Architecture changelog

Nothing yet.

Initial version. This is the first versioned architecture. Earlier drafts lived in the October 2026 research package and in the layer repos and were never versioned; they are not listed here. Version 1.0.0 is the target architecture for the first three plants: what is built today, what is planned, and what is still to be decided. Nothing in it claims a result at a plant. Verified savings to date are ₹0.

Added#

• Summary (section 1). Four changes to the five-layer thesis: two loops instead of one stack, governance as a plane, a separate action layer, and agents as clients rather than a layer. Also the seven-job analytics chain, the predictive safety filter, the NAMUR NE 175 / NE 178 write path, and sustainment. Names used throughout, and the target state for the first three plants.
• Current architecture map (section 2). What exists today across the eleven product repos, a component table, and the duplications and contradictions to resolve.
• Layer model (section 3). Six conceptual layers (Plant systems, Context, Analytics, Decision, Action, Value), the governance plane, the fast loop at the plant and the slow loop in the cloud, how repo layers L0 to L6 map onto the conceptual layers, and a page per layer.
• Plant context graph (section 4). The storage options for assets, topology, states, lots, genealogy and documents, with the recommendation in D1.
• Data contracts (section 5). Ten contracts between layers: PlantState, Evidence, Prescription, Action, ValueRecord, AutonomyPolicy, Lot and genealogy, WriteRequest, ClosureState and Envelope, plus the versioning rules.
• Autonomy levels (section 6). AL0 to AL5. A person decides and acts at AL1; AL2 and AL3 are staged writes that a plant grants per action class and asset and can revoke at any time. Safety instrumented systems and interlocks stay outside Stamped's envelope.
• Edge and cloud split (section 7). What runs on the Plant Box and what runs in the cloud, the write path as an additive monitoring-and-optimisation channel, and the edge-cloud sync protocol.
• Agent architecture (section 8). Agents sit downstream of structured intelligence, read through typed tools, and propose. Deterministic code and people dispose.
• Scope cut (section 10). Every component cut into Build now, Later (with the trigger that justifies it) and Not planned.
• Operating rules (section 11 to section 15). Non-functional targets, security and tenancy, degraded modes, observability, and deployment profiles and capacity.
• Running cost (section 16) and product context for builders (section 17).
• Deep docs. Layer pages for L1, L2, L5 and L6; the L1-L2 data plane; L3 detection (5 docs); the L4 decision runtime (31 docs and a glossary); the fast loop (7 docs).
• System views. SYS-00 to SYS-09 and EXT-01 to EXT-05, redrawn in house style.
• Diagrams. 55 Mermaid diagrams in one house style, each with a "How to read it" list and a Build now / Later line.
• Tour. Eight short stops in tour/ for engineers joining Stamped, each with its own diagram, plus an "In short" summary for every architecture chapter and layer page. They restate the architecture and add nothing to it.

Decisions#

• The decision board moved out of the architecture into DECISIONS.md: 21 decisions, D1 to D21.
• Decided by Vinayak on 7 October 2026: D13 (evidence labels), D14 (calibration thresholds), D19 (MHE solver).
• Deferred: D18 (probabilistic modelling stack), to be chosen with the first C15 template.
• Recommended and awaiting Vinayak: the other 17. The architecture follows each recommendation until he decides, and every entry says how to reverse it.

Contracts#

• All ten contracts are drafted as 0.x, marked experimental (D2). The first plant that uses one in production bumps it to 1.0.0.