Stamped reads data a plant already produces, finds where good parts, material and energy are being lost, and sends one named person a ranked card with options. The person decides and acts; the software recommends, calculates and records, then checks the outcome against telemetry.

What Stamped does with plant data

⑤ Action and ⑥ Value: L5

③ Analytics and ④ Decision: L3 and L4

② Context: L1 and L2

1 Ingest: meters, machine state, quality and production records

2 Normalise: join enough to name one condition

3 Detect: evidence with a verification plan

4 Recommend, withhold or abstain

5 Assign: role resolved to a person on shift

6 Record: accept, edit, reject or defer

7 Verify: named evidence, honest closure state

Named person at the plant

8 Propose learning: a named owner accepts any rule or threshold change

How to read it.

  1. The numbers are the eight steps of the operating loop in architecture section 17.1, grouped by the layer that owns them.
  2. The blue box is the only place a decision is made. Nothing in the loop acts on the plant without a person.
  3. The green return path is learning. It goes back into plant memory only when a named owner accepts it.

Build now: steps 1 to 7 on the slow loop at AL1. Later: the fast loop at the plant, shadow first.

View Mermaid source
flowchart TB
    %% house-style: tour-operating-loop
    subgraph ctx["② Context: L1 and L2"]
        direction LR
        s1["1 Ingest: meters, machine state, quality and production records"] --> s2["2 Normalise: join enough to name one condition"]
    end
    subgraph ana["③ Analytics and ④ Decision: L3 and L4"]
        direction LR
        s3["3 Detect: evidence with a verification plan"] --> s4["4 Recommend, withhold or abstain"]
    end
    subgraph act["⑤ Action and ⑥ Value: L5"]
        direction LR
        s5["5 Assign: role resolved to a person on shift"] --> s6["6 Record: accept, edit, reject or defer"] --> s7["7 Verify: named evidence, honest closure state"]
    end
    person["Named person at the plant"]
    s8(["8 Propose learning: a named owner accepts any rule or threshold change"])
    s2 --> s3
    s4 --> s5
    s5 --> person --> s6
    s7 --> s8

    classDef govc fill:#fff4d6,stroke:#c99a2e,color:#000
    classDef agentc fill:#e8f0ff,stroke:#5b7bd5,color:#000
    classDef loopc fill:#eef7ee,stroke:#4f9a4f,color:#000
    class person agentc
    class s8 loopc

The problem we start from#

Plants lose things they have already paid for. Material goes into parts that never ship, machines run with loads that serve nothing, and a press restarts on a cold die. The master document puts it as good parts first: energy and waste is one outcome, stated next to good parts, and tariff mechanics such as maximum demand and time-of-day pricing sit inside that outcome.

Most of the data needed to see these losses already exists. It sits in PLCs, energy meters, historians, check-weighers, MES and ERP exports, and on paper. Stamped reads it, never replaces it. MES, ERP, APS, QMS and CMMS are context we read, not systems we compete with.

One card, one person#

The output that matters is a card. A card names one condition, one recommended action, at most two alternatives (hold is always one of them), the role that should own it, the evidence behind it and how the result will be checked. L5 resolves the role to a person on shift and sends the card through one sender.

The first pilot shows the shape. A machine is confirmed idle while extra loads are still on. The card goes to the ops head. Autonomy stays off, and verification measures the named auxiliary circuit rather than the whole-facility meter, because the whole-facility meter is the wrong boundary for that claim.

Closed does not mean successful. A card can close as closed_no_change or rejected, and both are honest outcomes that feed learning.

What the software does not do#

Stamped does not silently control safety systems or critical equipment. It does not release quality holds, sign off conformance, authorise maintenance or bypass lockout, and it does not quietly change customer priority, promise dates, routing or the dispatch sequence. A quality correction or a near-term sequence is allowed only when a named person accepts it before anything is written back. Stop 4 shows how those rules are enforced in code.

Nothing in the architecture claims a result at a plant yet. Verified savings to date are ₹0, and the word "verified" is reserved for a counterfactual measurement that passed its model-fit checks.

Go deeper: operating loop and live card · architecture summary

Page history: last 1 change
  1. 2026-10-08 docs(technical): architecture tour and page summaries 79c8ea3

Diagram

100%

Search the architecture