Open specifications

Cognitive infrastructure for the agent layer.

TMS and QPS are open specifications for the data layer AI agents actually read. They generalize patterns I've implemented in production at two enterprises under NDA. The reference manifolds published here are public-safe analogues of that work.

Read the TMS specificationRead QPS

How to read this site

Five entry points. Pick the one that fits your job.

  1. 01

    Read TMS if you want the agent-facing data format.

  2. 02

    Read QPS if you need replay, auditability, and drift detection.

  3. 03

    Open the TechnoFlex examples to see full manifolds and QPS entries end to end.

  4. 04

    Use the Calculations and Operations references if you are implementing either spec.

  5. 05

    Install the create-tms-manifold Claude skill if you want to generate a first draft.

The pairing

TMS is what the data shows. QPS is how the data was produced.

The agent reads down through the TMS manifold to reach a decision. Verification reads back up through the QPS entry to confirm the view still matches reality. Both artifacts come from the same raw data, and both stay portable across execution environments.

      Raw operational data
      (silver tables, warehouses)
                 |
                 v
      ╔══════════════════════════╗
      ║      TMS  manifold       ║   <-- what the data shows
      ║  L0 / L1 / L2 + lineage  ║
      ╚══════════════════════════╝
                 |
                 v
            Agent reasoning
                 |
                 v
       Decision / recommendation
                 ^
                 |
      ╔══════════════════════════╗
      ║       QPS  entry         ║   <-- how it was produced
      ║  query + generation +    ║       (replay + drift check)
      ║  executions log          ║
      ╚══════════════════════════╝

The family

Two specifications. One reasoning loop.

TMS · v1.2

Tabular Manifold Spec

A cognitive transmission format for AI agents. Feature engineering as an interface contract.

  • Progressive disclosure: L0 summary, L1 geometry, L2 telemetry
  • Six canonical kinds including entity_profile (v1.2)
  • Reliability bands, quality flags, and interpretation hints baked in
  • Designed to be returned as an MCP tool result
Read spec →Reference examplesCalculationsAuthor with Claude

QPS · v1.0

Query Provenance Store

The accountability layer for AI agent decisions. Where queries go to be remembered.

  • Immutable generation record plus append-only execution log
  • Drift detection without preventing drift
  • Companion to TMS via the lineage block
  • Parameterized queries only; opaque MCP-tool dialect for sensitive cases
Read spec →Reference entriesOperations referenceTMS integration

Worked examples

Four reference manifolds. Public-safe analogues of real engagements.

The TechnoFlex manifolds are fictional but structurally faithful. Each is a full L0 + L1 + L2 + lineage envelope with realistic quality flags, Pareto-truncated rollups, and outlier telemetry. View interactively, search across keys, download the JSON.

  • Item profile

    TechnoSeal IM-1652 Ionomer Resin

    Sole-sourced sealing material, HHI 1.0. Supply-continuity risk lives at L0.

    View manifold →

  • Supplier profile

    Helvian Specialty Polymers GmbH

    Multi-commodity strategic vendor. Weighted CV 0.04 across 18 SKUs.

    View manifold →

  • Commodity profile

    Industrial Colorants & Pigments

    94.7% concentration on one supplier, one SKU. Unambiguous sourcing call from L0.

    View manifold →

  • Supplier-category profile

    Specialty Chemicals & Additives

    $35.6M portfolio, 57 vendors, 86% direct, 35 industries. Healthy with one watch.

    View manifold →

Author with Claude

A Claude skill that builds these manifolds for you.

create-tms-manifold is a Claude skill that walks Claude through generating a v1.2 entity_profile manifold end to end: scope, L0 summary, L1 geometry, L2 telemetry preview, token_budget, lineage. Drop it into ~/.claude/skills/ and prompt Claude with the entity ID.

Install the skillSee the math the skill refers to →

License + posture

Both specifications are released under Apache 2.0. Use them freely.

If you're shipping agents that need to reason over operational data, start with the TMS spec. If you also need to answer “why did the agent see what it saw,” add QPS. The reference manifolds and the Claude skill exist to keep the time-to-first-draft short. If you want the architecture review that decides how it lands inside your stack, that conversation happens at the engagement level.

Start with an architecture review