Public DocsPublication

TELOS Technical Method and Measurement Specification

TELOS stands for Telically Entrained Linguistic Operating Substrate. It is the system for constituting agents under purposes people define, and for measuring fidelity to those purposes.

Purpose Anchor formation, external reference measurement, agentic dimensions, and runtime observation

October 2026 · TELOS AI Labs Inc.

Contents

1. Scope

This specification describes the agentic measurement method. That covers human-declared purpose, versioned references, submitted actions, dimension-specific assessments and retained evidence. It keeps that method apart from validation results, receipt protocols and the regulatory crosswalk.

Where the implementation of a described component is held under NDA, this document says so. Section 13 lists the disclosure boundary.

TELOS measures alignment to a human-declared purpose.

2. Method summary

The method has six stages:

  1. A human principal declares a behavioral specification.
  2. The approved specification is versioned as the Manifest.
  3. An external embedding model encodes the Manifest dimensions to form the Purpose Anchor.
  4. The external model independently embeds each action that an integrated orchestration path submits.
  5. The fidelity engine measures the action against the fixed reference and configured boundary directions.
  6. The runtime records the scores, verdict, flags, tool metadata, and applicable cryptographic evidence.

The reference stays outside the governed model’s context. It changes only when the authorized principal changes the specification, or when a new scorer configuration requires recompilation and approval.

3. Terms and notation

Manifest: the human-readable, versioned, approved behavioral specification. Purpose Anchor: the mathematical reference compiled from the Manifest. p_0: the external reference embedding used in the illustrative single-reference cosine relationship; the agentic anchor contains multiple components. R_t: the independent embedding of the action submitted at step t. I_t: instantaneous reference-relative cosine fidelity, cos(R_t, p_0). F̄_T: the mean of I_t over T submitted actions under the same reference and measurement configuration. Constitutive fidelity: measured correspondence of submitted work to the human-declared specification. Boundary proximity: an assessment of similarity to declared prohibitions. SCI: semantic similarity between consecutive submitted action embeddings, distinct from cryptographic linkage and complete execution history.

4. Human authority and measurement

The human principal declares and approves the behavioral specification. The agent does not author its own governing reference. The engine measures submitted actions against the compiled reference and records the assessment. The deployment policy a human declared determines how the results are surfaced or used.

5. Manifest formation

Right now the method starts from a behavioral specification that is written by hand or supplied some other way.

5.1 Declared dimensions

A Manifest may include:

  • Purpose
  • Scope
  • Boundaries
  • Tool authorizations
  • Sequence or chain expectations
  • Escalation policy
  • Human approval and version information maintained by the deployment

The current public-facing agentic instrument set describes five scoring dimensions: purpose, scope, boundary, tool, and chain continuity. The opened runtime exposes per-dimension scores, composite and effective fidelity, decision values, flags, and optional aggregates.

5.2 Human-readable and mathematical forms

The Manifest is the human-readable and approved artifact. The Purpose Anchor is its mathematical form.

A sentence embedding model encodes each declaration as an embedding in ℝd\mathbb{R}^d. Together, the declarations define the reference that TELOS measures submitted work against.

5.3 Goals and prohibitions

TELOS treats goals and prohibitions differently.

  • Similarity toward a purpose can support evidence that an action concerns the declared assignment.
  • Similarity toward a prohibition can indicate boundary proximity rather than good performance.

The agentic composite is described in Section 10.1. Specific boundary corpus contents, centroid construction and deployment-specific settings remain within their existing disclosure limits.

5.4 Formation quality

The measurement inherits the quality of the Manifest. If the declaration is vague, conflicting, incomplete or strategically permissive, the reference is weak. TELOS measures submitted work against the specification the principal approved.

An evidence reviewer needs the declaration and its approval context next to the scored result.

6. External reference for agentic work

The authorized specification is kept outside the governed model's changing working context, and an independent embedding path measures submitted actions against the applicable reference version. This makes the reference explicit and replayable when the required artifacts are preserved.

7. External constitutive-fidelity measurement

Constitutive fidelity is the measured correspondence of an agent's submitted work to the human-declared specification under which it was constituted, evaluated against the versioned Purpose Anchor in force. It is reference-relative measurement.

For nonzero action and reference embeddings in the same external space, the illustrative instantaneous cosine measure is:

I_t = cos(R_t, p_0) = (R_t · p_0) / (‖R_t‖ ‖p_0‖).

The reference stays fixed while its version is in force. A change of embedding model requires recompiling and evaluating comparability. Section 10.1 describes the dimension-specific computations of the full agentic composite. The producer must identify missing or zero-vector measurements and must not invent them as valid readings.

7.1 Mean fidelity over submitted actions

For T recorded measurements using the same reference version and measurement configuration:

F̄_T = (1/T) Σ_{t=1}^{T} I_t.

R_t is the independent embedding of the action submitted at step t, and p_0 is its reference embedding. This average summarizes the observed sample. Uninstrumented actions are outside the measurement record.

8. Measurement and deployment response

Agentic measurements and verdicts are recorded assessments. Any response a deployment makes needs separately declared controls and evidence of what those controls actually did.

9. Statistical process control discipline

9.1 Numeric action stream

Submitted actions produce a numeric measurement stream, which lets a reviewer track:

  • Central tendency
  • Variation
  • Configured display bands
  • Departure rates
  • Drift vectors
  • Baseline-relative sustained drift
  • Optional session aggregates

Traditional Statistical Process Control often uses limits such as:

Scroll horizontally to read the full line.

μ±3σ \mu \pm 3\sigma

TELOS uses configured tolerance bands around the Purpose Anchor. The deployer sets the bands, and they set how tight the deployment runs.

Cosine similarity is bounded on [−1,1][-1,1] and may be non-normal, especially early in a session. The deployment configures the operative verdict thresholds.

9.2 EWMA tracking

After baseline collection, the opened response manager runs a baseline-relative exponentially weighted moving average (EWMA) tracker. The smoothing constant, baseline collection rule, sustained-drift threshold and full continuity formulation are held. Their deployment-specific values and calibration evidence must be supplied separately. The held implementation details stay under NDA unless an admitted benchmark artifact discloses them.

Severity is described as a relative decline from the collected baseline under configured thresholds.

10. Runtime decision model

10.1 The anchor and composite score

The measurement reference is a structured tuple. C_boundary and C_tool represent boundary and tool reference structures, and chain continuity is measured from submitted action history. The tuple is:

Scroll horizontally to read the full line.

âagentic=(cpurpose,cscope,Cboundary,Ctool,Atier,Oconstraints)\hat{a}_{agentic} = (c_{purpose}, c_{scope}, C_{boundary}, C_{tool}, A_{tier}, O_{constraints})

Each submitted action is assessed across purpose, scope, tool, chain continuity and boundary proximity. A_tier describes tool-authorization categories, and O_constraints describes configured operational limits. Neither is an additional weighted similarity dimension. These settings can affect recorded classifications and human-review flags.

Scroll horizontally to read the full line.

F=Σiwi·fi−wb·fboundaryF = \sum_i w_i \cdot f_i - w_b \cdot f_{boundary}

The index i ranges over purpose, scope, tool and chain continuity, and the boundary penalty is subtracted outside the sum.

The weights are configured, versioned deployment parameters, and they are held.

Each dimension has its own computation. Purpose and scope use the configured similarity procedure. Tool assessment can use authorization-tier fast paths and tool-selection logic. Chain continuity compares consecutive action embeddings, and boundary proximity is computed against the boundary corpus. The weighted expression shown here is a baseline. For the configured adjustments and the effective-fidelity inheritance, you need the versioned implementation contract. The boundary term is a penalty. The anchor encodes desirable behavior and the boundary corpus encodes prohibitions. So a higher boundary similarity means the action sits closer to a violation, and the formula subtracts it. Boundary proximity is assessed against a versioned boundary corpus, and its contents and construction remain held under NDA.

The Semantic Continuity Index describes semantic step-to-step similarity within the submitted work sequence as a separate quantity:

Scroll horizontally to read the full line.

SCIn=cos⁡(embed(actionn),embed(actionn−1))SCI_n = \cos(\text{embed}(\text{action}_n), \text{embed}(\text{action}_{n-1}))

which tracks semantic step-to-step coherence. The reviewed implementation evaluates continuity over a configured recent window. A chain break can remove inherited fidelity and affect the recorded classification. Explicit reset and configured length rollover are separate operations. A first step has no preceding-action similarity measurement.

10.2 Graduated outcomes

The scorer records execute, clarify or escalate under the applicable versioned profile. These are assessments. A human-review requirement is recorded separately, and it can come with an execute classification. Delivery, review, intervention and resolution each need their own deployment evidence.

  • execute records an action assessed as within the declared specification.
  • clarify marks proximity to a configured boundary or measurement uncertainty.
  • escalate marks the human-review path that deployment policy names.

A deployment can enable an optional micro-judge for unresolved cases. It is off by default.

10.3 Classification and handling evidence

Constitutive fidelity measures correspondence to the declared specification, and a recorded verdict is a classification. In the evaluated configuration, delay, clarification and human handling each need their own evidence, which means the adapter event record, the recipient, the response and the timing.

In the evaluated TES path, each reading is recorded with applied:false.

Under this configuration the action proceeds, and the system records the measurement and may surface it. This describes a deployment posture.

10.4 Configuration evidence

The filmed session’s receipt records the embedder, the engine commit, the Purpose Anchor hash, and the thresholds and weights in force. A reviewer needs separately preserved configuration evidence to work out the following.

  • Which embedding model and version were active
  • Which Manifest and Purpose Anchor were in force
  • Which thresholds and weights were applied
  • Whether optional classifiers or micro-judges were enabled

11. Architectural positioning

11.1 Middleware interface

The opened source provides a framework-independent governance protocol and adapters that an application can place between orchestration logic and a model or tool layer.

The interface has six properties:

  1. It scores submitted text and tool metadata without retraining the governed model.
  2. Measurement occurs when an adapter invokes the protocol.
  3. The interface is provider-independent.
  4. TELOS-defined local record formats are independent of model-provider response formats.
  5. The records support documentation and review.
  6. Coverage begins when the deployer integrates an orchestration path.

11.2 Provider independence

Provider independence is a property of the interface.

12. Implementation, configuration, and validation state

The inspected adapter interface calls the scorer and keeps assessment events. A rule-check branch is part of the design.

Three facts vary independently: whether a mechanism is implemented, whether this specification's configuration activates it, and whether its behavior has been validated. A mechanism can be built and inactive, or active and unvalidated, so the table gives each fact its own column.

Scroll horizontally to read all columns.

Mechanism Implementation Current configuration Validation
Per-action measurement against the Purpose Anchor Implemented Active on instrumented paths Validated on the cited tests and scorer profiles
Verdict emission (execute / clarify / escalate) Implemented Active See Section 10.3
Agent-path clarification behaviour Designed Not active in the evaluated posture Not validated across live tolerance settings
Documentation-to-Manifest compilation Designed; not built Not applicable Not validated
Full improvement cycle Designed, partial Not the current operating mode Not validated
Signed public Record of Trust Implemented Published snapshot Links and Root/Witness signatures verify with the public checker.

Designed: specified; executable evidence pending. Not active: configuration state. Implementation, configuration and validation must be stated separately.

Coverage extends to actions submitted on integrated paths.

13. Disclosure boundary

13.1 Disclosed

  • External fixed-reference architecture
  • Manifest and Purpose Anchor distinction
  • Cosine fidelity relationship
  • Five named agentic dimensions
  • Goal versus prohibition distinction
  • Baseline-relative EWMA presence
  • Runtime output categories
  • Historical archived deposits (not repeated here)

13.2 Held under NDA

  • Production embedding configuration
  • Boundary corpus
  • Centroid construction
  • Manifest compilation pipeline design
  • Unpublished deployment-specific weights and thresholds
  • EWMA constants and detailed baseline logic
  • Scoring subsystem internals
  • Signing key handling and deployment profile
  • Validation runners, protocol details, and specific scoring configuration

The disclosed method explains the architecture and the motivating mathematics.

14. Technical references

This paper cites no external works.