[ { "id": "01-system-context", "title": "AI Code Completion Service — System Context", "layout": "context", "colWidth": 260, "system": { "label": "Code Completion Service", "sub": "Ghost text, per keystroke" }, "groups": [ { "side": "left", "title": "People", "nodes": [ { "id": "dev", "label": "Developer", "kind": "actor", "rel": "types, accepts", "dir": "in" }, { "id": "admin", "label": "Platform admin", "kind": "actor", "rel": "policy, seats", "dir": "in" }, { "id": "mleng", "label": "Model engineer", "kind": "actor", "rel": "promotes models", "dir": "in" } ] }, { "side": "top", "title": "Editor surfaces (in scope: the contract only)", "nodes": [ { "id": "vscode", "label": "VS Code plugin", "kind": "external", "rel": "suggest / cancel", "dir": "in" }, { "id": "jb", "label": "JetBrains plugin", "kind": "external", "rel": "suggest / cancel", "dir": "in" }, { "id": "nvim", "label": "Neovim plugin", "kind": "external", "rel": "suggest / cancel", "dir": "in" } ] }, { "side": "right", "title": "Platform dependencies", "nodes": [ { "id": "entra", "label": "Microsoft Entra ID", "kind": "security", "rel": "identity", "icon": "azure_AzureActiveDirectory__Identity" }, { "id": "scm", "label": "Source hosting", "kind": "external", "rel": "repo read, ACLs" }, { "id": "gpu", "label": "Azure GPU capacity", "kind": "platform", "rel": "accelerator pools", "icon": "azure_AzureKubernetesService__Containers" } ] }, { "side": "bottom", "title": "Downstream and out of scope", "nodes": [ { "id": "chat", "label": "Editor chat / agents", "kind": "external", "rel": "separate service", "kind2": "error" }, { "id": "billing", "label": "Billing platform", "kind": "external", "rel": "usage figures", "kind2": "batch" }, { "id": "train", "label": "Model training", "kind": "external", "rel": "no data path", "kind2": "error" } ] } ], "note": "Red paths are deliberate non-connections: chat and agents are a separate service, and no tenant code reaches training.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "02-high-level-architecture", "title": "High-Level Architecture — Keystroke to Ghost Text", "layout": "flow", "chain": false, "align": "middle", "canvas": { "width": 1760 }, "stages": [ { "title": "Editor", "nodes": [ { "id": "plugin", "label": "Editor plugin", "sub": "debounce, cancel" }, { "id": "local", "label": "Local redaction", "sub": "ignore rules" } ] }, { "title": "Edge", "nodes": [ { "id": "afd", "label": "Front Door", "sub": "anycast, WAF", "kind": "integration" }, { "id": "gate", "label": "Seat & quota gate", "sub": "deadline stamped", "kind": "integration" } ] }, { "title": "Context", "nodes": [ { "id": "ctx", "label": "FIM assembly", "sub": "token budget" }, { "id": "retr", "label": "Repo retrieval", "sub": "permission scoped" } ] }, { "title": "Admission", "nodes": [ { "id": "route", "label": "Affinity router", "sub": "session to replica" }, { "id": "adm", "label": "Admission control", "sub": "deadline aware", "kind": "decision" } ] }, { "title": "Inference", "nodes": [ { "id": "pool", "label": "GPU pool", "sub": "continuous batching" }, { "id": "kv", "label": "KV prefix cache", "sub": "HBM / DRAM / NVMe", "kind": "store" } ] }, { "title": "Filter", "nodes": [ { "id": "post", "label": "Post-process", "sub": "trim, dedup" }, { "id": "filt", "label": "Secret & licence scan", "sub": "fail closed", "kind": "security" } ] }, { "title": "Measure", "nodes": [ { "id": "tel", "label": "Telemetry ingest", "sub": "accept, retain", "kind": "queue" }, { "id": "qual", "label": "Quality plane", "sub": "control arm" } ] } ], "edges": [ { "from": "plugin", "to": "afd", "label": "suggest" }, { "from": "afd", "to": "gate" }, { "from": "gate", "to": "ctx", "label": "deadline" }, { "from": "ctx", "to": "route", "label": "prompt" }, { "from": "adm", "to": "pool", "label": "admit" }, { "from": "pool", "to": "post", "label": "stream" }, { "from": "post", "to": "filt" }, { "from": "filt", "to": "tel", "label": "events", "kind": "async" }, { "from": "tel", "to": "qual", "kind": "async" }, { "from": "adm", "to": "plugin", "label": "empty, never late", "kind": "error", "route": "gutter" } ], "note": "The spine is the happy path; the suggestion streams back along it. The red return is the normal path under load: a fast, silent nothing.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "03-actors-and-journeys", "title": "Actors and the Journeys They Own", "layout": "actors", "cardWidth": 280, "canvas": { "width": 1740 }, "groups": [ { "title": "The people who feel the latency", "kind": "boundary", "actors": [ { "id": "dev", "label": "Working developer", "sub": "1.2 M daily", "goal": "Finish the line I was already typing, without stopping to read something wrong.", "journeys": [ { "id": "j-ghost", "label": "Accept a suggestion", "sub": "2.4 B requests/day" }, { "label": "Ignore one and keep typing", "sub": "55% cancelled" } ] }, { "id": "lead", "label": "Engineering lead", "sub": "per repo group", "goal": "Know whether this is helping my team or quietly putting plausible nonsense in our code.", "journeys": [ { "label": "Read acceptance by repo" }, { "label": "Exclude a generated path" } ] } ] }, { "title": "The people who own the estate", "kind": "trust", "actors": [ { "id": "admin", "label": "Platform admin", "sub": "40 tenants", "goal": "Switch it on for 4,000 engineers without our source code being retained anywhere.", "journeys": [ { "id": "j-switch", "label": "Switch on a tenant", "sub": "zero retention" }, { "label": "Revoke a seat", "sub": "60 s effective" } ] }, { "id": "sec", "label": "Security reviewer", "kind": "security", "sub": "approves adoption", "goal": "See where code goes, what is retained, and what stops a key being suggested back.", "journeys": [ { "label": "Trace the data path" }, { "label": "Audit filter suppressions" } ], "icon": "fa5_user_shield" } ] }, { "title": "The people who change the model", "kind": "cloud", "actors": [ { "id": "mleng", "label": "Model engineer", "sub": "weekly cadence", "goal": "Ship a better model and know within a day whether it actually helped anyone.", "journeys": [ { "id": "j-rollout", "label": "Promote a model", "sub": "1% canary, 24 h" }, { "label": "Roll one back", "sub": "10 min" } ] }, { "id": "sre", "label": "On-call SRE", "kind": "platform", "sub": "follows the sun", "goal": "Tell the difference between a fleet that is slow and one that is quietly useless.", "journeys": [ { "label": "Drain a GPU node" }, { "label": "Shed multi-line traffic" } ] } ] }, { "title": "Machines in the cast", "kind": "plain", "actors": [ { "id": "indexer", "label": "Workspace indexer", "kind": "platform", "sub": "per repo", "goal": "Keep a retrieval index that reflects the branch the developer is actually on.", "journeys": [ { "label": "Reindex a changed repo" } ] }, { "id": "gate", "label": "Promotion gate", "kind": "decision", "sub": "automated", "goal": "Refuse a model that loses retention, without waking anybody up.", "journeys": [ { "label": "Block or promote" } ] } ] } ], "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "04-journey-the-ghost-text", "title": "Journey — The Developer and the Grey Text", "layout": "journey", "actor": { "label": "Working developer", "sub": "mid-file, mid-thought", "goal": "Finish this line without breaking the thought I am holding", "trigger": "A pause of 60 ms after a keystroke, inside a function body", "success": "Tab pressed, and the line still there an hour later" }, "phases": [ { "title": "Type" }, { "title": "Pause", "sub": "60 ms" }, { "title": "Wait", "moment": true }, { "title": "Read", "moment": true }, { "title": "Accept" }, { "title": "Live with it" } ], "lanes": [ { "title": "What they do", "kind": "step", "cells": [ [ { "label": "Types a few chars" } ], [ { "label": "Stops for a moment" } ], [ { "label": "Keeps eyes on cursor" } ], [ { "label": "Scans the grey text" } ], [ { "label": "Presses Tab" }, { "label": "Or types over it" } ], [ { "label": "Edits or deletes it" } ] ] }, { "title": "What the platform does", "kind": "system", "cells": [ [ { "label": "Nothing yet" } ], [ { "label": "Assembles FIM prompt" }, { "label": "Stamps the deadline" } ], [ { "label": "Reuses warm prefix" }, { "label": "Streams first tokens" } ], [ { "label": "Trims and scans" } ], [ { "label": "Logs accept event" } ], [ { "label": "Measures retention" } ] ] }, { "title": "How it feels", "kind": "emotion", "levels": [ "Invisible help", "Neutral", "In the way" ], "points": [ 1, 1, 0, 1, 2, 1 ] }, { "title": "Where it hurts", "kind": "pain", "cells": [ [], [], [ { "label": "Cold cache: +180 ms" }, { "label": "Reads as a stall" } ], [ { "label": "Plausible and wrong" } ], [], [ { "label": "Deleted at review" } ] ] }, { "title": "What answers it", "kind": "gain", "cells": [ [], [ { "label": "60 ms debounce" } ], [ { "label": "Session affinity" }, { "label": "Empty, never late" } ], [ { "label": "Degeneracy check" } ], [ { "label": "Accept telemetry" } ], [ { "label": "Retention as metric" } ] ] } ], "chain": true, "note": "The trough is the wait, and it is the whole architecture: a cold prefix reads as a broken editor, not as a slower model.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "05-journey-promoting-a-model", "title": "Journey — Promoting a Model Nobody Asked For", "layout": "journey", "actor": { "label": "Model engineer", "sub": "weekly cadence", "goal": "Prove the new model helps before 1.2 M developers meet it", "trigger": "A candidate that wins on the offline corpus by 4%", "success": "Promoted on retention, or blocked with a number attached" }, "phases": [ { "title": "Register", "sub": "content addressed" }, { "title": "Evaluate", "sub": "held-out corpus" }, { "title": "Canary", "sub": "1% / 24 h", "moment": true }, { "title": "Judge", "moment": true }, { "title": "Ramp" }, { "title": "Live with it" } ], "lanes": [ { "title": "What they do", "kind": "step", "cells": [ [ { "label": "Pushes a digest" } ], [ { "label": "Runs offline suite" } ], [ { "label": "Opens 1% split" } ], [ { "label": "Reads the arms" } ], [ { "label": "Ramps to 100%" } ], [ { "label": "Watches retention" } ] ] }, { "title": "What the platform does", "kind": "system", "cells": [ [ { "label": "Signs the artifact" } ], [ { "label": "Refuses unevaluated" } ], [ { "label": "Stable assignment" }, { "label": "Separate GPU pool" } ], [ { "label": "Computes effect size" } ], [ { "label": "Warms prefix caches" } ], [ { "label": "Holds the control arm" } ] ] }, { "title": "How it feels", "kind": "emotion", "levels": [ "Confident", "Uneasy", "Blind" ], "points": [ 2, 1, 1, 0, 1, 2 ] }, { "title": "Where it hurts", "kind": "pain", "cells": [ [], [ { "label": "Offline score misleads" } ], [ { "label": "Cold caches skew p95" } ], [ { "label": "Acceptance up, retention down" } ], [ { "label": "Rollout dismisses caches" } ], [] ] }, { "title": "What answers it", "kind": "gain", "cells": [ [ { "label": "Immutable registry" } ], [ { "label": "Eval as a gate, not a verdict" } ], [ { "label": "Warm-up before measuring" } ], [ { "label": "Retention is canonical" } ], [ { "label": "Staged ramp by region" } ], [ { "label": "Auto-rollback in 10 min" } ] ] } ], "chain": true, "note": "The trough is Judge: acceptance rises when a model is more confident, and retention is the only signal that notices the difference.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "06-journey-switching-on-a-tenant", "title": "Journey — Switching It On For 4,000 Engineers", "layout": "journey", "actor": { "label": "Platform admin", "sub": "with a security reviewer behind them", "goal": "Give the whole estate completion without giving away the estate's code", "trigger": "A budget approval and a security review that has not started well", "success": "Seats live, zero retention proved, and a usage figure they trust" }, "phases": [ { "title": "Ask", "sub": "security review" }, { "title": "Configure", "sub": "policy + geo" }, { "title": "Pilot", "sub": "one repo group", "moment": true }, { "title": "Roll out" }, { "title": "Account for it", "moment": true }, { "title": "Revoke" } ], "lanes": [ { "title": "What they do", "kind": "step", "cells": [ [ { "label": "Asks where code goes" } ], [ { "label": "Sets zero retention" }, { "label": "Pins the geography" } ], [ { "label": "Enables 40 engineers" } ], [ { "label": "Assigns 4,000 seats" } ], [ { "label": "Reads cost per accept" } ], [ { "label": "Revokes a leaver" } ] ] }, { "title": "What the platform does", "kind": "system", "cells": [ [ { "label": "Shows the data path" } ], [ { "label": "Refuses cross-region spill" } ], [ { "label": "Scoped retrieval index" } ], [ { "label": "Caches policy, fails strict" } ], [ { "label": "Attributes GPU seconds" } ], [ { "label": "Effective in 60 s" } ] ] }, { "title": "How it feels", "kind": "emotion", "levels": [ "Trusting", "Wary", "Blocked" ], "points": [ 0, 1, 1, 2, 0, 2 ] }, { "title": "Where it hurts", "kind": "pain", "cells": [ [ { "label": "\"Disabled\" is not absent" } ], [ { "label": "Policy conflicts silently" } ], [ { "label": "Index leaks private repo" } ], [], [ { "label": "Bill is per request" } ], [ { "label": "Credential still valid" } ] ] }, { "title": "What answers it", "kind": "gain", "cells": [ [ { "label": "Privacy by absence" } ], [ { "label": "Declared precedence" } ], [ { "label": "ACL at request time" } ], [ { "label": "Restrictive on cold cache" } ], [ { "label": "Cost per accepted" } ], [ { "label": "Short-lived tokens" } ] ] } ], "chain": true, "note": "The first trough is the review, and it is answered structurally: the retaining path does not exist rather than being switched off.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "07-layered-architecture", "title": "Layered Architecture", "layout": "bands", "layerHeaderWidth": 160, "canvas": { "width": 1700 }, "bands": [ { "name": "Editor clients", "nodes": [ { "id": "b-vs", "label": "VS Code", "kind": "external", "sub": "reference client" }, { "id": "b-jb", "label": "JetBrains", "kind": "external" }, { "id": "b-nv", "label": "Neovim", "kind": "external" }, { "id": "b-deb", "label": "Debounce & cancel" }, { "id": "b-red", "label": "Local redaction" } ] }, { "name": "Edge", "nodes": [ { "id": "b-afd", "label": "Front Door", "kind": "integration", "sub": "anycast + WAF" }, { "id": "b-seat", "label": "Seat check", "kind": "security" }, { "id": "b-quota", "label": "Tenant quota", "kind": "integration" }, { "id": "b-geo", "label": "Geography pin", "kind": "integration" }, { "id": "b-ver", "label": "Client version gate", "kind": "decision" } ] }, { "name": "Context", "nodes": [ { "id": "b-fim", "label": "FIM assembly" }, { "id": "b-tabs", "label": "Open-tab sources" }, { "id": "b-rag", "label": "Repo retrieval" }, { "id": "b-budget", "label": "Token budget" }, { "id": "b-stable", "label": "Stable-prefix order" } ] }, { "name": "Admission", "nodes": [ { "id": "b-pick", "label": "Model choice" }, { "id": "b-aff", "label": "Affinity lookup", "kind": "store", "sub": "Redis" }, { "id": "b-adm", "label": "Deadline admission", "kind": "decision" }, { "id": "b-shed", "label": "Ordered shedding" } ] }, { "name": "Inference", "nodes": [ { "id": "b-batch", "label": "Continuous batching", "sub": "AKS GPU pools" }, { "id": "b-hbm", "label": "KV cache — HBM", "kind": "store" }, { "id": "b-dram", "label": "KV cache — DRAM", "kind": "store" }, { "id": "b-nvme", "label": "KV cache — NVMe", "kind": "store" }, { "id": "b-spec", "label": "Speculative decode" } ] }, { "name": "Post-process", "nodes": [ { "id": "b-trim", "label": "Stop-sequence trim" }, { "id": "b-dedup", "label": "Suffix dedup" }, { "id": "b-sec", "label": "Secret scan", "kind": "security", "sub": "fail closed" }, { "id": "b-lic", "label": "Public-code match", "kind": "security" }, { "id": "b-deg", "label": "Degeneracy check", "kind": "decision" } ] }, { "name": "Measurement", "nodes": [ { "id": "b-eh", "label": "Event Hubs", "kind": "queue" }, { "id": "b-scrub", "label": "Privacy scrub", "kind": "security" }, { "id": "b-adx", "label": "Data Explorer", "kind": "store" }, { "id": "b-exp", "label": "Experiment arms" }, { "id": "b-gate2", "label": "Promotion gate", "kind": "decision" } ] }, { "name": "Control plane", "nodes": [ { "id": "b-sql", "label": "Policy & seats", "kind": "store", "sub": "Azure SQL" }, { "id": "b-reg", "label": "Model registry", "kind": "store", "sub": "ACR, by digest" }, { "id": "b-flag", "label": "Flags & kill switch", "kind": "platform" }, { "id": "b-kv2", "label": "Key Vault", "kind": "security" }, { "id": "b-mon", "label": "Azure Monitor", "kind": "platform" } ] } ], "edges": [ { "from": "b-deb", "to": "b-afd", "label": "suggest", "route": "tb" }, { "from": "b-ver", "to": "b-fim", "label": "deadline", "route": "tb" }, { "from": "b-stable", "to": "b-pick", "route": "tb" }, { "from": "b-adm", "to": "b-batch", "label": "admit", "route": "tb" }, { "from": "b-spec", "to": "b-trim", "label": "stream", "route": "tb" }, { "from": "b-deg", "to": "b-eh", "label": "events", "kind": "async", "route": "tb" }, { "from": "b-gate2", "to": "b-reg", "label": "promote", "kind": "async", "route": "gutter" }, { "from": "b-sql", "to": "b-seat", "label": "cached policy", "kind": "async", "route": "gutter" } ], "note": "The control plane spans every layer and sits on none of them: it is read through a cache and never on the critical path.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "08-platform-components", "title": "Platform Components (C4 L2)", "layout": "nested", "canvas": { "width": 1760 }, "boxes": [ { "title": "Developer machine", "kind": "onprem", "dir": "row", "children": [ { "title": "Editor process", "kind": "boundary", "nodes": [ { "id": "c-plug", "label": "Completion plugin", "sub": "debounce, cancel" }, { "id": "c-gather", "label": "Context gatherer", "sub": "tabs, recent edits" }, { "id": "c-red", "label": "Redactor", "sub": "ignore rules", "kind": "security" } ] } ] }, { "title": "Azure — serving region", "kind": "cloud", "dir": "col", "children": [ { "title": "Edge (Front Door + API gateway)", "kind": "boundary", "nodes": [ { "id": "c-afd", "label": "Front Door", "sub": "anycast, WAF", "kind": "integration" }, { "id": "c-authz", "label": "Seat authoriser", "sub": "Entra token", "kind": "security" }, { "id": "c-quota", "label": "Quota shaper", "kind": "integration" }, { "id": "c-dl", "label": "Deadline stamper", "kind": "integration" } ] }, { "title": "Request plane (Container Apps, stateless)", "kind": "boundary", "nodes": [ { "id": "c-fim", "label": "Prompt assembler", "sub": "FIM, budgeted" }, { "id": "c-ret", "label": "Retrieval service", "sub": "BM25 + symbols" }, { "id": "c-router", "label": "Affinity router" }, { "id": "c-post", "label": "Post-processor", "sub": "trim, dedup" }, { "id": "c-filter", "label": "Safety filters", "sub": "secret, licence", "kind": "security" } ] }, { "title": "Inference plane (AKS GPU node pools, session affine)", "kind": "boundary", "nodes": [ { "id": "c-sched", "label": "Batch scheduler", "sub": "continuous batching" }, { "id": "c-small", "label": "Fast model server", "sub": "single-line" }, { "id": "c-large", "label": "Multi-line server", "sub": "escalated" }, { "id": "c-cache", "label": "KV cache manager", "sub": "3 tiers", "kind": "store" } ] }, { "title": "Measurement plane", "kind": "boundary", "nodes": [ { "id": "c-eh", "label": "Event Hubs", "sub": "two streams", "kind": "queue" }, { "id": "c-scrub", "label": "Scrubber", "kind": "security" }, { "id": "c-adx", "label": "Data Explorer", "sub": "aggregates", "kind": "store" }, { "id": "c-qual", "label": "Quality service", "sub": "arms, gating" } ] }, { "title": "Control plane (regional read, global write)", "kind": "trust", "nodes": [ { "id": "c-sql", "label": "Policy & seats", "sub": "Azure SQL", "kind": "store" }, { "id": "c-redis", "label": "Policy & affinity cache", "sub": "Redis", "kind": "store" }, { "id": "c-acr", "label": "Model registry", "sub": "ACR, by digest", "kind": "store" }, { "id": "c-flag", "label": "Flags & kill switch", "kind": "platform" }, { "id": "c-akv", "label": "Key Vault / HSM", "kind": "security" } ] } ] } ], "outside": [ { "id": "c-entra", "label": "Microsoft Entra ID", "kind": "security", "icon": "azure_AzureActiveDirectory__Identity" }, { "id": "c-scm", "label": "Source hosting", "kind": "external", "sub": "repo + ACLs" }, { "id": "c-mon", "label": "Azure Monitor", "kind": "platform" }, { "id": "c-blob", "label": "Blob Storage", "kind": "store", "sub": "artifacts, raw" } ], "edges": [ { "from": "c-plug", "to": "c-afd", "label": "suggest" }, { "from": "c-dl", "to": "c-fim", "label": "deadline" }, { "from": "c-fim", "to": "c-router" }, { "from": "c-ret", "to": "c-scm", "label": "ACL check", "kind": "sync" }, { "from": "c-router", "to": "c-sched", "label": "session key" }, { "from": "c-filter", "to": "c-eh", "label": "events", "kind": "async" }, { "from": "c-qual", "to": "c-acr", "label": "promote", "kind": "async" }, { "from": "c-acr", "to": "c-blob", "label": "weights", "kind": "batch" } ], "note": "Omitted for legibility: the per-tier Monitor exporters, private endpoints, the Entra token path and the affinity-map write, all drawn on views 09, 11, 15, 17, 18 and 20.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "09-integration-surface", "title": "Integration Surface", "layout": "hub", "canvas": { "width": 1700 }, "left": { "title": "Callers", "nodes": [ { "id": "i-vs", "label": "VS Code plugin", "kind": "external", "sub": "reference client" }, { "id": "i-jb", "label": "JetBrains plugin", "kind": "external" }, { "id": "i-nv", "label": "Neovim plugin", "kind": "external" }, { "id": "i-adminui", "label": "Admin console", "kind": "external" }, { "id": "i-ci", "label": "Model promotion CI", "kind": "platform" } ] }, "centre": { "title": "Code Completion Service", "nodes": [ { "id": "i-sug", "label": "Suggestion API", "sub": "streaming, cancellable", "kind": "integration", "icon": "azure_AzureAPIManagement__Web" }, { "id": "i-tel", "label": "Telemetry API", "sub": "accept / retain", "kind": "integration", "icon": "azure_AzureEventHub__Analytics" }, { "id": "i-adm", "label": "Admin API", "sub": "policy, seats, usage", "kind": "integration", "icon": "azure_AzureAPIManagement__Web" }, { "id": "i-mdl", "label": "Model API", "sub": "register, split, roll back", "kind": "integration", "icon": "azure_AzureAPIManagement__Web" } ] }, "right": { "title": "Dependencies", "nodes": [ { "id": "i-entra", "label": "Microsoft Entra ID", "sub": "identity", "kind": "security", "rel": "", "dir": "out", "icon": "azure_AzureActiveDirectory__Identity" }, { "id": "i-scm", "label": "Source hosting", "sub": "repo + ACL", "kind": "external", "rel": "", "dir": "out" }, { "id": "i-gpu", "label": "AKS GPU pools", "sub": "inference", "kind": "platform", "rel": "", "dir": "out", "icon": "azure_AzureKubernetesService__Containers" }, { "id": "i-lic", "label": "Public-code corpus", "sub": "match index", "kind": "external", "rel": "", "dir": "out", "kind2": "batch" } ] }, "note": "Four APIs, one of which is on a 300 ms budget; the other three are control-plane calls and may be slow. Monitor and the nightly usage export are drawn on views 01 and 18.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "10-request-path-data-flow", "title": "Data Flow — What Leaves the Machine and What Comes Back", "layout": "flow", "align": "top", "canvas": { "width": 1780 }, "stages": [ { "title": "On the machine", "nodes": [ { "id": "d-buf", "label": "Open buffers", "kind": "store", "sub": "never uploaded whole" }, { "id": "d-pick", "label": "Source selection" }, { "id": "d-redact", "label": "Redaction", "kind": "security", "sub": "secrets, ignores" } ] }, { "title": "In flight", "nodes": [ { "id": "d-prompt", "label": "FIM prompt", "sub": "≤ 8 k tokens" }, { "id": "d-hdr", "label": "Request envelope", "sub": "tenant, seat, deadline" } ] }, { "title": "Held in memory only", "nodes": [ { "id": "d-ctx", "label": "Assembled context" }, { "id": "d-kv", "label": "KV prefix state", "kind": "store", "sub": "recomputable" }, { "id": "d-out", "label": "Generated tokens" } ] }, { "title": "Split at the boundary", "nodes": [ { "id": "d-split", "label": "Stream split", "kind": "decision", "sub": "code vs not" } ] }, { "title": "Persisted — no code", "nodes": [ { "id": "d-ev", "label": "Interaction events", "kind": "queue", "sub": "ids, latency, verdicts" }, { "id": "d-agg", "label": "Quality aggregates", "kind": "store", "sub": "ADX, 400 days" }, { "id": "d-usage", "label": "Usage & cost", "kind": "store", "sub": "per tenant" } ] }, { "title": "Persisted — may hold code", "nodes": [ { "id": "d-diag", "label": "Diagnostic sample", "kind": "risk", "sub": "opt-in, ≤ 24 h" }, { "id": "d-eval", "label": "Eval corpus", "kind": "store", "sub": "licensed only" } ] }, { "title": "Never persisted", "nodes": [ { "id": "d-zero", "label": "Zero-retention tenants", "kind": "security", "sub": "no path exists" }, { "id": "d-notrain", "label": "Training", "kind": "risk", "sub": "no data path" } ] } ], "edges": [ { "from": "d-redact", "to": "d-prompt", "label": "redacted span" }, { "from": "d-hdr", "to": "d-ctx" }, { "from": "d-ctx", "to": "d-kv", "label": "prefix" }, { "from": "d-out", "to": "d-split", "label": "suggestion" }, { "from": "d-split", "to": "d-ev", "label": "metadata", "kind": "async" }, { "from": "d-split", "to": "d-diag", "label": "opt-in only", "kind": "error" } ], "note": "The red path is the only way code-bearing data is retained, and it is refused by default. Zero-retention tenants have no arrow because they have no path.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "11-data-stores-and-ownership", "title": "Data Architecture — Ownership and Rebuildability", "layout": "nested", "canvas": { "width": 1740 }, "boxes": [ { "title": "Systems of record — lose these and the platform is gone", "kind": "trust", "dir": "row", "children": [ { "title": "Control plane (global write, regional read replicas)", "kind": "boundary", "nodes": [ { "id": "s-sql", "label": "Policy, tenants, seats", "kind": "store", "sub": "Azure SQL, RPO 0" }, { "id": "s-consent", "label": "Consent & audit", "kind": "store", "sub": "append-only, 5 y" }, { "id": "s-reg", "label": "Model registry", "kind": "store", "sub": "ACR, immutable" } ] }, { "title": "Measurement (append-only)", "kind": "boundary", "nodes": [ { "id": "s-raw", "label": "Interaction log", "kind": "store", "sub": "Blob, 30 d raw" }, { "id": "s-eval", "label": "Eval corpora", "kind": "store", "sub": "licensed, 2 y" } ] } ] }, { "title": "Rebuildable projections — no backup, only a recompute time", "kind": "boundary", "dir": "row", "children": [ { "title": "Quality", "kind": "plain", "nodes": [ { "id": "s-adx", "label": "Quality aggregates", "kind": "store", "sub": "ADX, 400 d" }, { "id": "s-arm", "label": "Experiment arms", "kind": "store" } ] }, { "title": "Retrieval", "kind": "plain", "nodes": [ { "id": "s-idx", "label": "Workspace indexes", "kind": "store", "sub": "40 TB, per repo" }, { "id": "s-sym", "label": "Symbol graph", "kind": "store" } ] }, { "title": "Caches", "kind": "plain", "nodes": [ { "id": "s-pol", "label": "Policy cache", "kind": "store", "sub": "Redis, 30 s" }, { "id": "s-aff", "label": "Affinity map", "kind": "store", "sub": "Redis, session TTL" } ] } ] }, { "title": "Ephemeral — recomputed, never recovered", "kind": "lane", "dir": "row", "children": [ { "title": "Inference memory", "kind": "plain", "nodes": [ { "id": "s-hbm", "label": "KV cache — HBM", "kind": "store", "sub": "device memory" }, { "id": "s-dram", "label": "KV cache — DRAM", "kind": "store", "sub": "host spill" }, { "id": "s-nvme", "label": "KV cache — NVMe", "kind": "store", "sub": "node local" }, { "id": "s-inflight", "label": "Prompt & suggestion", "kind": "risk", "sub": "memory only" } ] } ] } ], "outside": [ { "id": "s-scm", "label": "Source hosting", "kind": "external", "sub": "the code's real owner" } ], "edges": [ { "from": "s-scm", "to": "s-idx", "label": "rebuilt from", "kind": "batch" }, { "from": "s-raw", "to": "s-adx", "label": "replayable", "kind": "batch" }, { "from": "s-sql", "to": "s-pol", "label": "cached", "kind": "async" }, { "from": "s-inflight", "to": "s-raw", "label": "metadata only", "kind": "async" } ], "note": "The platform owns no code. Every store that could hold code is either ephemeral, opt-in, or a rebuildable projection of a repository someone else owns.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "12-telemetry-and-policy-data-model", "title": "Data Model — Policy, Session and the Suggestion Event", "layout": "er", "canvas": { "width": 1760, "cols": 4 }, "rowGap": 230, "entities": [ { "id": "tenant", "name": "tenant", "kind": "store", "row": 0, "col": 1, "attrs": [ "tenant_id PK", "retention_mode", "geography", "model_tier", "spend_ceiling" ] }, { "id": "seat", "name": "seat", "kind": "store", "row": 0, "col": 2, "attrs": [ "seat_id PK", "tenant_id FK", "principal_oid", "state", "revoked_at" ] }, { "id": "policy", "name": "policy_rule", "kind": "store", "row": 0, "col": 3, "attrs": [ "rule_id PK", "scope tenant|group|seat", "languages", "filters", "precedence" ] }, { "id": "consent", "name": "consent_record", "kind": "store", "row": 0, "col": 0, "attrs": [ "record_id PK", "tenant_id FK", "change", "actor", "at immutable" ] }, { "id": "session", "name": "edit_session", "kind": "store", "row": 1, "col": 2, "attrs": [ "session_id PK", "seat_id FK", "workspace_id FK", "replica_id", "prefix_hash" ] }, { "id": "request", "name": "suggestion_request", "kind": "store", "row": 1, "col": 3, "attrs": [ "request_id PK", "session_id FK", "model_digest FK", "recipe_version", "deadline_ms" ] }, { "id": "outcome", "name": "suggestion_outcome", "kind": "store", "row": 2, "col": 3, "attrs": [ "request_id PK FK", "state", "ttft_ms", "filtered_by", "arm_id FK" ] }, { "id": "retention", "name": "retention_check", "kind": "store", "row": 2, "col": 2, "attrs": [ "request_id FK", "at_seconds", "chars_surviving", "measured_at" ] }, { "id": "workspace", "name": "workspace", "kind": "store", "row": 1, "col": 1, "attrs": [ "workspace_id PK", "tenant_id FK", "repo_ref", "index_version" ] }, { "id": "model", "name": "model_version", "kind": "store", "row": 1, "col": 0, "attrs": [ "model_digest PK", "family", "decoding_config", "eval_run_id", "signed_by" ] }, { "id": "arm", "name": "experiment_arm", "kind": "store", "row": 2, "col": 0, "attrs": [ "arm_id PK", "experiment_id", "model_digest FK", "traffic_pct", "is_control" ] }, { "id": "cost", "name": "cost_attribution", "kind": "store", "row": 2, "col": 1, "attrs": [ "day, tenant_id, model_digest PK", "gpu_seconds", "displayed", "accepted", "wasted_pct" ] } ], "relations": [ { "from": "tenant", "to": "consent", "label": "1 : N", "from_side": "w", "to_side": "e" }, { "from": "tenant", "to": "seat", "label": "1 : N", "from_side": "e", "to_side": "w" }, { "from": "seat", "to": "policy", "label": "0 : N", "from_side": "e", "to_side": "w", "kind": "optional" }, { "from": "tenant", "to": "workspace", "label": "1 : N", "from_side": "s", "to_side": "n" }, { "from": "workspace", "to": "session", "label": "1 : N", "from_side": "e", "to_side": "w" }, { "from": "session", "to": "request", "label": "1 : N", "from_side": "e", "to_side": "w" }, { "from": "request", "to": "outcome", "label": "1 : 1", "from_side": "s", "to_side": "n" }, { "from": "outcome", "to": "retention", "label": "1 : N", "from_side": "w", "to_side": "e" }, { "from": "model", "to": "arm", "label": "1 : N", "from_side": "s", "to_side": "n" }, { "from": "model", "to": "cost", "label": "1 : N", "from_side": "s3", "to_side": "n1" } ], "note": "No entity holds code: suggestion_request stores the recipe version, never the prompt. The model_digest on a request and the arm_id on an outcome are columns rather than drawn edges.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "13-keystroke-to-ghost-text", "title": "Critical Flow — One Keystroke, 300 Milliseconds", "layout": "sequence", "canvas": { "width": 1760 }, "lifelines": [ { "id": "dev", "label": "Developer", "kind": "actor" }, { "id": "plug", "label": "Editor plugin" }, { "id": "edge", "label": "Edge + seat gate", "kind": "integration" }, { "id": "ctx", "label": "Prompt assembler" }, { "id": "route", "label": "Affinity router" }, { "id": "gpu", "label": "GPU replica", "icon": "azure_AzureKubernetesService__Containers" }, { "id": "filt", "label": "Safety filters", "kind": "security" }, { "id": "tel", "label": "Telemetry", "kind": "queue", "icon": "azure_AzureEventHub__Analytics" } ], "messages": [ { "from": "dev", "to": "plug", "label": "keystroke", "kind": "call" }, { "from": "plug", "to": "plug", "label": "cancel previous in-flight", "kind": "self" }, { "from": "plug", "to": "plug", "label": "debounce 60 ms", "kind": "self" }, { "from": "plug", "to": "edge", "label": "POST /suggest (deadline 2000 ms)", "kind": "call" }, { "from": "edge", "to": "edge", "label": "seat, quota, version, geography", "kind": "self" }, { "from": "edge", "to": "ctx", "label": "forward, deadline 1940 ms", "kind": "call" }, { "from": "ctx", "to": "ctx", "label": "FIM assembly, budget, stable order", "kind": "self" }, { "from": "ctx", "to": "route", "label": "prompt + session key", "kind": "call" }, { "from": "route", "to": "route", "label": "affinity lookup; admit or refuse", "kind": "self" }, { "from": "route", "to": "plug", "label": "no capacity in deadline — empty", "kind": "error" }, { "from": "route", "to": "gpu", "label": "admit to in-flight batch", "kind": "call" }, { "from": "gpu", "to": "gpu", "label": "prefix cache hit (85%)", "kind": "self" }, { "from": "gpu", "to": "filt", "label": "first tokens, streaming", "kind": "call" }, { "from": "filt", "to": "filt", "label": "trim, suffix dedup, secret scan", "kind": "self" }, { "from": "filt", "to": "plug", "label": "displayable at 120 ms p50", "kind": "return" }, { "from": "plug", "to": "dev", "label": "ghost text painted", "kind": "return" }, { "from": "dev", "to": "plug", "label": "Tab, or keeps typing", "kind": "call" }, { "from": "plug", "to": "tel", "label": "accepted / cancelled", "kind": "async" }, { "from": "plug", "to": "tel", "label": "retention probe at 120 s", "kind": "async" } ], "note": "Message 10 is the one that keeps the contract: the refusal is issued before any accelerator work starts.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "14-context-assembly-pipeline", "title": "Context Assembly — Spending a Token Budget", "layout": "flow", "align": "top", "canvas": { "width": 1780 }, "stages": [ { "title": "Candidate sources", "nodes": [ { "id": "p-cur", "label": "Current file", "sub": "prefix + suffix" }, { "id": "p-tabs", "label": "Open tabs", "sub": "up to 20" }, { "id": "p-recent", "label": "Recent edits", "sub": "last 10 min" }, { "id": "p-imp", "label": "Imported symbols" }, { "id": "p-repo", "label": "Repo snippets", "sub": "retrieved" } ] }, { "title": "Admission", "nodes": [ { "id": "p-acl", "label": "ACL check", "kind": "security", "sub": "at request time" }, { "id": "p-ign", "label": "Ignore rules", "kind": "security" }, { "id": "p-poison", "label": "Poison detector", "kind": "decision", "sub": "minified, binary" } ] }, { "title": "Rank", "nodes": [ { "id": "p-lex", "label": "Lexical score", "sub": "BM25, in-process" }, { "id": "p-sym", "label": "Symbol proximity" }, { "id": "p-rec", "label": "Recency weight" } ] }, { "title": "Allocate", "nodes": [ { "id": "p-budget", "label": "Budget allocator", "sub": "priority order" }, { "id": "p-trunc", "label": "Lowest-priority truncation", "kind": "decision" } ] }, { "title": "Order", "nodes": [ { "id": "p-stable", "label": "Stable segment", "sub": "cacheable prefix" }, { "id": "p-vol", "label": "Volatile tail", "sub": "cursor neighbourhood" } ] }, { "title": "Emit", "nodes": [ { "id": "p-prompt", "label": "FIM prompt", "sub": "≤ 8 k tokens" }, { "id": "p-recipe", "label": "Recipe version", "sub": "logged, not the prompt" } ] } ], "edges": [ { "from": "p-repo", "to": "p-acl", "label": "scoped" }, { "from": "p-cur", "to": "p-ign" }, { "from": "p-poison", "to": "p-lex" }, { "from": "p-rec", "to": "p-budget", "label": "ranked" }, { "from": "p-trunc", "to": "p-stable", "label": "kept" }, { "from": "p-vol", "to": "p-prompt" }, { "from": "p-budget", "to": "p-tabs", "label": "tabs only", "kind": "error", "route": "gutter" } ], "note": "Order is not cosmetic: a volatile segment placed early invalidates the prefix cache on every keystroke and costs 180 ms.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "15-four-fates-of-a-request", "title": "Inference Plane — Four Fates of One Keystroke", "layout": "swimlane", "canvas": { "width": 1780 }, "laneHeaderWidth": 190, "stages": [ "Arrive", "Affinity lookup", "Admission", "Batch slot", "Prefix state", "Outcome" ], "lanes": [ { "title": "Warm session — 85%", "cells": [ [ { "label": "Session key known" } ], [ { "label": "Replica hit", "kind": "store" } ], [ { "label": "Queue wait < budget", "kind": "decision" } ], [ { "label": "Joins in-flight batch" } ], [ { "label": "HBM cache hit", "kind": "store" } ], [ { "label": "p50 120 ms", "kind": "opportunity" } ] ] }, { "title": "Cold session — 15%", "cells": [ [ { "label": "First keystroke" } ], [ { "label": "No replica held" } ], [ { "label": "Any replica accepted", "kind": "decision" } ], [ { "label": "Admitted, lower priority" } ], [ { "label": "Recompute prefix" } ], [ { "label": "+180 ms at p95", "kind": "risk" } ] ] }, { "title": "Drained mid-session", "cells": [ [ { "label": "Node upgrade or spot loss" } ], [ { "label": "Affinity invalidated", "kind": "risk" } ], [ { "label": "Re-placed, not errored", "kind": "decision" } ], [ { "label": "New replica" } ], [ { "label": "NVMe tier read", "kind": "store" } ], [ { "label": "Degrades, never fails", "kind": "opportunity" } ] ] }, { "title": "Refused at the door", "cells": [ [ { "label": "Peak, 90 k rps" } ], [ { "label": "Lookup skipped" } ], [ { "label": "Deadline unreachable", "kind": "risk" } ], [], [], [ { "label": "Empty, 0 GPU seconds", "kind": "opportunity" } ] ] }, { "title": "Cancelled — 55% of all", "cells": [ [ { "label": "Developer kept typing" } ], [ { "label": "Same replica" } ], [ { "label": "Already admitted", "icon": "fa5_check" } ], [ { "label": "Sequence evicted" } ], [ { "label": "Prefix kept warm", "kind": "store" } ], [ { "label": "Reclaimed in 50 ms", "kind": "opportunity" } ] ] } ], "note": "Row four costs nothing and row five costs most of the fleet. Budgeting for row five is what makes row one affordable.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "16-model-promotion-and-gating", "title": "Model Promotion — Offline Gate, Online Verdict", "layout": "flow", "align": "top", "canvas": { "width": 1780 }, "stages": [ { "title": "Register", "nodes": [ { "id": "m-art", "label": "Weights + config", "kind": "store", "sub": "content addressed" }, { "id": "m-sign", "label": "Sign & admit", "kind": "security", "sub": "Key Vault" } ] }, { "title": "Offline gate", "nodes": [ { "id": "m-corp", "label": "Held-out contexts", "kind": "store", "sub": "licensed only" }, { "id": "m-eval", "label": "Eval harness" }, { "id": "m-pass", "label": "Promotion allowed?", "kind": "decision" } ] }, { "title": "Canary", "nodes": [ { "id": "m-arm", "label": "1% arm", "sub": "stable assignment" }, { "id": "m-ctrl", "label": "Control arm", "sub": "held back" }, { "id": "m-warm", "label": "Cache warm-up", "sub": "before measuring" } ] }, { "title": "Online verdict", "nodes": [ { "id": "m-acc", "label": "Acceptance rate", "sub": "immediate, noisy" }, { "id": "m-ret", "label": "Retention at 120 s", "sub": "canonical" }, { "id": "m-lat", "label": "TTFT p95" } ] }, { "title": "Decide", "nodes": [ { "id": "m-gate", "label": "Regression > 1.5%?", "kind": "decision" } ] }, { "title": "Ramp or revert", "nodes": [ { "id": "m-ramp", "label": "Staged ramp", "sub": "by region" }, { "id": "m-back", "label": "Alias to previous", "kind": "opportunity", "sub": "≤ 10 min" }, { "id": "m-keep", "label": "Previous stays mounted", "kind": "store", "sub": "72 h" } ] } ], "edges": [ { "from": "m-sign", "to": "m-eval" }, { "from": "m-corp", "to": "m-eval" }, { "from": "m-pass", "to": "m-arm", "label": "allowed" }, { "from": "m-warm", "to": "m-ret", "label": "then measure" }, { "from": "m-ret", "to": "m-gate" }, { "from": "m-gate", "to": "m-ramp", "label": "no" }, { "from": "m-gate", "to": "m-back", "label": "yes", "kind": "error" } ], "note": "The offline suite is a gate, not a verdict: it can refuse a candidate before it reaches a canary, and it can never promote one. Only retention can.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "17-deployment-topology", "title": "Deployment Topology — Three Regions, One Working Set Each", "layout": "nested", "canvas": { "width": 1780 }, "boxes": [ { "title": "Global", "kind": "cloud", "dir": "row", "children": [ { "title": "Edge and control", "kind": "boundary", "nodes": [ { "id": "g-afd", "label": "Front Door", "sub": "anycast + WAF", "kind": "integration" }, { "id": "g-sql", "label": "Policy & seats", "sub": "SQL, write region", "kind": "store" }, { "id": "g-acr", "label": "Model registry", "sub": "ACR geo-replicated", "kind": "store" }, { "id": "g-akv", "label": "Key Vault / HSM", "kind": "security" } ] } ] }, { "title": "Serving region — West Europe (primary pattern, repeated per region)", "kind": "cloud", "dir": "row", "children": [ { "title": "Zone 1", "kind": "boundary", "nodes": [ { "id": "z1-req", "label": "Request plane", "sub": "Container Apps" }, { "id": "z1-gpu", "label": "GPU node pool", "sub": "session affine", "icon": "azure_AzureKubernetesService__Containers" }, { "id": "z1-redis", "label": "Redis replica", "kind": "store" } ] }, { "title": "Zone 2", "kind": "boundary", "nodes": [ { "id": "z2-req", "label": "Request plane" }, { "id": "z2-gpu", "label": "GPU node pool", "icon": "azure_AzureKubernetesService__Containers" }, { "id": "z2-redis", "label": "Redis replica", "kind": "store" } ] }, { "title": "Zone 3 (measurement and spare)", "kind": "boundary", "nodes": [ { "id": "z3-eh", "label": "Event Hubs", "kind": "queue" }, { "id": "z3-adx", "label": "Data Explorer", "kind": "store" }, { "id": "z3-idx", "label": "Retrieval indexes", "kind": "store", "sub": "per workspace" } ] } ] }, { "title": "Other serving regions", "kind": "cloud", "dir": "row", "children": [ { "title": "East US", "kind": "boundary", "nodes": [ { "id": "us-gpu", "label": "GPU pools", "sub": "own working set", "icon": "azure_AzureKubernetesService__Containers" }, { "id": "us-req", "label": "Request plane" } ] }, { "title": "Southeast Asia", "kind": "boundary", "nodes": [ { "id": "ap-gpu", "label": "GPU pools", "icon": "azure_AzureKubernetesService__Containers" }, { "id": "ap-req", "label": "Request plane" } ] }, { "title": "Sovereign option (phase 3)", "kind": "onprem", "nodes": [ { "id": "sv-gpu", "label": "Tenant-dedicated pool", "kind": "risk", "sub": "not in MVP", "icon": "azure_AzureKubernetesService__Containers" } ] } ] } ], "outside": [ { "id": "o-mon", "label": "Azure Monitor + Managed Grafana", "kind": "platform" }, { "id": "o-blob", "label": "Blob Storage", "kind": "store", "sub": "weights, raw telemetry" } ], "edges": [ { "from": "g-afd", "to": "z1-req", "label": "geography pinned" }, { "from": "g-acr", "to": "z1-gpu", "label": "weights by digest", "kind": "batch" }, { "from": "z1-redis", "to": "z2-redis", "label": "affinity map", "kind": "async" }, { "from": "z3-eh", "to": "o-blob", "label": "capture", "kind": "async" } ], "note": "A region is a complete working set, not a fraction of one: prefix caches are never replicated, so a region loss costs cold starts, not correctness.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "18-observability", "title": "Observability — Signal by Stage", "layout": "grid", "canvas": { "width": 1780 }, "laneHeaderWidth": 170, "columns": [ "Editor client", "Edge", "Context", "Admission", "Inference", "Filters & telemetry" ], "rows": [ { "title": "Latency", "cells": [ [ { "label": "TTFT, client measured", "kind": "platform" } ], [ { "label": "Edge overhead" } ], [ { "label": "Assembly p95" } ], [ { "label": "Queue wait p99" } ], [ { "label": "Tokens / second" } ], [ { "label": "Filter time p99" } ] ] }, { "title": "Saturation", "cells": [ [ { "label": "In-flight per session" } ], [ { "label": "Quota rejections" } ], [ { "label": "Budget truncations" } ], [ { "label": "Admission refusals", "kind": "risk" } ], [ { "label": "Batch occupancy" } ], [ { "label": "Event Hubs lag" } ] ] }, { "title": "Cache & affinity", "cells": [ [], [], [ { "label": "Prefix stability" } ], [ { "label": "Affinity hit rate", "kind": "store" } ], [ { "label": "Tier hit: HBM / DRAM / NVMe", "kind": "store" } ], [] ] }, { "title": "Quality", "cells": [ [ { "label": "Displayed, accepted" } ], [], [ { "label": "Recipe version" } ], [], [ { "label": "Per-model acceptance" } ], [ { "label": "Retention at 120 s", "kind": "opportunity" } ] ] }, { "title": "Safety", "cells": [ [], [ { "label": "Client version floor" } ], [ { "label": "Redaction hits", "kind": "security" } ], [], [ { "label": "Degeneracy ejections" } ], [ { "label": "Secret suppressions", "kind": "security" } ] ] }, { "title": "Cost", "cells": [ [], [ { "label": "Requests per seat" } ], [ { "label": "Tokens per prompt" } ], [ { "label": "Wasted generation %", "kind": "risk" } ], [ { "label": "GPU seconds per tenant" } ], [ { "label": "Cost per accepted", "kind": "opportunity" } ] ] } ], "note": "Only one row has no error signal behind it: Quality. That is why the control arm is held back permanently rather than per experiment.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "19-model-and-quality-lifecycle", "title": "Model and Quality Lifecycle", "layout": "cycle", "centre": { "label": "Completion quality" }, "rx": 440, "ry": 215, "nodes": [ { "id": "l-reg", "label": "Register by digest", "sub": "immutable" }, { "id": "l-eval", "label": "Offline evaluation", "sub": "gate only" }, { "id": "l-can", "label": "Canary at 1%", "sub": "24 h minimum" }, { "id": "l-meas", "label": "Measure retention", "sub": "against control" }, { "id": "l-dec", "label": "Promote or revert", "kind": "decision" }, { "id": "l-watch", "label": "Watch in production", "sub": "per language" }, { "id": "l-retire", "label": "Retire & reclaim", "sub": "free the working set" } ], "ringLabels": [ "candidate", "allowed", "live to 12 k devs", "effect size", "verdict", "drift", "capacity" ], "note": "The loop closes on retirement, not on promotion: a model left mounted holds working-set memory that the next candidate needs.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "20-trust-zones-and-network", "title": "Trust Zones and Network Segmentation", "layout": "zones", "canvas": { "width": 1760 }, "zones": [ { "title": "Untrusted — the developer's machine", "kind": "trust", "nodes": [ { "id": "t-ed", "label": "Editor process", "kind": "external", "sub": "untrusted client" }, { "id": "t-buf", "label": "Open buffers", "kind": "store", "sub": "never uploaded whole" }, { "id": "t-red", "label": "Local redaction", "kind": "security" }, { "id": "t-tok", "label": "Short-lived token", "kind": "security", "sub": "refreshed" } ] }, { "title": "Public edge — internet facing", "kind": "trust", "nodes": [ { "id": "t-afd", "label": "Front Door + WAF", "kind": "integration" }, { "id": "t-authz", "label": "Seat authoriser", "kind": "security" }, { "id": "t-ver", "label": "Client version gate", "kind": "decision" }, { "id": "t-rate", "label": "Per-seat rate limit", "kind": "integration" } ] }, { "title": "Request plane — private, no inbound internet", "kind": "trust", "nodes": [ { "id": "t-fim", "label": "Prompt assembler" }, { "id": "t-ret", "label": "Retrieval service" }, { "id": "t-post", "label": "Post-processor" }, { "id": "t-filt", "label": "Safety filters", "kind": "security" } ] }, { "title": "Inference plane — no egress, tenant-partitioned cache", "kind": "trust", "nodes": [ { "id": "t-gpu", "label": "GPU node pools", "sub": "egress denied", "icon": "azure_AzureKubernetesService__Containers" }, { "id": "t-kv", "label": "KV cache", "kind": "store", "sub": "partitioned by tenant" }, { "id": "t-sched", "label": "Batch scheduler" } ] }, { "title": "Control and data plane — private endpoints only", "kind": "trust", "nodes": [ { "id": "t-sql", "label": "Policy & seats", "kind": "store" }, { "id": "t-acr", "label": "Model registry", "kind": "store" }, { "id": "t-adx", "label": "Quality aggregates", "kind": "store" }, { "id": "t-akv", "label": "Key Vault / HSM", "kind": "security" }, { "id": "t-aud", "label": "Audit store", "kind": "store", "sub": "append-only" } ] } ], "edges": [ { "from": "t-ed", "to": "t-afd", "label": "TLS, token bound" }, { "from": "t-authz", "to": "t-fim", "label": "authorised, deadline" }, { "from": "t-ret", "to": "t-sql", "label": "private endpoint", "kind": "async" }, { "from": "t-fim", "to": "t-sched", "label": "prompt, tenant tagged" }, { "from": "t-filt", "to": "t-adx", "label": "metadata, scrubbed", "kind": "async" } ], "note": "The inference plane has no egress at all, and returns tokens rather than context: a model that cannot reach the internet cannot exfiltrate what it was given.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } }, { "id": "21-identity-seat-and-policy", "title": "Identity Flow — Seat, Policy and Repository Permission", "layout": "sequence", "canvas": { "width": 1780 }, "lifelines": [ { "id": "dev", "label": "Developer", "kind": "actor" }, { "id": "plug", "label": "Editor plugin" }, { "id": "entra", "label": "Microsoft Entra ID", "kind": "security", "icon": "azure_AzureActiveDirectory__Identity" }, { "id": "edge", "label": "Edge authoriser", "kind": "integration" }, { "id": "cache", "label": "Policy cache", "kind": "store" }, { "id": "sql", "label": "Policy store", "kind": "store" }, { "id": "ret", "label": "Retrieval service" }, { "id": "scm", "label": "Source hosting", "kind": "external" } ], "messages": [ { "from": "dev", "to": "plug", "label": "signs in once", "kind": "call" }, { "from": "plug", "to": "entra", "label": "OIDC device code", "kind": "call" }, { "from": "entra", "to": "plug", "label": "token, 60 min", "kind": "return" }, { "from": "plug", "to": "edge", "label": "suggest + bearer token", "kind": "call" }, { "from": "edge", "to": "edge", "label": "verify signature, extract tenant", "kind": "self" }, { "from": "edge", "to": "cache", "label": "effective policy for seat", "kind": "call" }, { "from": "cache", "to": "sql", "label": "miss — read through", "kind": "call" }, { "from": "sql", "to": "cache", "label": "rules, precedence resolved", "kind": "return" }, { "from": "cache", "to": "edge", "label": "policy, ≤ 30 s stale", "kind": "return" }, { "from": "edge", "to": "plug", "label": "seat revoked or cold cache — refuse", "kind": "error" }, { "from": "edge", "to": "ret", "label": "authorised, with principal", "kind": "call" }, { "from": "ret", "to": "scm", "label": "may this principal read this repo?", "kind": "call" }, { "from": "scm", "to": "ret", "label": "scoped repo list", "kind": "return" }, { "from": "ret", "to": "ret", "label": "retrieve within scope only", "kind": "self" }, { "from": "ret", "to": "edge", "label": "context, provenance tagged", "kind": "return" } ], "note": "Message 12 is evaluated per request, not per index build: an index that outlives a permission change must not serve it.", "meta": { "v": "1.0", "owner": "Developer Experience Architecture", "date": "2026-10" } } ]