many more tool actions than user questions and rewarded the gold next call at each prefix. Token-level likelihood favored locally imitating the trajectory, not minimizing unnecessary actions or maximizing eventual task success.
Does the context already support an answer?
Can the model emit grounded, valid tool actions?
The environment requires both abilities at once.
Why native training failed to transfer
- Capability interference.A narrow synthetic tool distribution displaced parts of the base model’s broad function-calling behavior.
- Template saturation.Near-zero training loss indicated that the model had learned the trajectory surface faster than the general policy beneath it.
- Schema non-transfer.Correct abstract plans did not guarantee correct field names, types, required arguments, or environment-specific identifiers.
- Exposure bias.Gold prefixes taught recovery from authored states, not from the model’s own malformed actions and cascading errors.
- Objective mismatch.Assistant-token cross-entropy rewards the next reference token; an agent needs grounded task success with penalties for guessing, repetition, and unnecessary calls.
what the evidence does—and does not—say
A narrow result is more useful than a vague success
- Evidence-sufficiency judgment can move under post-training.
- The effect survives unseen wording and fully balanced answer positions.
- Better epistemic choice does not imply better agent execution.
- Explicit decision traces are not sufficient by themselves.
- A general “theory of knowledge” capability.
- Broad gains across all 12 epistemic families.
- Better calibration on real future events.
- Better ToolSandbox agents than base.
- A claim about every model size or training objective.
EpiState is a mechanism benchmark, not a naturalistic agent benchmark. It intentionally makes information boundaries auditable. A publishable claim needs both: controlled checkpoints that identify what changed, and external trajectories that test whether the change matters in real workflows.
07 · the next decisive experiment
Train the transition, preserve the executor
The next run should not simply add more epistemic prose to more native trajectories. It should preserve the base model’s tool distribution while explicitly optimizing the boundary between information acquisition and execution.
The strongest external test would transform blinded ToolSandbox or τ-bench trajectory checkpoints into EpiState-style choices without exposing their gold calls. That would ask whether the sufficiency policy transfers beyond authored templates. The corresponding end-to-end run would then determine whether better schema preservation and recovery carry the policy through execution.
Show a controlled improvement in multiple epistemic transitions, preserve general tool execution, and predict which external agent trajectories improve before running them.
definitions
Terms used in this report
- Epistemic state
- The facts, uncertainties, sources, dependencies, and sufficiency conditions relevant to the current task.
- Epistemic control
- Selecting a justified information or action transition from that state.
- Grounding
- Binding an action’s claims and arguments to values actually supplied or observed.
- Evidence sufficiency
- The point at which the task can be answered or acted on without acquiring another fact.
- Agent success
- Completing the task through valid actions while respecting information, authorization, and stopping constraints.