--- title: Plumbob — Execution Infrastructure for AI-Native Products description: Plumbob LLC builds reusable distributed execution infrastructure for AI-native software and edge systems like Cremaflow. An early-stage company based in Illinois, United States. --- # Plumbob LLC > Grounding AI in deterministic execution. > Build the Product. Plumbob owns the execution machinery. Products submit work through one durable boundary. Plumbob handles immutable inputs, model execution, recovery from uncertain outcomes, spend limits, provenance, cancellation, and erasure—and is being extended toward multi-step programs, grounded research, independent evaluation, and controlled actions. ## Quick Facts - **Execution Kernel:** Cloudflare Workers + Workflows - **Storage & Audit:** D1 (Audit & State) + R2 (Content-Addressed Artifacts) - **Frontier AI:** Claude via Pinned WorkloadReleases - **Company Stage:** Early Stage · Pre-Commercial (Illinois, United States) - **Approved Company Contact:** hello@plumbob.ai - **Primary Domain:** https://plumbob.ai --- ## The Platform Thesis **Products own meaning. Plumbob owns execution.** > "Products own meaning. Plumbob owns reusable mechanism. Cloudflare owns infrastructure mechanics." ### Plumbob Cloud (The Platform) Product-agnostic execution infrastructure that turns arbitrary AI interactions into bounded, durable, and auditable units of work. - **Tenant-Bound Machine Identity:** Scoped credentials isolate machine callers without cross-tenant data leakage. - **Durable Workflows:** Orchestrates multi-step jobs across interruptions with auditable terminal states. - **Atomic Spend Governance:** Monthly budget ceilings verified in micro-USD before inference dispatch. - **Lineage-Aware Erasure:** Purges input artifacts and all derived outputs through fenced cascades. ### Cremaflow (The First Product) An in-development reference Product for real-world physical telemetry, sensor fusion, and deterministic replay. - **Deterministic Local Core:** High-frequency BLE sensor loop (scales, pressure, flow) operates 100% offline. - **Frozen Evidence Bundles:** Packages physical extraction telemetry into SHA-256 hashed immutable artifacts. - **Claude Advisory Workloads:** User-triggered shot analysis and recipe generation executed through Plumbob. - **Deterministic Advisory Validation:** Rejects or quarantines model claims that conflict with trusted sensor telemetry. --- ## What Plumbob Is Building Toward Plumbob is being built so Products can submit **intent** rather than orchestrate infrastructure. The platform owns the distributed execution required to fulfill that work: immutable inputs, pinned model and workload releases, durable recovery, bounded spend, provenance, cancellation, and erasure. Above that execution kernel, Plumbob is developing reusable higher-order capabilities for multi-step AI systems: 1. **Execution Envelopes** *(Idempotent Kernel)*: A Product submits immutable input under one idempotency identity. Plumbob owns Artifact admission, workload execution, retries and recovery, provider attempts, spend limits, cancellation, and erasure underneath it. 2. **Released Multi-Step Programs** *(Orchestration Graphs)*: Products can compose immutable Workloads, search operations, and evidence acquisition into bounded directed acyclic graphs (DAGs) without turning the model itself into an unrestricted runtime orchestrator. 3. **Grounded Research & Evidence** *(Provenance Capture)*: Search discovers candidate sources; Plumbob retrieves exact source bytes itself, records cryptographic provenance, and distinguishes model claims from evidence that the platform actually observed. 4. **Independent Evaluation** *(Separation of Concerns)*: A Product can use separately released Workloads to evaluate outputs produced by other Workloads, preserving both results and their provenance rather than asking the same model to self-certify. 5. **Bounded Adaptive Execution** *(Controlled Loops)*: Future loops and repair paths have declared iteration, operation, time, and spend ceilings. A model may select among released choices; it cannot invent new runtime authority. 6. **Controlled Effects** *(Consequential Actions)*: Longer term, Plumbob can turn consequential external actions into immutable commands with explicit authorization, idempotency, uncertain-outcome reconciliation, and optional human release. --- ## First Product Integration — In Development: Cremaflow Cremaflow is the first Product being built to consume Plumbob's execution platform, bringing nuanced reasoning to espresso telemetry without compromising local deterministic truth. **Claude is the first frontier model integrated into Plumbob's live execution kernel.** Cremaflow is the first Product being built to consume that capability. Its espresso telemetry, replay, analysis, and live guidance remain deterministic and local. When the user explicitly requests AI analysis, Cremaflow freezes the relevant evidence, sends that immutable request through Plumbob, and treats Claude's result as provenance-tracked advisory output subject to deterministic Product validation. ### Architecture Chain ``` Deterministic physical system → Immutable evidence → Bounded Claude execution → Deterministic validation → Advisory user experience ``` ### System Specifications - **Local Runtime:** Docker · Python 3.14 · SQLite - **Hardware Loop:** ESPHome BLE · 50 Hz Telemetry - **AI Workloads:** Session Review · Recipe Drafts (M7C) - **Advisory Safety:** Deterministic Advisory Validation ### Frozen Evidence Bundle Session telemetry + deterministic analysis + recipe/equipment context: - **SHA-256 bound before egress:** User consent durably binds to the exact byte digest. Replays with modified content are rejected. - **Exact bytes become the AI request boundary:** The model sees only the admitted evidence value, with no access to SQLite or physical devices. - **Lineage-aware lifecycle:** Calling delete on this bundle cascades through all derived model outputs across platform storage. ### Model Advisory - **Model:** Claude via pinned Plumbob WorkloadRelease (Strict Advisory) - **Measured Pattern (High Confidence):** Puck resistance decay accelerated between t=18s and t=24s, indicating late-stage erosion without catastrophic channeling. - **Comparison (Medium Confidence):** Flow velocity remained 0.3 ml/s higher than the 14-day washed Gesha baseline, suggesting grind distribution is slightly coarse for target extraction yield. - **Actionable Experiment:** Adjust grinder setting 1 step finer; maintain 18.0g dose to target 34.5g yield at 31s extraction duration. ### Deterministic Advisory Validation Deterministic advisory validation rejects or quarantines model claims that conflict with trusted telemetry or exceed the advisory contract: - Resistance Decay: Verified by flow & pressure derivative → Admitted (High Confidence) - Grind Setting Step: Valid within grinder step resolution → Admitted (1 Step Finer) - Temperature Shift: Forbidden without physical sensor probe → Expunged (Conflict Guard) Model outputs can never alter canonical SQLite session facts or execute physical machine actions without explicit user acceptance. --- ## How Plumbob Executes Work Execution Sequence: ``` Product Submits Work → Admits Immutable Input → Execution Survives Failures → Claude Runs Inside Bounds → Result & Provenance Return ``` 1. **One execution boundary:** Products submit work instead of orchestrating storage, runs, and recovery. A Product submits intent under one tenant-bound credential. Plumbob derives authority, enforces rate and flood protections, and admits immutable inputs without leaking underlying storage or machine mechanics. 2. **Durable by default:** Accepted work survives request loss, worker recycling, and infrastructure interruptions. Execution is orchestrated across stateful lifecycle steps that guarantee every job reaches an auditable terminal state. If a client disconnects, the work continues to completion or until explicit cancellation. 3. **Frontier models under bounds:** Models, spend, and structured outputs are governed by immutable releases. Live inference is executed under pinned model policies with pre-dispatch monetary spend reservations in micro-USD. Provider attempts have a retry limit of zero, preventing double-billing on uncertain network outcomes. 4. **Traceable results:** Inputs, execution identity, provider usage, and outputs retain write-once provenance. Outputs are strictly validated against released schemas before delivery. Every result records exact provider usage, token costs, model identity, and input digests in append-only audit storage. 5. **Lifecycle control:** Cancel work without destroying state; revoke when content needs to be erased. Cancellation stops queued and subsequent steps while safely reconciling in-flight calls. Lineage-aware cascading erasure purges inputs, raw outputs, and downstream derivations across storage with fenced database markers. --- ## Non-Negotiable Engineering Guarantees - **Atomic Spend Reservations:** Worst-case cost reservations before provider dispatch. Every live execution atomically reserves estimated spend in micro-USD against Product and Tenant monthly ceilings before invocation. If a request cannot fit within budget, it fails immediately without reaching the provider. Retries have a strict limit of 0. - **Exact-Bytes Authorization:** Cryptographic binding to SHA-256 byte digests. User egress consent authorizes the exact immutable admitted body hash, selected privacy classes, and actor. Attempting to upload altered bytes under the same authorization triggers an immediate conflict error. - **No Hallucinated Authority:** Advisory outputs downstream of canonical truth. Probabilistic AI outputs are strictly advisory. A model can never mutate canonical domain state, edit SQLite journal tables, publish recipes automatically, or trigger physical machine actions without explicit user review. - **Deterministic Foundation First:** 100% offline edge operability. Physical sensor acquisition, deterministic extraction algorithms, and live guidance run entirely local-first. Cloud infrastructure or external model provider outages never disrupt instrument operation or brewing. - **Lineage-Aware Fenced Erasure:** Fenced deletion markers across storage. Tenant content is retained only until erased. Calling DELETE /v2/artifacts/{id} triggers a cascading purge across all raw model outputs and derivative results in R2, leaving only immutable D1 audit metadata. - **Model Pinning & Strict Schemas:** JSON Schema 2020-12 enforcement. Execution policies pin exact provider model identifiers, gateway parameters, and strict JSON Schema definitions. Workloads reject silent prompt drift, non-conforming responses, and unverified model changes. --- ## Built for Frontier-Model Products **Plumbob LLC** is an early-stage company established in 2025, based in Illinois, United States. We design and build execution infrastructure for AI-native software and edge systems. **Early-Stage Transparency:** Plumbob Cloud and our initial products (including Cremaflow) are in active internal engineering. We are pre-commercial and under active development, focused entirely on distributed-systems rigor, deterministic validation, and safety. ### Why Claude Matters to Our Architecture Deploying frontier models like Claude in real-world applications requires moving from “respond to a prompt” into software infrastructure: - **Deterministic Evidence Boundaries:** Claude receives immutable evidence bundles extracted from ground-truth local systems, never direct or unmetered database access. - **Bounded Execution & Structured Outputs:** Pinned releases, strict JSON Schema 2020-12 enforcement, and pre-execution monetary spend caps prevent runaway token consumption and prompt drift. - **Downstream Advisory Validation:** Model outputs are treated as probabilistic hypotheses, verified by product-owned deterministic rules before reaching users or physical actuators. --- ## Company & Contact Information - **Legal Entity:** Plumbob LLC - **Jurisdiction:** Illinois, United States - **Primary Domain:** https://plumbob.ai - **Approved Company Address:** hello@plumbob.ai - **Website Terms:** https://plumbob.ai/terms - **Privacy Policy:** https://plumbob.ai/privacy