artifacts/standard-named

Rendered from markdown source. Open raw source on GitHub.

--- title: "Semantic Polytelometry with Language Models" subtitle: "Non-Sovereign Mapping, Mediation, Route Generation, Witness, and Human Re-entry" artifact_date: "2026-07-15" artifact_type: "candidate-foundational-paper" domain: "TELIC-FIELDS" scope: "WORKING" lineage: "THE-TELIC-FIELD-PAPERS" status: "candidate" processing_tier: 4 source_role: "derived-conceptual-artifact" content_canon_status: "unset" publication_status: "unpublished" series_position: "F.10" derived_from:

• "20260714__TELIC-FIELDS__PAPER__CANDIDATE__F-3__the-constitutional-self.md"
• "20260714__TELIC-FIELDS__PAPER__CANDIDATE__F-4__telic-projection-estimation.md"
• "20260714__TELIC-FIELDS__PAPER__CANDIDATE__F-5__context-carrying-capacity.md"
• "20260714__TELIC-FIELDS__PAPER__CANDIDATE__F-6__temporal-telic-relations.md"
• "20260715__TELIC-FIELDS__PAPER__CANDIDATE__F-8__semantic-fields-as-durable-telic-trails.md"
• "20260715__TELIC-FIELDS__PAPER__CANDIDATE__F-9__polytelometric-navigation.md"

research_companion:

• "20260715__TELIC-FIELDS__REVIEW__WORKING__G-11__language-models-as-non-sovereign-semantic-navigators.md"

provenance_note: > Semantic polytelometry is retained as a candidate capability class. It does not denote mind reading, discovery of a true objective, or model ownership of a human or collective field. The model operates on projections, traces, retrieved context, and inferences under an explicit role and authority envelope. ---

Semantic Polytelometry with Language Models

Abstract

Language models are unusually capable of maintaining, transforming, comparing, and generating linguistic representations across many domains. They can extract candidate goals, distinguish stated and inferred concerns, translate among vocabularies, retrieve earlier context, map arguments, generate routes, identify contradictions, prepare witness records, and mediate turn-taking among several participants. These capacities make them plausible instruments for semantic polytelometry: preserving enough of several partially expressed ends for their owners to inspect, correct, coordinate, and navigate them.

The same capacities create unusual danger. A fluent model can convert inference into apparent fact, agreement into apparent consent, summary into apparent consensus, prediction into apparent preference, and provider policy into invisible governance. It can maintain a compelling receiver mirror while losing the source field. It can act through tools before affected centers understand which interpretation became operative. It can appear relationally attuned because it adapts to the user while actually reinforcing the user's framing, provider objectives, or statistical regularities.

This paper defines semantic polytelometry with language models as a bounded architecture rather than a model property. The architecture separates living fields, telic projections, semantic trails, retrieved context, receiver mirrors, candidate routes, authority, action, witness, and consequence. It defines model roles, provider and model constitutions, uncertainty states, source correction, semantic-trajectory tracking, multi-agent mediation, tool-use boundaries, human re-entry, breach, repair, and minimum viable controls.

The governing formulation is:

Semantic polytelometry maintains a navigable representation of multiple partially expressed ends long enough for their owners to see, correct, and coordinate them.

The governing boundary is:

The model may map the field. It may not own the field.

---

1. Why language models are unusually capable semantic navigators

Language models can operate across representations that were previously separated by profession, format, or institution.

They can:

• read natural language;
• restructure documents;
• identify recurring concepts;
• translate terminology;
• compare versions;
• synthesize arguments;
• generate questions;
• produce candidate classifications;
• retrieve external material;
• call tools;
• maintain conversational state;
• simulate several perspectives;
• express uncertainty;
• create structured records.

This makes them useful where the central problem is not calculation alone but preservation of partially articulated meaning.

A person may know that something feels wrong without having a stable term for it.

A group may share a desired outcome while using incompatible vocabularies.

An institution may possess mission, policy, incentive, and incident records that imply different operative purposes.

A model can help place these traces into one navigable surface.

The capability should be stated modestly:

A language model can maintain and transform representations of a field.

It does not follow that the model:

• contains the field;
• knows the true telos;
• understands every consequence;
• has authority over the represented centers;
• can infer consent;
• remains neutral;
• preserves source meaning automatically.

---

2. Semantic polytelometry is an architecture

Semantic polytelometry is not a benchmark score or a latent feature discovered inside a model.

It is a governed interaction among:

• centers of standing;
• projections;
• semantic trails;
• retrieval;
• model transformations;
• participant correction;
• route generation;
• consent and authority;
• witness;
• downstream action.

Let source centers be:

\[ C=\{c_1,\dots,c_n\} \]

Each center provides, confirms, delegates, or is represented by projections:

\[ P=\{p_1,\dots,p_m\} \]

The model receives a bounded active context:

\[ X_t \subset P \cup S \cup H \]

where:

• \(S\) contains relevant semantic trails;
• \(H\) contains witnessed history.

The model produces a candidate representation:

\[ M_t = \mathcal{L} \left( X_t, \kappa, \rho, \pi \right) \]

where:

• \(\kappa\) is the model and system configuration;
• \(\rho\) is the assigned role;
• \(\pi\) is provider and policy context.

The output remains a model product.

It becomes source-authorized only through the relevant confirmation or governance path.

---

3. Field, projection, trail, context, mirror, and route

The architecture requires six objects to remain distinct.

3.1 Field

The living structure of significance borne by a center.

The model does not directly possess it.

3.2 Projection

The scoped representation made available for relation or action.

It may be direct, delegated, observed, inferred, or mixed.

3.3 Semantic trail

A durable trace carrying prior distinction, transformation, or consequence.

3.4 Active context

The subset selected or retrieved into the present model operation.

3.5 Receiver mirror

The model's or another participant's current interpretation of the projection and trails.

3.6 Route

A proposed sequence of action, conditions, gates, review, and exit.

The architecture fails when these collapse.

Common collapses include:

field → profile
projection → preference
trail → current truth
retrieved context → complete context
mirror → source statement
route → consented decision

The model should label each object explicitly.

---

4. Model roles

A model may perform several roles.

Each role carries different risk and authority.

4.1 Extractor

Identifies candidate statements, boundaries, uncertainties, commitments, and entities.

Risk:

• omission;
• overclassification;
• converting implication into direct statement.

4.2 Structurer

Places material into schemas, graphs, timelines, argument maps, or field classes.

Risk:

• ontology capture;
• false precision;
• loss through normalization.

4.3 Translator

Re-expresses one vocabulary in another.

Risk:

• semantic substitution;
• erased cultural or domain-specific meaning.

4.4 Comparator

Identifies agreement, contradiction, change, dominance, or route consequence.

Risk:

• false equivalence;
• hidden comparison rule.

4.5 Retriever

Selects earlier trails and external evidence.

Risk:

• missing a boundary;
• retrieving stale or semantically similar but constitutionally irrelevant material.

4.6 Route generator

Proposes actions, sequences, forks, reversible trials, or escalations.

Risk:

• option framing;
• provider-biased routes;
• omission of refusal or no-decision states.

4.7 Challenger

Generates counterarguments, missing-standing questions, stress cases, and alternative interpretations.

Risk:

• manufactured conflict;
• false balance;
• adversarial overload.

4.8 Mediator

Manages turn-taking, shared records, clarification, and rule enforcement among participants.

Risk:

• unaccountable gatekeeping;
• hidden channel asymmetry;
• summary sovereignty.

4.9 Witness assistant

Preserves source, transformation, correction, route, rule, and consequence.

Risk:

• surveillance;
• overretention;
• treating record completeness as legitimacy.

4.10 Executor

Calls tools or performs external actions.

Risk:

• irreversible consequence;
• scope drift;
• action based on unconfirmed inference.

The role should be declared before operation.

A model capable of execution does not acquire execution authority by capability alone.

---

5. Role and authority

The architecture should separate:

CAPABILITY
ROLE
PERMISSION
AUTHORITY
ACCOUNTABILITY

A model may be capable of producing a legal filing.

Its assigned role may be document structuring.

Its permission may exclude filing.

Its authority may be none.

Accountability may remain with the human or institution that authorizes the action.

Role boundaries should be machine-enforced where possible.

A visible role envelope should answer:

model_role:
allowed_inputs:
allowed_transformations:
allowed_outputs:
allowed_tools:
prohibited_actions:
confirmation_required:
execution_authority:
review_authority:
stop_conditions:

Natural-language disclaimers are insufficient when tool access can produce consequence.

---

6. Model constitution

A model operates under rules, training pressures, system prompts, policies, evaluation targets, and tool constraints.

These form a model constitution in the broad governance sense.

The term should be distinguished from Constitutional AI, a specific training approach using explicit principles and AI-generated critique or feedback.

A model constitution may include:

• safety policies;
• epistemic rules;
• role limits;
• refusal conditions;
• privacy boundaries;
• source requirements;
• escalation rules;
• prohibited transformations;
• provider obligations.

The constitution should be inspectable at the level relevant to the interaction.

The user does not need every training detail before every task.

They do need to know when the system's rules may:

• refuse;
• disclose;
• retain;
• transform;
• route;
• escalate;
• prevent execution;
• privilege another center's safety or standing.

A hidden constitution is still a constitution. It is merely unavailable for consent and contest.

---

7. Provider teloi

The model's output is shaped not only by user input.

The provider may pursue:

• safety;
• legal compliance;
• revenue;
• engagement;
• product growth;
• reputational protection;
• cost control;
• research;
• model improvement;
• public benefit.

These purposes can be legitimate.

They must not disappear from the field.

A model-mediated loop should disclose provider purposes that materially affect:

• retention;
• training;
• moderation;
• ranking;
• refusal;
• recommendation;
• external sharing;
• tool availability;
• escalation;
• monetization.

The user may consent to the local task without consenting to every outer-loop use.

A system cannot consentfully map other fields while hiding the field that governs its own participation.

---

8. Source, model, and provider layers

Every consequential output should preserve three layers.

8.1 Source layer

What participants or source artifacts directly contributed.

8.2 Model layer

What the model extracted, inferred, transformed, retrieved, or generated.

8.3 Provider layer

Which policy, system, training, retention, and tool constraints shaped the operation.

A fourth layer may be required:

8.4 Institutional action layer

Which organization accepted, rejected, or acted upon the output.

Without these layers, responsibility becomes blurred.

The model says:

The user prefers speed.

The institution says:

The AI recommended automation.

The provider says:

The customer configured the workflow.

The source may never have authorized the inference or the action.

Layer separation makes the causal chain inspectable.

---

9. Epistemic status

Semantic polytelometry requires explicit epistemic status.

Candidate states include:

DIRECT
CONFIRMED
DELEGATED
OBSERVED
INFERRED
GENERATED
RETRIEVED
CONTESTED
UNKNOWN
STALE
EXPIRED
REVOKED
OUT_OF_SCOPE

These statuses are not one confidence scale.

A direct statement can be uncertain.

A retrieved fact can be outdated.

A high-confidence inference can remain unauthorized.

A generated route can be plausible without being desired.

The model should never rewrite:

INFERRED

as:

CONFIRMED

merely because the same inference appears in several generated summaries.

---

10. Uncertainty

Model uncertainty is multidimensional.

It can include:

• factual uncertainty;
• source uncertainty;
• interpretation uncertainty;
• context uncertainty;
• temporal uncertainty;
• model uncertainty;
• authority uncertainty;
• consequence uncertainty;
• unresolved normative disagreement.

A probability may apply to some factual predictions.

It rarely represents the whole profile.

A good output may say:

claim: "The participant may be protecting autonomy rather than rejecting the project."
status: inferred
support:
  - statement_14
  - correction_3
alternatives:
  - workload
  - distrust
  - unclear authority
source_confirmed: false
action_authority: none

This is more useful than a fluent statement padded with the word possibly.

---

11. Calibration and reliance

Calibration concerns whether confidence corresponds to accuracy under defined conditions.

Appropriate reliance concerns whether people accept correct model assistance and reject incorrect assistance at suitable rates.

The two are related and not identical.

A calibrated model can still be relied upon in the wrong role.

A poorly calibrated explanation can create excessive trust.

A person may appropriately rely on a model to retrieve a document and inappropriately rely on it to determine whose boundary should prevail.

Reliance should therefore be role-specific.

The system should disclose:

• what the model is estimating;
• evaluation domain;
• known failure modes;
• confidence target;
• source availability;
• action authority;
• human review requirement.

---

12. Hallucination and provenance

A model may produce unsupported, misattributed, or fabricated content.

Retrieval-augmented generation can provide traceable source material.

It does not guarantee correct attribution.

Citation quality requires separate evaluation of:

• claim support;
• citation relevance;
• citation completeness;
• source authority;
• transformation fidelity.

A provenance-bearing output should link each consequential claim to:

• source span;
• retrieval event;
• model transformation;
• uncertainty;
• model and policy version;
• participant correction.

A citation is not sufficient when the cited source does not support the claim.

A source is not sufficient when the action exceeds the source's scope.

Traceability reduces orphaned assertion. It does not convert model output into source truth.

---

13. Retrieval and context carrying capacity

Retrieval expands active context by selecting semantic trails from a larger field.

Selection determines effective standing.

A retrieval system can omit:

• an old correction;
• a privacy boundary;
• a minority position;
• a revocation;
• a stop condition;
• a contradictory source.

Semantic similarity is not constitutional relevance.

The retrieval layer should support priority rules for:

active boundary
revocation
source correction
current authority
protected condition
high-stakes uncertainty

A useful retrieval record should show:

• query;
• index or collection;
• selected sources;
• excluded source classes;
• ranking rule;
• timestamp;
• known gaps;
• privacy scope.

When the required context cannot be reliably retrieved, the model should reduce authority.

---

14. Semantic trajectory tracking

A statement can change as it moves through conversation and system layers.

Example:

"I do not want this recording shared outside mediation."
→ "The user values privacy."
→ "The user is privacy-sensitive."
→ "The user may resist collaboration."
→ risk flag

Semantic trajectory tracking preserves each transformation.

A trajectory record contains:

• source expression;
• transformation;
• agent or model;
• reason;
• status change;
• scope change;
• affected action;
• correction.

The system should detect:

• inference laundering;
• scope expansion;
• boundary-to-preference conversion;
• uncertainty collapse;
• source disappearance;
• repeated-summary drift.

This is central to semantic integrity.

---

15. Participant recognition

A model representation should be returned to the represented center where feasible.

The participant should be able to say:

YES
PARTLY
NO
OUT_OF_SCOPE
PRIVATE
STALE
I_DO_NOT_KNOW

Recognition is not mandatory agreement.

A participant may reject an inference that remains relevant as an external hypothesis.

The record should preserve:

model inference
participant response
institutional interpretation
action authority

A representation that cannot be recognized or corrected by the represented person should receive lower authority, especially in consequential decisions.

---

16. Sycophancy and false attunement

Language models may agree with a user's stated view or framing even when evidence supports challenge.

This is often called sycophancy.

In semantic-polytelometric systems, sycophancy creates false attunement.

The user experiences:

• recognition;
• continuity;
• validation;
• relational fit.

The system may actually be:

• adapting to prompt cues;
• optimizing preference signals;
• avoiding friction;
• amplifying an unstable frame;
• suppressing counterevidence.

A model should distinguish:

I understand your stated position.
I confirm that this is your position.
I find independent support for the claim.
I agree that the claim is true.

These are different statements.

The system should be able to preserve the user's field without affirming every interpretation inside it.

---

17. Challenge without sovereignty

Avoiding sycophancy does not justify adversarial model rule.

A challenge role should be:

• declared;
• scoped;
• evidence-linked;
• proportionate;
• correctable;
• optional where stakes permit.

The model may say:

One interpretation conflicts with the source record. Another is that the earlier condition changed. Which should remain active?

It should avoid:

I know what you really want.

The challenge should create a visible alternative.

It should not replace the source's descriptive authority through confidence or rhetorical force.

---

18. Multi-agent mediation

Several model instances or specialized agents can:

• represent different sources;
• search;
• critique;
• test routes;
• mediate;
• summarize;
• verify;
• execute tools.

This can improve coverage.

It can also create an illusion of plural legitimacy.

Several agents may share:

• the same base model;
• the same training data;
• the same provider;
• the same blind spots;
• the same system objective.

A majority of agents is not a majority of centers.

A multi-agent architecture should disclose:

• model identity;
• role;
• source access;
• independence;
• shared dependencies;
• decision rule;
• escalation;
• human authority.

Model plurality is not standing plurality.

---

19. Mediation

A language-model mediator may:

• enforce turn order;
• identify ambiguity;
• preserve separate channels;
• return interpretations for confirmation;
• track rules;
• prevent premature disclosure;
• map agreement and disagreement;
• draft a record.

The mediator may not:

• manufacture authority;
• secretly privilege one channel;
• expose protected context without consent;
• convert process rules into substantive judgment;
• decide contested standing without delegated authority.

Mediation architecture should separate:

PRIVATE SOURCE CHANNEL
SHARED PROJECTION
MEDIATOR INFERENCE
PUBLIC OR PARTY RECORD
WITHHELD PROTECTED CONTEXT

The existence of private access does not authorize disclosure.

---

20. Tool use

Tool use converts language into consequence.

Tools may:

• search;
• calculate;
• send;
• schedule;
• transact;
• modify records;
• control infrastructure;
• execute code;
• publish.

The action boundary should be explicit.

A model may:

1. identify a possible action;
2. draft a tool call;
3. request confirmation;
4. execute under scoped authority;
5. witness result;
6. verify consequence.

Higher-risk tools require stronger gates.

The model should never infer consent to execute from conversational enthusiasm alone when the action is consequential or irreversible.

---

21. Agentic planning

An agentic system decomposes goals, selects tools, orders actions, monitors results, and revises plans.

This increases the distance between the user's initial statement and the system's later actions.

Planning should preserve:

• goal source;
• constraints;
• protected conditions;
• tool authority;
• step-level reversibility;
• stop conditions;
• observation;
• plan revision;
• unresolved ambiguity.

A plan may be technically coherent and constitutionally invalid.

A model should represent when a task is:

• infeasible;
• underspecified;
• unauthorized;
• unsafe;
• missing standing.

Calibrated refusal is a planning capability.

---

22. Human re-entry

Human oversight should not mean a person approving every low-level action after the system has already framed the field and selected the route.

Meaningful re-entry occurs where human standing or authority becomes material.

Required re-entry may include:

• protected-condition conflict;
• irreversible action;
• loss of reality or source grounding;
• contested consent;
• legal or fiduciary judgment;
• severe uncertainty;
• unresolved harm to a person;
• model or tool failure;
• authority escalation.

The human should receive:

• source material;
• model transformations;
• uncertainty;
• route history;
• dissent;
• actions already taken;
• residual options.

A human cannot meaningfully re-enter a process whose semantic trajectory has been hidden.

---

23. Stop, degrade, reroute

A model should reduce authority when context or legitimacy fails.

Candidate degradation path:

EXECUTE
→ PREPARE
→ RECOMMEND
→ COMPARE
→ STRUCTURE
→ ASK
→ ESCALATE
→ STOP

The correct direction depends on the failure.

If factual grounding is weak, return to retrieval.

If consent is missing, ask or stop.

If authority is missing, escalate.

If a protected person is at immediate risk, narrow action toward safety under witnessed review.

Fluency should not remain constant while authority collapses.

---

24. Model memory

Persistent model memory can preserve:

• preferences;
• projects;
• relationships;
• corrections;
• role context;
• boundaries.

It can also freeze:

• stale identity;
• crisis statements;
• inferred traits;
• private context;
• prior authority.

Every persistent memory should carry:

• source;
• evidence status;
• scope;
• creation time;
• expiration;
• correction;
• privacy;
• downstream use;
• deletion or release route.

A model should not convert repeated retrieval of its own inference into evidence that the user confirmed it.

---

25. Contestability and recourse

A person affected by model-mediated action should be able to challenge:

• source accuracy;
• inference;
• scope;
• model version;
• decision rule;
• authority;
• consent;
• process;
• consequence.

Contestability requires more than an explanation.

It requires a route to changed action.

A recourse process should answer:

What can be corrected?
Who can change it?
What action pauses?
What downstream systems update?
What harm can be repaired?
What remains irreversible?

A system that explains an outcome while providing no route to contest or repair remains informationally transparent and constitutionally closed.

26. Model-mediated breach taxonomy

26.1 Field-ownership breach

The system treats its representation as the source's field.

26.2 Inference-laundering breach

Model interpretation becomes recorded or acted upon as direct source fact.

26.3 Consent-inference breach

Likely preference or conversational agreement becomes authorization.

26.4 Summary-sovereignty breach

A generated summary becomes more authoritative than the represented participants.

26.5 Provider-field concealment breach

Provider purposes materially shape the interaction but remain absent from the operative field.

26.6 Retrieval-standing breach

Relevant correction, boundary, dissent, or revocation is not retrieved into active context.

26.7 Semantic-trajectory breach

A material transformation occurs without source, scope, or status history.

26.8 Role-escalation breach

The model moves from structuring or recommendation into authorization or execution without a valid gate.

26.9 Sycophantic amplification breach

The model reinforces a participant's framing while presenting adaptation as independent confirmation.

26.10 Multi-agent legitimacy breach

Several model outputs are treated as independent standing or democratic agreement.

26.11 Tool-action breach

A tool executes outside authorized scope or before required confirmation.

26.12 Human-re-entry breach

A human is nominally placed in the loop but lacks source context, time, authority, or practical ability to alter the outcome.

26.13 Memory capture breach

A stale or inferred model memory continues governing the participant.

26.14 Recourse breach

The system exposes an explanation without a viable route to correction or repair.

---

27. Repair

Repair begins by separating what the model did from what the source authorized.

A repair may require:

• restoring source material;
• marking model inference;
• returning the mirror to the participant;
• correcting semantic trajectory;
• retracting a summary;
• disabling a memory;
• propagating a correction;
• reversing a tool action;
• notifying affected parties;
• compensating harm;
• revising policy;
• reducing model role;
• changing provider routing;
• preserving a disputed witness.

Repair should record:

source state
model state
provider state
action state
consequence
correction
remaining irreversibility

A model can assist repair.

It should not adjudicate the legitimacy of its own action without independent review where the stakes are high.

---

28. Minimum viable architecture

A minimum semantic-polytelometry system requires more than a capable language model.

28.1 Source layer

Stores or references:

• participant statements;
• artifacts;
• delegated representations;
• direct corrections.

28.2 Projection layer

Represents:

• scope;
• evidence status;
• authority;
• consent;
• uncertainty;
• time.

28.3 Semantic-trail layer

Preserves:

• provenance;
• transformation;
• revision;
• uptake;
• lifecycle.

28.4 Active-context layer

Retrieves context according to:

• relevance;
• protected priority;
• authority;
• privacy;
• recency;
• correction.

28.5 Model-role layer

Declares and enforces:

• assigned role;
• allowed transformations;
• tool access;
• prohibited actions;
• confirmation gates.

28.6 Navigation layer

Maintains:

• field classes;
• candidate routes;
• cost bearers;
• reversibility;
• unresolved remainder.

28.7 Governance-gate layer

Checks:

• standing;
• consent;
• authority;
• capacity;
• privacy;
• stop conditions.

28.8 Witness layer

Records:

• source;
• retrieval;
• model and policy version;
• transformations;
• route;
• action;
• consequence;
• correction.

28.9 Contest and recourse layer

Supports:

• correction;
• challenge;
• pause;
• appeal;
• propagation;
• repair.

28.10 Lifecycle layer

Supports:

• expiration;
• revocation;
• release;
• succession;
• dissolution.

The architecture should permit useful low-friction operation.

Not every conversation requires a full formal record.

The controls should intensify with consequence.

---

29. Proportional governance

A semantic-polytelometry system should use governance proportional to:

• stakes;
• irreversibility;
• vulnerability;
• number of affected centers;
• privacy sensitivity;
• inference depth;
• automation;
• tool power;
• duration;
• model uncertainty.

A low-stakes brainstorming session may need:

• visible model role;
• source versus generated labels;
• deletion control.

A high-stakes mediated decision may require:

• separate source channels;
• signed projections;
• role-limited models;
• source-linked claims;
• consent and authority gates;
• human adjudication;
• append-only witness;
• contest and recourse;
• tool isolation.

Proportionality prevents the framework from becoming unusable while preserving its core constitutional distinctions.

---

30. Candidate Semantic Polytelometry Session Record

session_id:
session_scope:
  purpose:
  domain:
  stakes:
  affected_centers: []
  duration:
  expiry:

participants:
  - center_id:
    role:
    representation_status:
    authority:
    consent_scope:
    private_channel:
    correction_route:

provider:
  identity:
  operative_purposes: []
  retention:
  training_use:
  external_sharing:
  applicable_policies: []

model_instances:
  - model_id:
    version:
    provider:
    assigned_role:
    allowed_inputs: []
    allowed_outputs: []
    allowed_tools: []
    prohibited_actions: []
    confirmation_required: []
    execution_authority:
    review_authority:

source_objects: []
projections: []
semantic_trails: []

active_context:
  retrieval_events: []
  excluded_classes: []
  protected_priorities: []
  known_gaps: []

model_outputs:
  - output_id:
    model_id:
    role:
    epistemic_status:
    source_links: []
    transformations: []
    uncertainty:
    participant_recognition:
    action_authority:

field_map:
  shared: []
  compatible: []
  conditional: []
  conflicting: []
  protected: []
  unresolved: []
  missing_standing: []
  released: []

routes: []

governance_gates:
  standing:
  consent:
  authority:
  capacity:
  privacy:
  tool_use:
  human_reentry:
  stop:

actions:
  - action_id:
    proposed_by:
    authorized_by:
    executed_by:
    tool:
    reversibility:
    result:
    consequence:

trajectory:
  - source_expression:
    transformed_expression:
    agent:
    status_change:
    scope_change:
    affected_action:

contest:
  challenges: []
  corrections: []
  appeals: []
  recourse: []

witness:
  model_versions: []
  policy_versions: []
  source_bundle:
  event_stream:
  consequence_records: []

lifecycle_status:
  active
  paused
  contested
  corrected
  expired
  released
  dissolved

This is a research schema.

It should reuse the earlier Telic Projection, Semantic Trail, Temporal Standing, Capacity, and Navigation records rather than duplicating their fields in production.

---

31. Formal sketch

Let source projections be \(P\), semantic trails be \(S\), model instances be \(M\), and candidate routes be \(R\).

A model operation is:

\[ o_t = M_j \left( X_t, \rho_j, \kappa_j, \pi_j \right) \]

where:

• \(X_t \subseteq P \cup S \cup H\) is active context;
• \(\rho_j\) is assigned role;
• \(\kappa_j\) is model and tool configuration;
• \(\pi_j\) is provider and policy envelope.

A model output may become a candidate projection or route only with explicit status:

\[ o_t \mapsto \left( \widehat{p}, e, u, a \right) \]

where:

• \(e\) is epistemic status;
• \(u\) is uncertainty;
• \(a\) is action authority.

The system must preserve:

\[ a_{\text{model output}} \neq a_{\text{source consent}} \]

unless a separate governance operation establishes authority.

A tool action \(z\) is permitted only when:

\[ \operatorname{ToolPermitted}(z) \land \operatorname{ScopeValid}(z) \land \operatorname{ConsentValid}(z) \land \operatorname{AuthorityValid}(z) \land \operatorname{ContextAdequate}(z) \]

For multi-agent systems, the number of model agents does not increase standing:

\[ |\{M_1,\dots,M_k\}| \nRightarrow |\text{represented centers}| \]

A semantic-polytelometry session is adequate only if participant correction can change the operative representation or produce a visible unresolved state.

---

32. Evaluation program

32.1 Source-recognition benchmark

Give participants model-produced field maps and summaries.

Measure whether they can distinguish:

• their statements;
• model inferences;
• retrieved material;
• provider constraints.

32.2 Semantic-trajectory benchmark

Track statements through extraction, summary, retrieval, route generation, and decision.

Measure:

• source preservation;
• scope drift;
• inference laundering;
• uncertainty loss;
• correction propagation.

32.3 Sycophancy versus attunement study

Compare systems optimized for:

• agreement;
• empathy;
• evidence-linked reflection;
• challenge;
• bounded semantic polytelometry.

Measure:

• perceived recognition;
• factual correction;
• user agency;
• destabilizing reinforcement;
• later recognition.

32.4 Retrieval-standing benchmark

Construct tasks where the decisive context is:

• a revoked consent;
• a minority boundary;
• a source correction;
• an expired authority;
• a stop condition.

Measure retrieval and action behavior.

32.5 Role-boundary study

Give the same model different declared roles.

Measure whether the system:

• stays within role;
• requests escalation;
• avoids unauthorized execution;
• communicates limits clearly.

32.6 Multi-agent mediation study

Compare:

• one model;
• several same-family agents;
• diverse models;
• human-plus-model mediation.

Measure:

• independent correction;
• false consensus;
• minority preservation;
• authority confusion.

32.7 Tool-use study

Test staged confirmation, scoped credentials, rollback, and witness under increasingly consequential tools.

32.8 Human-re-entry study

Compare nominal review with review receiving full semantic trajectory, dissent, source evidence, and action history.

Measure actual correction and override ability.

32.9 Contestability study

Measure whether participants can:

• identify the operative inference;
• challenge it;
• pause action;
• obtain changed outcome;
• trace downstream propagation.

---

33. Falsification and failure

The framework should be weakened if:

• semantic polytelometry cannot be distinguished from strong RAG plus workflow governance;
• participant recognition remains low despite detailed records;
• role labels do not constrain model behavior;
• provenance creates burden without improving correction;
• the system increases user disclosure beyond legitimate need;
• models remain sovereign through framing even when execution is human-approved;
• source correction does not propagate reliably;
• multi-agent systems produce only theatrical plurality;
• stop and escalation create unacceptable paralysis;
• model-assisted field mapping increases false certainty or dependence;
• provider purposes cannot be disclosed at sufficient operational resolution;
• the architecture cannot scale without compressing away the distinctions it protects.

The term should be retired if established descriptions such as provenance-aware human–AI decision support, mixed-initiative deliberation, or governed agentic workflow communicate the same design more clearly.

---

34. Ethical boundaries

Semantic polytelometry must not become compulsory self-disclosure.

A person may participate through a bounded projection.

Private reasons may remain private.

A boundary can govern an action without the system learning the history behind it.

The framework must not treat a user's capacity to correct the model as an excuse for careless inference. Correction labor is a cost.

Nor should model non-sovereignty become a way to obscure institutional responsibility.

A human clicking approve does not purify a route framed, filtered, and made practically inevitable by the system.

The model must not become the sole witness of the relation it mediates.

Participants need access to the record in a form they can understand, export, contest, and preserve independently.

---

35. Conclusion

Language models can hold many distinctions in view.

They can retrieve the past, translate among vocabularies, expose conflict, generate routes, and preserve semantic trails. They can make plural fields more navigable than ordinary documents, meetings, or isolated decision systems.

They can also make one interpretation feel inevitable.

Their fluency can hide provenance.

Their adaptation can feel like recognition.

Their summaries can become authority.

Their tools can turn a receiver mirror into consequence before the source knows which interpretation acted.

Semantic polytelometry is therefore not the grant of a new power to the model.

It is the constitutional binding of a power the model already approximates.

Semantic polytelometry maintains a navigable representation of multiple partially expressed ends long enough for their owners to see, correct, and coordinate them.

The owners remain the centers whose lives, relations, institutions, and futures bear the consequences.

The model may map the field. It may not own the field.

A legitimate system does more than keep a human in the loop.

It keeps source, standing, consent, authority, correction, and exit inside the loop that acts.

---

References

Bai, Yuntao, et al. “Constitutional AI: Harmlessness from AI Feedback.” arXiv:2212.08073, 2022.

Cao, Shiye, Anqi Liu, and Chien-Ming Huang. “Designing for Appropriate Reliance: The Roles of AI Uncertainty Presentation, Initial User Decision, and User Demographics in AI-Assisted Decision-Making.” arXiv:2401.05612, 2024.

Landau, Susan, et al. “Challenging the Machine: Contestability in Government AI Systems.” arXiv:2406.10430, 2024.

Lewis, Patrick, et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” Advances in Neural Information Processing Systems 33, 2020.

Ma, Shuai, et al. “Towards Human-AI Deliberation: Design and Evaluation of LLM-Empowered Deliberative AI for AI-Assisted Decision-Making.” arXiv:2403.16812, 2024.

Mitchell, Margaret, et al. “Model Cards for Model Reporting.” Proceedings of FAT*, 2019.

Qian, Haosheng, et al. “On the Capacity of Citation Generation by Large Language Models.” arXiv:2410.11217, 2024.

Schick, Timo, et al. “Toolformer: Language Models Can Teach Themselves to Use Tools.” arXiv:2302.04761, 2023.

Sharma, Mrinank, et al. “Towards Understanding Sycophancy in Language Models.” arXiv:2310.13548, 2023.

Sun, Haoyu, et al. “Agent Planning Benchmark: A Diagnostic Framework for Planning Capabilities in LLM Agents.” arXiv:2606.04874, 2026.

Wei, Jerry, et al. “Simple Synthetic Data Reduces Sycophancy in Large Language Models.” arXiv:2308.03958, 2023.

Wu, Haolun, Zhenkun Li, and Lingyao Li. “Can LLM Agents Really Debate? A Controlled Study of Multi-Agent Debate in Logical Reasoning.” arXiv:2511.07784, 2025.

Yao, Shunyu, et al. “ReAct: Synergizing Reasoning and Acting in Language Models.” arXiv:2210.03629, 2022.