The rest of the conversation after the survey in the first comment: the voice layer, the 50% compaction, and the three diagrams.
The project can be framed as epistemic compaction: an iterative research practice that transforms an exploratory humanâlanguage-model session into a smaller, versioned, evidence-linked state from which another inquiry cycle can begin.
The important word here is not only compression. It is epistemic, because the material being compressed includes claims, evidence, interpretations, decisions, uncertainty, relationships, and future questions.
A ZIP archive provides a useful starting analogy, but it is not an exact model. ZIP compression is designed to be lossless: the original files can be reconstructed exactly. A language-model summary is normally lossy. Once a long conversation becomes a short summary, certain distinctions may disappear: who proposed an idea, how confident they were, which evidence supported it, what was rejected, and what remained unresolved.
This creates the first distinction.
Token compression reduces the number of tokens sent to a model.
Semantic compaction converts a session into a smaller research checkpoint.
Terminological compaction gives a recurring conceptual structure a shorter expression.
Archival compression preserves material for exact recovery.
These operations may support one another, but they do not preserve the same things.
A coined term, for example, does not contain a complete definition in the way a ZIP archive contains reconstructible bytes. Its meaning depends on convention, morphology, prior discourse, examples, social history, and community uptake. A term is therefore less like a container and more like an address: it points toward a larger conceptual territory.
The Indigenous-language analogy needs to remain careful and situated. It would be inaccurate to claim that Indigenous peoples simply invent words to compress concepts. Every language creates compact expressions, but those expressions belong to particular linguistic, historical, territorial, and social systems.
A safer formulation is that some linguistic practices demonstrate how a compact expression can index a much larger relational field. That observation can generate questions for context engineering. It cannot be treated as a generic technical method to extract from Indigenous languages without language-specific scholarship, community involvement, and appropriate authority.
For the Narrative Context Protocol and the Wampum narrative engine, this distinction becomes especially important. The information carried into the next cycle may not be only a proposition. It may also be a relational commitment: who established the knowledge, where it applies, who may interpret it, what restrictions accompany it, and what responsibilities follow from reusing it.
Several academic fields meet around this problem.
Computational linguistics contributes prompt compression, summarization, discourse segmentation, information extraction, and long-context evaluation.
Pragmatics contributes implicature, presupposition, deixis, speaker commitment, and common ground. These concepts help identify meaning that may never have been explicitly stated but was nevertheless active in the conversation.
Discourse studies contribute coherence and rhetorical relations. They help preserve whether one statement supported, contradicted, qualified, explained, or resulted from another.
Terminology studies examine concept systems, definitions, naming practices, synonym control, and the conditions under which a phrase becomes a stable term.
Morphology and lexicology examine compounding, derivation, lexicalization, and neologism: the processes through which complex meanings become reusable linguistic units.
Information theory contributes the distinction between lossless and lossy coding, along with the idea of a rateâdistortion trade-off. A representation may become smaller, but the relevant question is how much task-important information becomes distorted.
Knowledge organization contributes ontologies, controlled vocabularies, concept maps, and machine-readable relations.
Qualitative methodology contributes coding, memo-writing, comparison, and progressive category formation. Grounded-theory practices are especially relevant because they do not move directly from raw observations to a final theory. They preserve analytic memos, provisional codes, emerging categories, and unresolved tensions across multiple cycles.1
Humanâcomputer interaction and computer-supported cooperative work contribute common-ground theory, external cognition, boundary objects, and handoff design.
Software engineering contributes checkpoints, event logs, schemas, provenance, branching, version control, and reproducible state transitions.
Archival studies contribute appraisal and traceability: deciding what can be discarded, what must remain recoverable, and which record remains authoritative.
Indigenous methodologies contribute relationality, situated authority, knowledge stewardship, and the recognition that some knowledge should not be detached from the relationships that make it meaningful.
Existing prompt-compression research addresses only part of this wider problem.
LLMLingua uses a smaller model to identify less important prompt material. It first reduces larger prompt components and then performs token-level compression. Its researchers report compression ratios of up to twenty times on selected tasks, but those results should not be interpreted as a universal level of safe compression. The appropriate ratio depends on the task and on what information must survive.2
Other systems encode contexts into special gist, memory, or compressed tokens. Research suggests that even the positional relationship between source tokens, compressed representations, and later queries can affect how well the original context remains usable.3
A particularly relevant direction decomposes text into smaller discourse units before compression. This is close to the decomposition work from episode forty-four. Rather than asking a model to summarize an undifferentiated transcript, the protocol first identifies atomic statements and their relationships. Compression happens only after decomposition.
Commercial agent systems increasingly distinguish active context from persistent memory. OpenAI describes compaction as reducing context size while preserving the state required for subsequent turns. Anthropic pairs context editing with persistent memory so that active tool results can be removed while information needed over longer periods remains externally available.45
This produces three separate questions.
Compaction asks: what must remain active for the next cycle?
Memory asks: what might become relevant again?
The archive asks: what must remain recoverable even if the current model does not select it?
The distinction matters because abstractive summaries can sound coherent while changing entities, quantities, causal relations, temporal order, negation, or confidence. Traditional overlap metrics are not sufficient to detect all of these failures.67
A compressed checkpoint therefore cannot be validated only by asking whether it sounds accurate. It must be tested by attempting to continue the work from it.
The working definition becomes:
Epistemic compaction is the deliberate transformation of an exploratory research trace into a reduced, structured, provenance-bearing state intended to support reliable interpretation and continuation across humanâmodel work cycles.
The process begins with exploration.
During this phase, the researcher and model can work relatively openly. They can retrieve sources, decompose questions, propose interpretations, invent provisional expressions, use tools, and follow unexpected connections.
The full event stream should remain in an external archive. The active context should not be the only surviving record.
As the session progresses, significant events can be marked by type: claim, source, decision, definition, contradiction, assumption, question, procedure, relationship, restriction, or next action.
The second phase detects a boundary.
Compaction should not occur only because a token limit is approaching. It should also happen at meaningful research transitions.
A boundary may appear when a question has been provisionally answered, a design decision has been made, a literature-search phase has ended, a contradiction has been isolated, the inquiry changes discipline, another person or agent must take over, or the next action no longer requires the entire exploratory sequence.
Completion should not be expressed as a falsely precise percentage. A better approach is to examine several dimensions.
Have the important claims been identified?
Is their evidence recorded?
Are decisions and their rationales explicit?
Are uncertainties attached to particular claims?
Is there an executable next action?
Compaction can begin when every dimension is at least partially represented and the next action is clear enough for another cycle to start.
The third phase is decomposition.
Before generating a summary, the session is converted into atomic records.
A claim record might contain an identifier, a statement, a provisional status, a confidence level, supporting sources, contradictions, dependencies, the session in which it appeared, and any restrictions on reuse.
This atomic representation prevents a fluent paragraph from hiding differences between evidence, inference, speculation, and decision.
The fourth phase produces three layers.
The first is a human abstract. It explains what happened in readable language.
The second is a machine handoff. It records claims, decisions, constraints, open questions, retrieval pointers, and next actions in a structured form.
The third is a terminology ledger. It records each working term, its definition, provenance, aliases, examples, relationships, revisions, and governance conditions.
One representation should not be forced to perform all three functions. A polished abstract is optimized for human comprehension. A machine handoff is optimized for operational fidelity. A terminology ledger is optimized for conceptual stability and revision.
The fifth phase is validation.
The model generates questions from the original session. Some questions test central claims. Others test uncertainty, rejected alternatives, chronology, contradictions, source relationships, and restrictions.
A second model answers those questions using only the compacted checkpoint.
The answers are then compared with the original record.
This reveals omissions, unsupported additions, altered confidence, broken relationships, and information that must remain available through retrieval.
This process acts as a semantic checksum. It does not reconstruct every word. It tests whether the information required for responsible continuation remains recoverable.
The sixth phase is resumption.
The next cycle begins with the machine handoff, the relevant terminology entries, retrieval pointers, the active research question, and a statement of what the model must not assume.
Before doing new work, the model restates the research state and identifies anything it cannot reconstruct. This makes failures visible before they become embedded in another cycle.
The seventh phase is recompaction.
Earlier checkpoints are never silently overwritten. Each cycle produces a new version.
The lineage moves from the raw trace to checkpoint one, then to cycle-two additions, checkpoint two, cycle-three additions, and checkpoint three.
Every checkpoint records what was added, removed, generalized, renamed, contradicted, or downgraded in confidence.
The experiment should compare several context conditions.
The first condition uses the full original transcript.
The second uses an ordinary rolling summary.
The third uses selected passages copied directly from the transcript.
The fourth uses the structured epistemic capsule.
The fifth uses an automated token-compression method such as LLMLingua.
Each condition receives the same continuation tasks.
The evaluation records token reduction, claim recall, provenance coverage, downstream task success, contradiction rate, unsupported claims, preservation of dissenting observations, preservation of uncertainty, human resumption time, archive retrievals, repair operations, terminology stability, and stewardship violations.
The objective is not maximum compression. It is to find a useful position on a trade-off curve.
A representation that saves many tokens but destroys provenance is not successful.
A representation that preserves factual claims but erases uncertainty is not successful.
A representation that allows the model to continue while preventing a human from auditing its state may also be unacceptable.
Several research questions follow.
Which kinds of meaning disappear first during compaction: explicit propositions, pragmatic implications, chronology, causal relations, confidence, or social obligations?
Does decomposition into discourse units preserve research continuity better than direct summarization?
Can preserving unresolved questions sometimes be more valuable than preserving provisional answers?
When does a recurring phrase become a stable research term?
Can a term include what might be called decompression instructions: a definition, examples, provenance, relationships, exclusions, and revision history?
How does terminology invented by a model differ from terminology negotiated within a scholarly or community context?
Is the best checkpoint for a human also the best checkpoint for a model?
Should the human-readable and model-operational versions be separate but linked?
How can a capsule preserve who knew what, what was implied, what was contested, and what participants treated as common ground?
How should deixis be handled when expressions such as âthis approach,â âthe earlier problem,â or âwhat we decidedâ lose their referents after the original conversation is removed?
Can speaker commitments and pragmatic context be represented without reproducing the entire dialogue?
For Indigenous and community-held knowledge, which material should not be compacted, generalized, detached from place, or transferred between systems?
Who has the authority to approve a compressed representation?
How should restricted circulation, refusal, ceremony, contextual integrity, and relational accountability appear in a machine handoff?
Can completion confidence be estimated from missing provenance and unresolved questions?
Are epistemic phase boundaries more reliable than token thresholds?
What signs reveal that a session is being compacted too early?
The systems involved should have different responsibilities.
GitHub can hold the canonical schemas, protocol versions, experiment code, and checkpoint history.
Notion can hold the human-facing literature map, glossary, concept pages, and article drafts.
Slack can support discussion, review requests, checkpoint notifications, and disagreement threads.
An external archive can preserve full transcripts, source documents, and restricted records.
The active language-model context should contain only the current capsule and the evidence retrieved for the present task.
The first repository structure can include a protocol directory for schemas and validation rules, a session directory for episode forty-four, a literature directory organized by field, and an experiments directory containing each comparison condition.
The articleâs abstract can state the problem this way:
Large-language-model research sessions accumulate prompts, retrieved sources, tool traces, provisional interpretations, decisions, and unresolved questions. Existing prompt-compression methods primarily optimize token use or downstream performance, while conventional summaries may produce readable but insufficiently grounded accounts.
The article proposes epistemic compaction as an iterative method for transforming an exploratory humanâmodel interaction into a reduced, versioned, and provenance-bearing research state.
The method decomposes the session into typed claims, evidence, decisions, uncertainties, relationships, restrictions, and next actions.
It then produces separate human-readable, machine-operational, and terminological representations.
Compaction occurs at epistemic phase boundaries and is evaluated through question-based reconstruction, downstream continuation tasks, claim-level provenance, contradiction detection, and human resumption effort.
The framework connects prompt compression with pragmatics, discourse studies, terminology, qualitative memo-writing, knowledge organization, agent memory, and research governance.
It distinguishes technical compression from language-specific and Indigenous practices of relational meaning, requiring contextual authority wherever Indigenous knowledge is involved.
The experiment does not seek a universally minimal prompt. It asks what must remain recoverable for inquiry to continue responsibly.
The current working term is epistemic compaction.
The central claim is that a research session should become a structured, provenance-bearing state rather than only a prose summary.
The proposed cycle is: explore, detect a boundary, decompose, generate three representations, validate, resume, and recompact.
The three representations are the human abstract, the machine handoff, and the terminology ledger.
The primary evaluation dimensions are token reduction, claim recall, grounding, provenance coverage, continuation performance, and human resumption effort.
The unresolved problems are the final terminology, language-specific Indigenous collaboration, calibration of the stopping rule, representation of implicature and common ground, and the balance between human readability and machine efficiency.
The next cycle is to build the checkpoint schema and apply it retrospectively to episode forty-four, comparing it against an ordinary rolling-summary baseline.
â
The project can be framed as epistemic compaction: an iterative practice that transforms an exploratory humanâlanguage-model session into a smaller, versioned, evidence-linked state from which another inquiry cycle can begin.
The material being compacted is not only text. It includes claims, evidence, interpretations, decisions, uncertainty, relationships, restrictions, and future questions.
A ZIP archive is a useful but incomplete analogy. ZIP compression is lossless: the original files can be reconstructed exactly. Language-model summaries are usually lossy. When a conversation becomes a short summary, distinctions may disappearâwho proposed an idea, what supported it, how confident participants were, what was rejected, and what remains unresolved.
This produces four distinct operations.
Token compression reduces the tokens sent to a model.
Semantic compaction converts a session into a research checkpoint.
Terminological compaction gives a recurring conceptual structure a shorter expression.
Archival compression preserves material for exact recovery.
A term does not contain its definition as a ZIP archive contains reconstructible bytes. Its meaning depends on convention, context, examples, social history, and community uptake. A term is therefore less like a container than an address pointing toward a larger conceptual territory.
The Indigenous-language analogy must remain situated. It would be inaccurate to say that Indigenous peoples simply create words to compress concepts. Indigenous expressions belong to particular languages, histories, territories, communities, and systems of responsibility.
A safer formulation is that some linguistic practices demonstrate how a compact expression can index a larger relational field. This can generate questions for context engineering, but it cannot become a technique extracted from Indigenous languages without language-specific scholarship, community participation, and appropriate authority.
For the Narrative Context Protocol and Wampum narrative engine, the unit carried forward may therefore be more than a proposition. It may include who established the knowledge, where it applies, who may interpret it, and what responsibilities accompany its reuse.
Several academic fields converge here.
Computational linguistics contributes prompt compression, summarization, discourse segmentation, and long-context evaluation.
Pragmatics contributes implicature, presupposition, deixis, speaker commitment, and common groundâmeaning that may be active without being explicitly stated.
Discourse studies contribute coherence and relations among claims: support, contradiction, qualification, explanation, and consequence.
Terminology studies examine concept systems, naming, definitions, synonyms, and the conditions under which phrases become stable terms.
Morphology and lexicology examine lexicalization, compounding, derivation, and neologism.
Information theory contributes lossless and lossy coding, along with the trade-off between reduction and distortion.
Knowledge organization contributes ontologies, controlled vocabularies, concept maps, and machine-readable relationships.
Qualitative methodology contributes coding, memo-writing, comparison, and category formation. Grounded-theory practices are relevant because they preserve provisional codes, analytic memos, emerging categories, and unresolved tensions across iterative cycles.8
Humanâcomputer interaction contributes common ground, external cognition, boundary objects, and handoff design.
Software engineering contributes checkpoints, event logs, schemas, provenance, branching, and version control.
Archival studies contribute appraisal and traceability: determining what can be removed, what must remain recoverable, and which record is authoritative.
Indigenous methodologies contribute relationality, situated authority, stewardship, and the recognition that some knowledge cannot responsibly be detached from the relationships that give it meaning.
Existing prompt-compression research addresses only part of this problem. LLMLingua uses a smaller model to identify less important prompt material and reports compression of up to twenty times on selected tasks. That result should not be treated as a universal safe ratio because acceptable compression depends on the task and on what must survive.9
Other methods encode long contexts into gist, memory, or compressed tokens. Their effectiveness may depend partly on the positions of compressed representations relative to source material and later queries.10
A particularly relevant direction decomposes context into smaller discourse units before compression. This closely matches the episode-forty-four principle: decompose first, compact second.
Agent systems also increasingly separate active context from persistent memory. OpenAI describes compaction as reducing context while preserving state needed for later turns. Anthropic combines context editing with external memory so active tool results can be removed while durable information remains available.1112
This creates three questions.
Compaction asks: what must remain active for the next cycle?
Memory asks: what may become relevant later?
The archive asks: what must remain recoverable even when the current model does not select it?
Abstractive summaries can sound coherent while changing entities, quantities, chronology, causality, negation, or confidence. Lexical-overlap metrics alone cannot reliably detect those failures.1314
The working definition is:
Epistemic compaction is the deliberate transformation of an exploratory research trace into a reduced, structured, provenance-bearing state intended to support reliable interpretation and continuation across humanâmodel work cycles.
The protocol begins with exploration. The researcher and model retrieve sources, decompose questions, propose interpretations, invent provisional terms, use tools, and follow unexpected connections. The complete event stream remains in an external archive.
Important events are marked as claims, sources, decisions, definitions, contradictions, assumptions, questions, procedures, relationships, restrictions, or next actions.
Compaction begins at an epistemic boundary, not merely at a token limit. A boundary may occur when a question has been provisionally answered, a decision has been made, a search phase has ended, a contradiction has been isolated, the inquiry changes discipline, or another person or agent must take over.
The readiness test asks:
Have the important claims been identified?
Is their evidence recorded?
Are decisions and rationales explicit?
Are uncertainties attached to specific claims?
Is there an executable next action?
The next phase is decomposition. The session becomes atomic records containing identifiers, statements, status, confidence, sources, contradictions, dependencies, session provenance, and reuse restrictions. This prevents fluent prose from concealing differences between evidence, inference, speculation, and decision.
Every boundary produces three representations.
The human abstract explains what was learned.
The machine handoff records claims, decisions, constraints, open questions, retrieval pointers, and next actions.
The terminology ledger records terms, definitions, provenance, aliases, examples, relationships, revisions, and governance conditions.
These representations should remain separate but linked. Human readability, operational fidelity, and terminological stability are different objectives.
Validation then functions as a semantic checksum.
Questions are generated from the original session. They test central claims, uncertainty, rejected alternatives, chronology, contradictions, source relations, and restrictions.
Another model answers using only the compacted checkpoint. Its answers are compared with the original record to identify omissions, unsupported additions, altered confidence, and broken relationships.
The next cycle begins with the machine handoff, relevant terminology, retrieval pointers, the active question, and a statement of what the model must not assume. Before continuing, the model restates the research state and identifies what it cannot reconstruct.
Earlier checkpoints are never overwritten. Each cycle creates a new version recording what was added, removed, generalized, renamed, contradicted, or downgraded in confidence.
The experiment should compare five conditions:
The full transcript.
An ordinary rolling summary.
Selected original passages.
A structured epistemic capsule.
Automated token compression.
Each condition receives the same continuation tasks. Evaluation records token reduction, claim recall, provenance coverage, task success, contradiction rate, unsupported claims, preservation of dissent and uncertainty, human resumption time, archive retrievals, repair operations, terminology stability, and stewardship violations.
Maximum compression is not the objective. A capsule that saves tokens but destroys provenance has failed. One that preserves facts while erasing uncertainty has also failed. A machine-efficient state that humans cannot audit may be unacceptable.
The inquiry can then ask which meanings disappear first: propositions, implications, chronology, causal relations, confidence, or obligations.
It can test whether discourse decomposition preserves continuity better than direct summarization, and whether preserving unresolved questions is sometimes more valuable than preserving provisional answers.
It can examine when a phrase becomes a stable term and whether its âdecompression instructionsâ can be represented through definitions, examples, provenance, relations, exclusions, and revision history.
It can ask whether humans and models need different but linked checkpoints, and how a capsule can preserve who knew what, what was implied, what was contested, and what participants treated as common ground.
For Indigenous and community-held knowledge, it must ask what should not be compacted, generalized, detached from place, or transferred between systems; who may authorize a compressed representation; and how refusal, restricted circulation, ceremony, contextual integrity, and relational accountability should appear in a machine handoff.
GitHub can contain canonical schemas, protocol versions, experiment code, and checkpoint history. Notion can contain literature maps, concept pages, glossaries, and article drafts. Slack can support discussion, review, and disagreement. An external archive can preserve transcripts, sources, and restricted records. The active model context should contain only the current capsule and retrieved evidence.
The proposed cycle is:
Explore. Detect a boundary. Decompose. Generate three representations. Validate. Resume. Recompact.
The working term is epistemic compaction.
The central claim is that a research session should become a structured, provenance-bearing state rather than only a prose summary.
The next cycle is to build the checkpoint schema and apply it retrospectively to episode forty-four, comparing it with an ordinary rolling-summary baseline.
â
Shows the iterative protocol: explore, detect a boundary, decompose, generate representations, validate, resume, and recompact.
Shows the three synchronized outputs:
- Human abstract
- Machine handoff
- Terminology ledger
It also situates them within the archive and governance layers.
Shows how five context strategies receive identical continuation tasks and are evaluated for meaning fidelity, provenance, continuation quality, cost, and human effort.
miadisabelleauthorOct 15, 2026, 09:23 AM