Gateway observations
472.0k
Observed at the HEARTH boundary
The studio lab · published operating evidence
Find the constraint.
Complete the circuit.
Leave a trail.
I build working systems with AI and the people who understand the work. We turn intent into experiments, keep the evidence, and use what we learn to improve the next decision.
The aim is to expand what people can do while keeping ownership and judgment visible. This is an active studio and lab; its snapshots and field reports keep their dates, limits, and corrections attached. See how the experiments changed the work →
Hands-on software architect and technical lead with more than two decades building and modernizing production systems in .NET and Azure.
checksummed aggregate
bounded fixture
recorded evidence
approved career claim
A result should change the next decision.
LINUX FIELD NOTES · SEPTEMBER 24–OCTOBER 4, 2026Recent JEPA research exposed a shortcut behind an apparent perfect score. Later thinking experiments showed that a report's final rewrite could discard its best reasoning. Those findings changed the tests, the records we keep, and the way the workflow delivers a model's answer. My direction, the agents' work, and the limits of the results stay visible in the account.
Follow the intent, experiments, and decisions →Execution requests admitted
31,485
Entered the workflow; not an output-review verdict
Artifacts recorded
67,105
Inputs and results with receipts
Model calls
23,876
Research + other use, through the gateway
Public repos
50
GitHub profile · 2026-10-07
Published ADRs
69
48 commandcenter + 21 lumberjacks · 2026-10-07
Hardware field lab · the measured system beneath the system
29 INDEXED LABS · MEASURED FIELD REPORTPublished receipt 01 · dual Intel Arc Pro B70
A model file is not a service.
Twenty-nine indexed local-model labs plus the August rebuild and Qwen 3.8 campaigns sit behind this panel. The published aggregate includes 11,388 research inference attempts outside the gateway, captured in server logs and dated run records. The report preserves its measured operating points and dates. The old conclusions were rerun, and one apparent bottleneck became three. The useful proof is the chain from workload to constraint to changed decision.
Causal result
19.7 → 131.5 tok/s
One Vulkan dispatch constant moved the ninth-request result by 6.7×.
Operating point 0.69× → 5.49×The “slower” model reversed the verdict as real prompt depth increased.
Whole system 847 waves · 0 resetsServing held through concurrent inference and disk traffic; the box survived the claim.
Service choice 95.2 vs 14.7 tok/sThe largest model that fit was not the most useful model to keep serving.
HEARTH pulse · research and recorded activity
WITHHELD CELLS MARKED · DATED SNAPSHOTPublished aggregate through 2026-10-07. Integrity checked before release; this is a dated view of ongoing work, not live service health.
Much of this local-model activity is research: testing model fit, improving the harness, and learning what support helps models succeed. Research runs through the gateway as well as outside it. These counts show how activity was recorded; they do not separate research from everyday use.
Observed calls by fixed family · logarithmic scale
Status checks count as operations. Workflow control covers submissions, planning, reviews, cancellations, and result retrieval; it does not count completed model work. Other retains calls without an assigned category.
The scale is logarithmic so model calls and other work remain visible beside polling and health automation. Hollow: research attempts recorded outside the gateway, harvested from server logs and run records. Gateway model-call rows also include research.
Weekly observations by category · shared logarithmic scale
Corrected October 7: categories now have independent bars on a shared scale. Weekly counts from 1 through 9 and selected additional cells are withheld to limit what totals reveal. Earlier published values and comparisons between releases can still reveal counts.
Local work · review outcomes
Review outcomes are unavailable until the published snapshot is loaded.
A successful execution can still await review or be rejected. Execution failures remain separate. Research can involve repeated attempts as the harness and model support improve; these outcomes do not count independent benchmark trials or human decisions.
Direction and review
A dated decision record distinguishes Derek's scoped delegation, decisions made by Codex under that delegation, and the later review of model outputs.
Documentary evidence of direction is available. Verified human-activity counts are unavailable in this feed; agent decisions are not counted as human actions.
Read the direction and review note →System topology · roles and boundaries
MACHINE ROLES ABSTRACTED · RAW CONTENT WITHHELDCALLERS
Tools + agents
Typed operations enter through one observable boundary.
DOOR
HEARTH gateway
Admission, request identity, validation, and receipts.
CONTROL
Policy + idempotency
Retries stay attempts beneath one durable job.
COMPUTE
Local models + workers
Capability, capacity, health, and policy select a local lane.
TRUTH
Append-only ledger
Events and content-addressed artifacts survive workers.
PUBLIC
Safe projection
Aggregates are checked, suppressed, staged, and reviewed.
Guided trace replay · one job, one retry, one durable result
CURATED FIXTURE · FABRICATED IDS · NO OPERATIONAL CONTENTThis trace is an honest fixture. The state machine mirrors the released execution contract; identifiers and event content are fabricated. Use delayed delivery and duplicate injection to test the consumer.
Job statenot started
Attempts0 / 0 failed
Artifacts0
Buffered0
Working range · four routes through the evidence
DEREK FIRST · CLAIMS FROM THE REVIEWED CAREER FEED
ROUTE 01
Runtime + hardware
Measure the service, not the box.
Built a self-hosted local inference environment on Intel Arc hardware, working across model runtime, GPU memory, Vulkan compute, and Windows driver constraints.
Intel ArcVulkanllama.cppmeasured labs
Recorded evidence
Follow route →
ROUTE 02
Distributed state
Identity, leases, replay, suppression.
Built and physically tested a bounded replacement for selected native networking and replication paths in a Unity/Valheim stack using BepInEx and Harmony; retained receipts cover durable replay, logical peer identity, ownership leases, idempotent delivery, and native-path suppression across a dedicated server and two clients.
UnityBepInExHarmonyValheim
Recorded evidence
Open receipt →
ROUTE 03
Agent control plane
Durable work beneath changing models.
Building and testing greenfield .NET 9 multi-agent systems with durable messaging, explicit workflow stages, and distributed tracing.
.NET 9OpenTelemetrydurable jobsreplay
Recorded evidence
Follow route →
ROUTE 04
Enterprise identity
Modernize across real team boundaries.
Led enterprise customer-identity consolidation and Auth0 externalization across three engineering teams at a consumer-credit platform.
CIAMAuth0Azurecross-team delivery
Recorded evidence
Career context →
The domains change. The engineering shape holds.
| Mechanism | HEARTH control plane | Community runtime work | Enterprise platforms |
|---|---|---|---|
| Stable identity | Request, job, invocation | Logical peer and object identity | Customer and service identity |
| Duplicate safety | Idempotency key + attempt lineage | Idempotent delivery + suppression | API and event contract boundaries |
| Eventual truth | Ledger → derived projection | Replay → converged runtime state | System of record → channel views |
| Failure evidence | Append-only events + artifacts | Packet and lifecycle receipts | Telemetry, ADRs, audit boundaries |
| Handoff | Schema, runbook, reproducible replay | Bounded mechanism notes | Operable platform + team ownership |
Receipt wall · inspect the work behind the claims
REPORTS + SOURCES · CORRECTIONS STAY ATTACHEDDOCUMENTARY SUMMARYDirection
Direction, delegation, and output review
A dated Linux workflow note keeps operator authorization, delegated agent decisions, and acceptance of model outputs distinct.
FIELD REPORTMeasured
What Consumer Hardware Actually Does
Twenty-nine indexed labs, a hardware rebuild, corrected constraints, full tables, and explicit unmeasured boundaries.
UPSTREAM PATCHCode
llama.cpp pull request #27652
A concrete upstream artifact from investigating Vulkan inference behavior and GPU-memory admission.
SOURCE + ADRSBuild
HEARTH in commandcenter
The gateway, contracts, execution ledger, projections, runbooks, and decisions behind this released aggregate.
BUILD GUIDEMethod
Build Your Own HEARTH
The earlier Windows + NVIDIA tutorial: a small gateway, a tool-call ledger, and step-by-step checks.
FIELD LABTested
Native network replacement landscape
Durable replay, logical peer identity, ownership leases, idempotent delivery, and native-path suppression tested across three processes.
ARCHITECTURETrace
MAF + OpenTelemetry dispatch stack
A message-driven agent workflow whose stages, failure paths, and correlations are visible instead of implied.
DECISION RECORDProtocol
Five-byte input payload
One public decision record shows the representation, comparison boundary, and why the smaller protocol was chosen.
LONGITUDINALContext
Twenty years, one operating method
Identity, healthcare messaging, factory systems, product platforms, and local AI—organized around the same inspectable method.
PUBLIC INDEXWriting
Engineering field reports
Failures, reversals, decisions, and build notes written for people who want enough detail to disagree.
Operating method · how the receipts get made
REPEATABLE ACROSS DOMAINSStart where work and state actually cross boundaries, not where the diagram says they do.
Measure where reality stops matching the dashboard and preserve the failed assumptions.
Prototype the difficult path, then stay through retries, telemetry, cutover, and ownership.
Ship the schema, decision record, runbook, trace, and evidence the next engineer can replay.