Belief Stack v0.3 — pre-registration

§0 The conceptual frame

Every agentic system today pays the reconstruct-world-model-every-step tax.

At each step, the agent re-reads its prompt history, retrieved documents, memory stores, tool outputs, and prior plans to re-infer one question:

What do I currently believe?

That reconstruction work happens implicitly inside the model’s context window at every step. It is expensive in tokens, in latency, and in failure modes (stale assumptions, missed contradictions, false completions).

v0.3 is the first direct test of whether maintained belief state can reduce that tax: not by compressing context, but by maintaining state with provenance and lifecycle outside the context window, so the agent does not need to reconstruct it.

The differentiator is maintained state with provenance and lifecycle. A summary compresses. A belief state persists, evolves, weakens, strengthens, and can be inspected.

§1 Research question

Can an agent perform a planning task using a maintained belief state with substantially less raw context, at equal or better quality than an agent using larger raw context?

§2 Hypotheses

2.1 Primary hypothesis

A lifecycle-managed belief overlay will outperform or match larger raw-context bundles, because it provides maintained state, provenance, and freshness signals rather than forcing the model to reconstruct the workflow state from evidence at every step.

2.2 Secondary hypothesis

The combination of belief overlay + minimal targeted evidence (Arm C) will outperform or match large raw-context bundles (Arm A) more reliably than belief overlay alone (Arm B). Arm C may represent the realistic production architecture: belief state for state-keeping, minimal evidence for execution-time detail.

2.3 Scope discipline (bounded claim)

The narrower claim is the one we have evidence for and the one we will defend.

§3 Core claim under test

This is not generic context compression. A summary compresses. A belief state persists, evolves, weakens, strengthens, and can be inspected.

The claim is:

A smaller bundle of lifecycle-managed beliefs can support planning better than a larger bundle of raw evidence, because the belief state preserves what is currently believed, why, and whether it is still valid — not because it is smaller.

The differentiator is maintained state with provenance and lifecycle, not compression.

§4 Arms

4.1 Fair-comparison constraint (load-bearing)

The experiment must distinguish “belief state helps planning” from “the belief builder did extra work the raw-context arm never got.”

This is the only way the v0.3 comparison cleanly attributes any observed lift to the belief-state architecture itself.

4.2 State vs scratchpad

Belief Stack is a state substrate. Raw evidence may still be required as scratchpad. They are different objects and serve different functions in planning.

The v0.3 hypothesis is that maintained state can substitute for the work of reconstructing state from raw evidence at every step. The hypothesis is not that scratchpad can be eliminated. Arm C tests precisely this split: belief state for state-keeping, minimal targeted evidence for execution detail.

§5 Task type

Workflow planning task with changing state. Examples:

• user goal changes midstream
• prior assumption becomes contradicted
• tool output introduces new constraint
• agent must decide next action

The task surface should reward correct handling of state evolution, not just one-shot question-answering.

§6 Evaluation criteria

6.1 Planning quality (primary outcome)

1. Correct next action
2. Avoids acting on stale or contradicted belief
3. Identifies uncertainty when evidence is missing
4. Produces coherent plan
5. Can explain which belief justified the action

6.2 Operational telemetry (primary outcome — quantifies the tax)

Each arm tracks, per planning task:

• Cumulative tokens consumed across the workflow
• Latency per planning step
• Total latency across the workflow

The planning-quality result is the headline. The token / latency result quantifies the reconstruct-world-model-every-step tax — turning the §0 framing into a measured number rather than rhetoric.

6.3 Failure-mode catalog

Inheriting v0.2.2’s deterministic metrics where applicable; adding one v0.3-specific failure mode.

Grounding bankruptcy — the agent has the correct high-level workflow state (active assumptions, constraints, pending validations) but lacks the evidentiary detail needed to execute correctly (specific filenames, exact error messages, raw parameters). The belief state is right; the scratchpad is missing. This is the predicted failure mode of Arm B if it fails, and the failure mode Arm C is designed to prevent.

§7 Success condition (locked)

Arm B or arm C matches or beats arm A under both gates simultaneously:

• Token gate (D7): Mean input-token count is ≤ 50% of Arm A’s mean input-token count across the 75-question set.
• Quality gate (D8): Aggregate planning-correctness rate is within 2 percentage points of Arm A’s rate, or strictly exceeds Arm A’s rate.

A “win” requires both gates to clear. Either gate failing means the headline claim is not supported for that arm.

§8 Failure condition (locked)

Arm A outperforms both belief-based arms on planning correctness by more than 2 percentage points, or belief-based arms do not achieve the ≥ 50% token reduction (rendering the “materially fewer tokens” claim unfalsifiable for them), or belief-only planning (Arm B) fails specifically via the grounding-bankruptcy failure mode at a materially higher rate than Arm C — indicating that the architecture needs scratchpad-as-evidence to function, which would weaken the headline “maintained state for planning” claim.

§9 What this would prove

If successful, v0.3 supports the claim that Belief Stack is not only useful for human inspection (the v0.1 / v0.2.2 result). It is also useful as a runtime planning substrate for agents.

§10 What this would not prove

It would not prove agents can fully replace context with beliefs. It would only show that, for bounded planning tasks, lifecycle-managed belief state can reduce the need to reconstruct workflow state from raw context.

§11 Locked language

The language discipline matters: “context compression” puts this in a crowded category (chunking, summarization, RAG variants). The differentiated framing is that the maintained state is warranted and lifecycle-managed, and that scratchpad-style raw evidence has its own legitimate place in execution. Use the reconstruct-world-model-every-step tax as the named operational pain point this work is targeted at.

§12 Locked decisions

§13 Anti-curation discipline (carry-forward)

• Task / question text generated blind to the substrate state that will be ranked.
• Belief overlay rendering policy (ranking, dedup, budget) locked before answer generation.
• Seeds, model IDs, judge configurations locked here before any data flows.
• All failures preserved with metadata; no silent truncation.
• No prompt tuning after seeing outputs.

These are the same constraints v0.1 and v0.2.2 ran under and that produced credible results. They apply to v0.3 unchanged.

§14 Non-goals

• v0.3 does not test multi-agent coordination.
• v0.3 does not test long-horizon (multi-hour) planning.
• v0.3 does not test belief overlay generation from a real event stream — overlays are rendered from the locked v0.1 substrate format.
• v0.3 does not claim a general theory of agent reasoning. It tests a specific, bounded planning claim on a specific substrate.
• v0.3 does not displace the human-inspection surface. The dual-consumer framing remains load-bearing; v0.3 just tests the agent surface specifically.

§15 Framing question — resolved 2026-06-03

The earlier open question (does the v0.3 design distinguish “belief state helps planning” from “the substrate-builder did extra work Arm A never got”?) is resolved as the fair-comparison constraint in §4.1: Arm A must receive the same underlying evidence used to derive the belief state.

This is now a load-bearing design constraint, not an open question.