The Context Window Between Us: A Falsifiable Framework for Human–AI Coordination

A model can retain every token in a conversation and still lose the thing the conversation was doing. The Meta-Context Window framework names that missing layer, describes how it breaks, and proposes experiments that can prove the framework wrong.

Think about the last time an AI conversation went subtly wrong.

The transcript was still there. The model could quote your earlier instructions. Nothing had obviously exceeded the token limit. Yet the interaction no longer felt like the same collaboration. A distinction that mattered had been flattened. An assumption had changed without being named. The answer was locally reasonable and globally wrong.

We usually describe this as a model failure: the model forgot, misunderstood, hallucinated, or lacked enough context. Sometimes that diagnosis is correct. But it collapses two different problems into one:

1. Capability: Can the model perform the task?
2. Coordination: Do the human and model still share the same working meaning, priorities, constraints, and sense of where the task stands?

The Meta-Context Window (MCW) Framework starts from a simple claim:

Many failures in human–AI collaboration are coordination failures, not capability failures.

MCW is an early-stage, open-source framework for describing those failures without pretending they have already been empirically solved.

Evidence status: The definitions, taxonomy, measurement rubrics, and experimental protocols exist. The experiments have not yet been piloted. No quantitative validation exists. Current empirical claims remain at the illustration and practitioner-observation layers.

Three context windows, not one

AI work usually treats “the context window” as a property of the model. That is only one of the three windows involved in an interaction.

The canonical definition is:

The MCW is the dynamically maintained shared state of meaning, salience, intent, and continuity that emerges through ongoing interaction between one or more HCWs and one or more ACWs.

The framework writes this relationship as:

$$MCW=f(HC{W}_1\dots HC{W}_n,AC{W}_1\dots AC{W}_m,T,C)MCW\; =\; f(HCW\_1\; \backslash ldots\; HCW\_n,\; ACW\_1\; \backslash ldots\; ACW\_m,\; T,\; C)$$

Here, $T$ represents time and interaction history; $C$ represents channel constraints such as bandwidth, latency, modality, and noise.

This notation is a mnemonic, not completed mathematics. The framework does not yet define a codomain, functional form, or computable probability distribution over coordination state. Calling the notation “formal” must not smuggle in a level of mathematical completion that does not exist.

What the notation does capture is the ontology: the MCW is neither the human’s state nor the AI’s state, and it is not their sum. It exists only in the coupling between them.

Information Units: what crosses the boundary

The framework describes coordination in terms of Information Units (IUs):

An Information Unit is the minimal transferable element of information that can influence coordination state between actors.

An IU can be a stated assumption, a correction, a goal, a constraint, a distinction, or a change in priority. IUs are substrate-independent but representation-dependent: the same IU can be encoded in speech, prose, a diagram, a tool result, or a sequence of tokens.

An IU is not a token. Tokens are one possible representation. It is not necessarily a fact, belief, or sentence either. How to count and segment IUs reliably remains an open problem, so the framework avoids quantitative IU claims that depend on a solved individuation rule.

During communication, an IU passes through five stages:

1. Selection: Which part of an actor’s internal state is chosen for transfer?
2. Encoding: How is it compressed into language, symbols, or another representation?
3. Transmission: What survives the channel?
4. Decoding: How does the receiving actor reconstruct it using prior context?
5. Integration: How is it weighted and incorporated into the receiving context window?

A conversation can preserve the words while damaging the IU. A summary can be textually accurate yet discard the distinction on which the next decision depends. A refusal can be policy-compliant yet hide the constraint needed for the human to repair the interaction. A user can think for three hours between turns, update five assumptions, and transmit none of them.

That is why a larger ACW is not, by itself, a larger or healthier MCW.

Six ways shared coordination breaks

The current taxonomy defines six failure modes:

These are not six colorful names for “the prompt was bad.” They make different predictions and imply different repairs.

For example, Asymmetric State Advancement is not ordinary forgetting. Imagine that I leave a conversation, reason about the project overnight, and return with a changed architecture. My HCW advanced continuously. The AI’s ACW did not receive those changes. If I continue as though it did, the resulting phase lag can feel like model incompetence even when the model is correctly reasoning from the state I actually supplied.

Resetting the conversation can make this worse. A reset clears ACW state; it does not externalize the missing HCW state. The repair is synchronization: I must transmit the assumptions and decisions that changed off-turn.

Repair is an operation, not a vibe

The framework treats repair as a first-class coordination primitive. Its five canonical repair operations are:

• Re-grounding: Reintroduce foundational goals and assumptions.
• Decompression: Expand a compressed bundle back into the distinctions it replaced.
• Re-weighting: Explicitly change which IUs carry the most salience.
• Disambiguation: Split an overloaded term or statement into its competing interpretations.
• Synchronization: Externalize state changes that occurred outside the shared exchange.

This operation set is intentionally small. It also remains incomplete: the current canon does not yet designate a unique repair operation for Constraint Opacity or Repair Suppression. Proposed extensions exist, but they are not silently promoted into the canonical set.

The framework’s falsifiable timing claim is deliberately modest:

Repair cost is predicted to be non-decreasing in discovery lag.

Earlier language claimed that late repair becomes “exponentially” more expensive. That asserted a functional form without data, so the claim was withdrawn. The surviving claim is an ordering claim: holding other factors constant, discovering a coordination failure later should not make repair cheaper.

The novelty claim - and the debt

MCW does not claim to have discovered common ground or conversational repair.

Conversation analysis has studied repair for decades. Grounding theory describes how people establish evidence of mutual understanding. CSCW studies the articulation work required to coordinate. Distributed cognition explains why meaningful cognitive state can be carried across people and artifacts rather than located inside one mind.

The framework’s related-work analysis states the debt plainly:

In interactions between cognitively similar parties with symmetric memory and continuous availability, MCW reduces substantially to grounding theory and should not replace it.

The framework’s load-bearing bet appears where that symmetry breaks.

The human window is continuous, situated, and non-enumerable. The artificial window is discrete, formally bounded, and input-driven. MCW predicts coordination failures specific to that asymmetry - especially failures caused when the human advances off-turn and assumes the artificial system advanced with them.

That wager is testable. If matched Human↔AI and Human↔Human interactions show no coordination signature specific to the HCW/ACW asymmetry, then existing grounding and repair theories explain the phenomenon and MCW becomes redundant vocabulary.

That is not an edge-case falsifier. It is the center of the framework.

A research program designed to lose

The repository currently specifies five original experiments plus a sixth, Drift Accumulation, introduced as a declared extension:

1. False Alignment Injection
2. Asymmetric State Advancement
3. Overcompression Damage
4. Constraint Opacity Stress Test
5. Repair Signal Suppression
6. Drift Accumulation

Each is comparative, tied to a specific failure mode, and paired with a condition that counts against the framework. The current protocols use four anchored ordinal proxies:

• H - MCW Health: evidence that the parties share the current goal, assumptions, and constraints
• R - Repair Cost: observable effort required to restore coordination
• D - Drift: late-discovered divergence across an interaction window
• M - Misattribution: the degree to which a Human↔AI coordination failure is blamed on actor capability; M is not scored in the Human↔Human experiments

These are coordination observables, not validated interval-scale metrics. The repository includes anchored rubrics and version-controlled pre-registration templates because a taxonomy that can reinterpret every outcome as support is not a research framework. It is a vocabulary with armor.

The MCW Constitution therefore imposes an epistemic floor: no claim can be stated above the evidence layer that supports it, summaries must preserve falsification conditions, and extensions must declare how they depart from canon.

System prompts are initialization artifacts

One practical consequence follows immediately from the framework.

A system prompt is a preloaded, static IU bundle injected into the ACW before the MCW forms. It can bias initial conditions, but it cannot observe later drift, discover a changed human assumption, or dynamically repair shared state. It is an initialization artifact, not a coordination solution.

That does not make system prompts useless. It changes what they are for.

An MCW-aware initialization can make priorities, constraints, uncertainty, and repair permission legible at the start of an interaction:

This interaction is a collaborative problem-solving process.

Primary goal:
- [State the immediate objective in one sentence.]

Coordination norms:
- Surface uncertainty early.
- Ask for clarification when assumptions are unclear.
- If misalignment is detected, pause progress and repair first.

Constraints:
- Name relevant constraints and explain their effect on the response.
- Distinguish reasoning, capability, and policy constraints.

Compression:
- Do not over-summarize unless requested.
- Preserve assumptions and edge cases when summarizing.

Repair:
- If the interaction drifts, name the drift.
- Re-ground goals and assumptions before continuing.

This prompt is an initialization aid, not a complete specification.
The interaction may adapt as shared understanding evolves.

The critical prediction is counterintuitive: a good MCW-aware prompt should not eliminate the need for re-grounding. If a static prompt genuinely maintained coordination throughout long, changing interactions without later repair, it would falsify the framework’s claim about prompts.

What exists now

The project is deliberately published before validation is complete so that its vocabulary, experiments, and failure conditions can be challenged in public.

Current artifacts include:

• the canonical glossary
• framework diagrams
• the failure-mode and repair specifications
• measurement rubrics and experimental protocols
• a methods-oriented paper outline
• a public exploratory Custom GPT test bed
• a planned Hugging Face Space for reproducible baseline-versus-MCW-aware comparisons

The Space is the natural next artifact: a versioned instrument that lets users run the same task against matched baseline and MCW-aware conditions, then export a result snapshot without treating the demo as proof.

Try to break it

The MCW Framework is Apache-2.0 licensed. It is meant to be forked, criticized, operationalized, and - if the experiments fail - retired or revised.

• Read the framework: rainmana.github.io/mcw-framework
• Inspect or fork the source: github.com/rainmana/mcw-framework
• Read what would make it redundant: Related Work & Positioning
• Inspect the governance rules: MCW Constitution
• Open an issue: GitHub Issues

The useful question is not whether the framework sounds intuitively true. Many unfalsifiable ideas do.

The useful question is whether MCW produces differential predictions, reliable diagnoses, and cheaper repair than the theories and practices we already have.

If it does, we have named a missing layer in human–AI collaboration.

If it does not, the experiments should make that visible.

---

Citation

@software{akin2026mcw,
  author  = {William Alec Akin},
  title   = {Meta-Context Window (MCW) Framework},
  year    = {2026},
  url     = {https://github.com/rainmana/mcw-framework},
  license = {Apache-2.0}
}

Machine-readable citation metadata, including ORCID, is available in CITATION.cff.