Applied research

Contextual parity
across schemas

Different systems should not leave agents with different versions of what the evidence means.

Context as an agentic virtue

For this research direction, contextual parity means preserving the decision-relevant meaning and limits of evidence for each authorized agent, even when their input schemas differ. Context includes where information came from, when it was observed, what it describes, and what remains uncertain.

A proposed context service would have to preserve those properties through ingestion, transformation, and delivery. The example below makes a small part of that problem inspectable.

One observation. Two source formats.

A fictional warehouse exports a CSV in Fahrenheit. A separate telemetry API reports Celsius. The identifiers and timestamp formats differ. Before either record can inform an agent, the relationship between them must be explicit.

Synthetic worked example

These fixed records and transformations run only in this page. They illustrate a research direction; they are not live integrations, an AI evaluation, or a demonstration of Matrix functionality.

Change the evidence

Choose a scenario to inspect what reaches both agents.

Aligned sourcesDifferent formats describe the same observation.Conflicting readingOne source reports a different temperature.Missing readingOne source supplies no temperature.

Keep the source evidence

Inventory export

Source reference: inventory.csv:row-2

site_code,temp_f,recorded_at
WH-17,68,2026-06-01T09:00:00-04:00

The export does not include a receipt time. That stays unknown.

Telemetry response

Source reference: telemetry.json:record-17

{
  "locationId": "warehouse-17",
  "temperatureC": 20,
  "observedAt": "2026-06-01T13:00:00Z",
  "receivedAt": "2026-06-01T13:04:00Z"
}

The receipt time describes delivery, not when the observation occurred.

Normalize meaning; preserve its limits

Shared temperature20 °CAligned

68 °F converts to 20 °C. Both records name the same warehouse and instant. The example produces a shared value without flagging a mismatch. Agreement alone does not establish that either source is correct.

Identity: An explicit fixture mapping connects WH-17 to warehouse-17. Similar names alone would not justify a match. Unknown identifiers would need resolution before joining records.
Units: (68 − 32) × 5/9 = 20 °C. The original value and unit remain in the source record; a renamed field would not perform this conversion.
Time: 09:00 −04:00 and 13:00 Z describe the same instant. Observation time stays separate from receipt time. No freshness judgment is inferred from arrival order.
Provenance: Both source references, original records, and the mapping version stay attached. An agent view points back to this shared evidence rather than replacing it.

Inspect the normalized context record

{
  "context_id": "synthetic-warehouse-17/aligned",
  "entity": "warehouse-17",
  "observed_at_utc": "2026-06-01T13:00:00.000Z",
  "measurement": {
    "kind": "ambient_temperature",
    "value": 20,
    "unit": "Celsius"
  },
  "evidence_state": "aligned",
  "review_required": false,
  "mapping": {
    "version": "illustration-v1",
    "identity_rule": "WH-17 maps to warehouse-17",
    "unit_rule": "(F - 32) × 5/9"
  },
  "observations": [
    {
      "source_ref": "inventory.csv:row-2",
      "original": {
        "site_code": "WH-17",
        "temp_f": 68,
        "recorded_at": "2026-06-01T09:00:00-04:00"
      },
      "normalized_c": 20,
      "received_at": null
    },
    {
      "source_ref": "telemetry.json:record-17",
      "original": {
        "locationId": "warehouse-17",
        "temperatureC": 20,
        "observedAt": "2026-06-01T13:00:00Z",
        "receivedAt": "2026-06-01T13:04:00Z"
      },
      "normalized_c": 20,
      "received_at": "2026-06-01T13:04:00Z"
    }
  ]
}

Project the context into each agent’s schema

The first agent expects flat planning fields. The second expects a nested review record. Both receive the same entity, instant, measurement state, review flag, and evidence references.

Planning-agent input

{
  "context_ref": "synthetic-warehouse-17/aligned",
  "warehouse_id": "warehouse-17",
  "observed_at_utc": "2026-06-01T13:00:00.000Z",
  "temperature_c": 20,
  "evidence_state": "aligned",
  "review_required": false,
  "evidence_refs": [
    "inventory.csv:row-2",
    "telemetry.json:record-17"
  ]
}

Review-agent input

{
  "context": {
    "id": "synthetic-warehouse-17/aligned",
    "subject": "warehouse-17",
    "time": "2026-06-01T13:00:00.000Z"
  },
  "measure": {
    "name": "ambient_temperature",
    "value": 20,
    "unit": "Celsius"
  },
  "assessment": {
    "state": "aligned",
    "requires_review": false
  },
  "lineage": [
    "inventory.csv:row-2",
    "telemetry.json:record-17"
  ]
}

Different schemas. The same evidence state.

A conflict must stay a conflict in both views. A missing reading must stay missing. A transformation that removes either warning would change the context an agent can reason from.

Both illustrative agents are assumed to have access to the same evidence. A real service would need to apply permissions before projection and make permitted omissions explicit. This fixture makes no operational recommendation and does not test agent reasoning.

Matrix: questions to test next

Could a decision-support system help people revisit a conclusion when its evidence changes, while keeping the assumptions and human review points visible? Context preservation is one question within that broader research direction.

Current prototype and future work

Matrix is an early, privately funded internal prototype. Early portions have been prototyped internally; Matrix has not received a government award or been deployed for customers. Broader capabilities would require additional funding, development, and evaluation.

The example above is an independent educational illustration. It establishes no result about Matrix. The tests below are proposed evaluation work, not completed experiments or measured performance.

Read the Matrix overview

A proposed evaluation plan

Does meaning survive a schema change?

Test setup: Create paired records with known equivalent meaning, plus deliberately incompatible units, ambiguous identifiers, and missing fields.
Evidence to inspect: Compare the preserved facts, relationships, uncertainty, and source references across each projection. Record any silent loss or false equivalence.

What happens when evidence changes?

Test setup: Replace a synthetic source observation after a downstream view has been produced.
Evidence to inspect: Check whether affected views are identified, refreshed, or explicitly marked stale. Keep observation and receipt times distinct.

Can an agent exceed its context permissions?

Test setup: Define two synthetic roles with different access to source fields and supporting records.
Evidence to inspect: Inspect generated views and trace references for unauthorized disclosure, including information revealed indirectly through summaries.

Can a person reconstruct the reasoning inputs?

Test setup: Ask a reviewer to trace an output back to its source versions, transformation rules, and unresolved conflicts.
Evidence to inspect: Record missing links, ambiguous mappings, and whether the reviewer can explain why a value was withheld or included.

Any future evaluation would need a defined task, representative data, explicit acceptance criteria, and a documented account of failures. Preserving context is necessary to study these questions; it does not by itself establish that an AI answer is correct.

Start with a transformation you can inspect

For a federal teaming or applied R&D discussion, bring a public description of the source formats, the agents or people who need the information, and the meaning that must survive between them.

Discuss a research scope