Engineer-citable reference architectures from iSimplifyMe. Each paper documents a production pattern we deploy for clients — private LLM, AWS Bedrock, regulated-industry posture, and the engineering tradeoffs that shape model selection, isolation, and compliance.
Paper Nº 0123 min read
The serving-path experiment crosses a vendor boundary — 5,850 pre-registered calls sending GPT-5.6 (Sol) through OpenAI's first-party API, Amazon's OpenAI-compatible and translated Bedrock doors, and the codex subscription harness in one measured span. Every door emits the same two byte-variants of the same JSON; which one is modal depends on the door — and pinning reasoning effort high flips it behind Amazon's doors while doing nothing on OpenAI's own.
Published 2026-08-23·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 0215 min read
A stealth frontier model appeared on a public aggregator with its operator undisclosed. Differential token counting — 95 probe strings, fourteen candidate vocabularies, 126 API calls, $0 — identified the vocabulary it runs (GLM, 84/95 exact vs 46/95 for the best alternative), the escaping behavior of its serving stack, and the two signals that mark a next generation. The attribution is registered as a prediction before any official claim, and every number regenerates from the public companion repo. A dated follow-up against Zhipu's released GLM-5 vocabulary scores 95/95 exact at zero error, upgrading the vocabulary match from lineage to identity; a second battery then found the chat-template grammar identical to Zhipu's served GLM-5.2, with every runtime divergence localized to one added input-hardening layer.
Published 2026-08-22·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 0318 min read
A five-arm, three-family local benchmark in which every dramatic result was manufactured by a setting: an output cap scored a 122B model at 68% of its capability and inverted the ranking; greedy decoding turned a half-solvable item into a hard 0%; and an infinite reasoning loop proved to be a property of the serving stack, not the weights. The accuracy-versus-budget curve — derived free from recorded token counts and validated to ±2.2 points — plus the triage protocol that caught all three artifacts before publication.
Published 2026-08-11·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 0420 min read
A codebase forked from a sibling carries the parent's identity in layers no review can see — copy is only one of ten surfaces. The residue taxonomy, six design rules for guards that do not depend on anyone remembering, and the fleet-scale enforcement results, with the public check that runs them.
Published 2026-08-05·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 0517 min read
Six companion runs resolve the tension the control-ceiling study disclosed: the long-generation instability was a mixture of three hidden prompt-cache states, each internally byte-deterministic at every sample size measured. Which state serves a call is selected by instance history — a single interposed different-prompt call flipped the output bytes at exactly that call in five of five fresh-instance cycles, under a passing manipulation gate — and the bytes each state produces are additionally scoped to the server session. The result is a state-pinning pattern for byte-reproducible long-form generation, the surface no output validator covers.
Published 2026-07-31·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 0616 min read
A zero-new-calls reanalysis of the reproducibility series: all 33,444 confirmatory outputs from the three studies run through a stated deterministic validator. Parse-and-canonicalize recovers nearly everything on deterministic configurations — every frontier structured-JSON cell closes to exactly 1.0, including a 51-point single-cell recovery — while the residue that survives is semantic and lawful: budget-tier field flips, rare thinking-mode label flips, a 20% object-level disagreement rate under sampling, and one greedy local flip that forks the same ambiguous phrase the same two ways as the budget API model.
Published 2026-07-30·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 0717 min read
The pre-registered baseline under the first two studies — 6,513 calls to open-weight models on owned hardware with every determinism knob pinned. Full control reaches exact byte-reproducibility in 48 of 65 cells, and the exceptions are the finding: the largest model pays an 11-point reasoning-mode cost the smaller models do not, identical weights answer byte-identically across Apple and NVIDIA silicon until generation length breaks the spell, and asking the engine for log-probabilities changes what it generates.
Published 2026-07-30·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 0820 min read
The pre-registered sequel to The Same Request Twice — 13,950 identical requests sending the same Claude models through AWS Bedrock, Claude Platform on AWS, and the first-party Anthropic API simultaneously. The thinking-mode reproducibility cost follows the serving door, not the operator and not (mostly) the model: a null on the first-party API for Opus 5, certified behavioral equivalence between the two AWS-fronted doors, and a forbidden markdown fence whose firing rate moves seventy points with the infrastructure.
Published 2026-07-29·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 0921 min read
A pre-registered measurement of AWS Bedrock inference reproducibility — 13,100 identical requests across three Claude models, two routing scopes, two thinking modes, and three load windows: where byte-level determinism holds perfectly, where it never holds at all, why adaptive thinking turns structured-output serialization into a near-fair coin, and why the budget tier was the only place the answers themselves changed.
Published 2026-07-28·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1019 min read
A measurement architecture for evaluation panels — why a win rate is uninterpretable without position bias and inter-rater agreement, why balance must be guaranteed by construction rather than left to random assignment, and why a directional verdict needs both a Wilson interval that excludes chance and an exact binomial test that reaches it.
Published 2026-07-26·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1115 min read
A production account of four AI agents maintaining a live pipeline through a shared message board — and the argument that a multi-agent system should be bounded by its topology rather than its guardrails: only humans originate work, the agent chain terminates at a node that triggers nothing, and the verification step is deterministic code rather than a second model.
Published 2026-07-25·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1217 min read
A promotion architecture for AI code review — shadow, advisory, soft-gate, hard-gate — where a non-deterministic reviewer earns each rung on measured precision, availability, latency, and false-block rate, with automatic demotion, fails-neutral wiring, and an availability watch that makes silent gate death impossible.
Published 2026-07-23·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1314 min read
The 100-point, seven-section Answer Engine Optimization rubric — gating rules, atomic answer specifications, and score thresholds — as a print-ready white paper. The living standard stays open at its canonical home in the Lab.
Published 2026-07-10·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1417 min read
A reference architecture for controlling production AI cost on AWS Bedrock — prompt caching, per-task model routing, cache-aware routing, cheaper defaults, and spend observability: the cost layer that holds spend flat as usage scales across an organization.
Published 2026-06-28·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1522 min read
A reference architecture for deploying production AI inside HIPAA-regulated workflows, drawn from our work building healthcare AI infrastructure on AWS Bedrock and SageMaker.
Published 2026-05-05·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1628 min read
A reference architecture for the data and retrieval layer of LLM-native AI systems on AWS Bedrock — pipelines, permissioned retrieval, hybrid search, context engineering, memory, and feedback loops — drawn from iSimplifyMe production deployments in regulated and mid-market work.
Published 2026-05-06·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1718 min read
A reference architecture for the reliability layer of LLM-native systems on AWS Bedrock — layered guardrails, atomic content integrity, investigate-only audit agents, circuit breakers, retries, and quality gates — the engineering that decides whether a deployed AI system holds up in regulated production or decays into a demo.
Published 2026-06-28·Joseph W. Elstner·Founder & Principal Architect
Paper Nº 1817 min read
A reference architecture for the business-integration layer of an LLM-native platform on AWS — single-table multi-tenancy with isolation by construction, domain-routed tenant resolution, a unified lead pipeline, role-permissioned dashboards, and synchronized billing — the layer that turns AI capability into a product many clients run on one platform.
Published 2026-06-28·Joseph W. Elstner·Founder & Principal Architect
