Source-linked AI summary
Tuning the Stochastic Machine: A Systems Engineer's Operating Model for Human-AI Engineering
George Andrikopoulos
TL;DR
Corrections to LLM assistants often remain trapped in volatile sessions, leaving a gap in governed accumulation of expert knowledge. The paper maps the LLM stack to operating-system concepts and uses practice cases to derive a discipline centered on persistent, governed corrections and symmetric review.
Problem
Expert corrections and learned error knowledge lack a governed persistent store, while existing practice supplies mechanisms but not the operations discipline to manage them.
Method
The paper maps LLM components to operating-system layers and derives an operating discipline using persistent configuration, recurrence monitoring, provenance, and review.
Results
Practice cases show corrections persisted with provenance, framework gaps surfaced through structured review, and security guarantees treated as hypotheses until validated against intent.
Takeaways & Limitations
Human-AI engineering should treat capability as governed state, persist corrections with provenance, prefer deterministic controls, and monitor tail behavior and drift.
Takeaways & Limitations
The evidence is from one practitioner’s environments and reports no quantitative comparisons, leaving production baselines, cross-team transfer, and independent grading for future work.
Abstract
from arXiv · showhide
When an expert corrects an LLM assistant's error, the correction usually dies with the session, and the error class returns. I argue this is an operations problem, not a tooling problem: mechanisms for persisting corrections exist and are shipping, but the discipline for governing them -- versioning with provenance, recurrence monitoring, counter-metrics, retirement of stale rules -- does not. Writing as a systems engineer of thirty years, I map the LLM stack onto the machines my profession already operates (frozen silicon, firmware, loadable modules, persistent configuration, volatile memory), identify where the mapping fails (stochastic generation, configuration that binds only probabilistically, no general-purpose retirement (verification) stage by default), and derive from the failures a seven-principle operating discipline with an error loop at its core. Three cases from my own practice illustrate the mechanism, among them a control that silently became the exact harm it was built to prevent. I close with the measurement framework this view implies and the lab study required to test it.
1 The reboot problem: capability is state without a store
Expert corrections to AI mistakes typically disappear between sessions, allowing the same costly error to recur. The paper frames this as a state-management and operational-discipline problem: pair capability should accumulate, but prevailing practice measures activity instead.
- The reboot problem: A corrected, subtle, costly mistake recurs a month later in another session and colleague’s chat, made by the same tool with the same confidence.The correction was accepted, but it did not persist beyond the session.
- The reboot problem: The central systems question is where state lives, because nothing malfunctioned when the correction disappeared.The paper treats the issue as a state question rather than a system failure.
- The reboot problem: Human-AI pair capability is inherently stateful and should accumulate knowledge of caught errors and made corrections, yet prevailing practice leaves it without governance.The model and conversation shed state by design, making explicitly rehousing evicted capability necessary.
- The reboot problem: Usage dashboards measure licences active and prompts issued, rewarding activity over capability and making improving output indistinguishable from generating more drafts for review.The paper identifies this as a Goodhart’s-law failure in adoption measurement.
- The reboot problem: The paper presents an experience report and operating model, claiming that missing operational discipline—not tooling—is the needed layer and that systems engineers already possess it.Empirical validation is deferred to the study specified in §8.
2 Related work, and the claim stated precisely
Persistence mechanisms for correcting and extending LLM behavior already exist, but working teams lack the operations discipline to govern them. The paper’s narrow claim is that persistence without governance does not compound, and that this missing layer can be borrowed from operations management.
- Related work: Mechanisms for persistence already exist across memory architectures, agent critique and revision, skill libraries, simulated memory streams, and feedback-based training.Examples include long-term state, cross-episode revision, executable skill artifacts, retrieved memory streams, RLHF, and Constitutional AI.
- The claim stated precisely: The missing capability is operations discipline for governing persistence mechanisms at the level of a working team.The passage states this absence directly, while the surrounding argument specifies what that discipline must govern.
- The claim stated precisely: Governance requires provenance-bearing rules, recurrence monitoring, and explicit recognition of returning errors as drift rather than novelty.Rules should record their date, triggering incident, and rationale so they can be audited and retired.
- Related work: Evaluation should focus on behavioral trajectories rather than release snapshots because systems continue learning after deployment through in-context learning and memory features.This frames the claim from the measurement perspective rather than the operator’s perspective.
- The claim stated precisely: The paper is written from the operator’s seat: persistence mechanisms are not lacking, but persistence without governance does not compound.The missing governance layer is presented as nameable and borrowable from operations management.
3 The machine, in the operator’s terms
The paper models the LLM stack using five machine-room axioms, while identifying critical breaks: natural-language configuration is probabilistic, generation is stochastic, and no general-purpose retirement stage ships by default. From the volatile-session model, it derives a central operational rule: corrections must be persisted as configuration or they disappear at reboot.
- Five axioms: Five axioms map the LLM stack onto machines the author has operated for thirty years.The stack is described through machine vocabulary rather than as an abstract software analogy.
- Five axioms: Every session is a cold boot: prompts and modules load at startup, persistent rules are read from /etc, and conversation remains volatile RAM.The conversation is lost when the session ends, while rules and memory files are intended to persist across boots.
- Five axioms: Natural-language configuration is probabilistic rather than law, except where grammar-expressible constraints can be enforced by constrained decoding.Conflicting instruction files silently widen the output distribution, and prose-only guarantees are unenforced.
- Five axioms: The engine is stochastic, and no general-purpose retirement stage checks speculative work before it commits state.Verifier and critic models, guardrail classifiers, and self-consistency sampling are partial analogues that are valuable where they apply.
- Operational consequence: A conversational correction is runtime sysctl -w: effective immediately but gone at reboot unless written to /etc.The paper argues that organisations currently run expensive expertise in this non-persistent mode.
4 The stochastic engine, and the tail
LLM output is a distribution shaped by sampling configuration, with residual run-to-run variance even at temperature zero. Evaluating average performance hides the tail where harmful failures occur, so the relevant object is the output distribution rather than a single value.
- Sampling configuration controls how much of the probability distribution’s tail is reachable, while batching and routing preserve residual run-to-run variance even at temperature zero.Temperature, nucleus, and top-k truncation shape tail access; jitter is inherent rather than incidental.
- LLM output should be evaluated as a distribution, not an average or single value, because tail behavior determines the system’s damaging failures.The paper frames hallucination as a wild shot in the tail and run-to-run variance as mechanism jitter.
- A model brilliant at p50 but dangerous at p99 can appear fast on average while missing its SLA once a day.The latency analogy emphasizes p50, p99, p99.9, and max as the relevant distributional views.
5 The discipline
The discipline comprises seven principles for making human-AI engineering corrections persistent, layered, measurable, governed, and aimed at real targets. Its core operating loop moves corrections from expert detection into durable artifacts, stronger controls, monitoring, and eventual retirement without losing provenance.
- Persistence and layering: P2 assigns each rule one home across global, domain, and project layers, preventing duplicated domain rules from forking and decaying.Project logic and invariants remain versioned with the repository, while domain rules load wherever their domain appears.
- Probabilistic control: P3 engineers the output distribution rather than seeking determinism, shifting the mean, reducing variance, bounding the tail, and evaluating that tail.Its control hierarchy prefers types, property-based tests, unit tests, constrained decoding, and only then natural-language instructions.
- Persistence and layering: P1 makes every meaningful correction a durable, versioned instruction-layer artifact; runtime-only fixes disappear at reboot.A correction made only in the running workload is effective immediately but does not survive the next session.
- Human retirement and write-back: P4 makes the human the last-resort retirement stage: experts detect and squash undetectable mispredicts, then write fixes back so the pipeline does not replay them.The write-back path makes the human-machine pair compound rather than merely coexist.
- Measurement, governance, and targeting: P5–P7 require counter-metrics, governance against contextual decay, and verification that a measurable target exists before engineering a distribution.Governance includes provenance, recurrence monitoring, review, and a cut list; rules progress from incidents to checks and then an attestable attic, never deletion solely because they have zero fires.
6 The loop
The loop turns every meaningful miss into an operational write-back: fix the instance immediately, abstract recurring errors into rules, and anchor those rules with rationale, checks, incidents, and an appropriate home. One-off corrections remain exempt, while recurring classes require durable treatment.
- 6 The loop: On every meaningful miss, the write-back path begins with fixing the instance before any process or ceremony.The correction must not be delayed by the surrounding process.
- 6 The loop: Name the error class by abstracting until the rule prevents at least two future errors, while exempting genuinely one-off corrections.Every recurring class deserves a rule, but not every correction does.
- 6 The loop: Draft each rule as an imperative carrying its reason, a built-in check for unverified assumptions, and a one-line incident example.The operational check asks under what configuration the rule could cause the exact harm it was meant to prevent.
- 6 The loop: Choose the rule’s home by layer after checking whether a related rule already exists.Layer selection is identified as part of the P2-aligned write-back process.
7 Evidence from practice
Three practice cases show that the error-correction loop is operable in real environments and reveal how failures arise, propagate, and become governed corrections. They illustrate mechanism and use, but do not establish effect sizes.
- Experience report: The experience report presents three cases from the author’s environments as illustrations of the loop’s operation, not evidence of effect sizes.The cases establish that the loop is operable and what it feels like in use; effect-size evaluation is deferred to §8.
- Case 1 — a mechanism chosen by reputation: An io_uring recommendation failed because unsupported descriptors fell back to io-wq workers that could run on the CPU cores the mechanism was meant to protect.The assistant relied on io_uring’s reputation as “async and low-latency” without knowing the target descriptors’ FMODE_NOWAIT support.
- Case 2 — the control that became its own negation: An IP allowlist review caught both inert newline-separated output where JSON was expected and a trailing character that changed /24 into /0.The resulting /0 matched everything, while the malformed handoff would have passed smoke tests without enforcing the allowlist.
- Case 2 — the control that became its own negation: The allowlist case generalized a rule that security-boundary parse results must be checked against intent, while guarantees that have never fired remain hypotheses.The review asked the same failure-mode question prescribed for design rather than checking syntax alone.
- Case 3 — the loop running in both directions: Two episodes show the discipline operating symmetrically: an assistant correction was persisted with provenance, while the author identified and corrected a gap in his own delivery methodology.One correction prohibited estimating biographical facts in deliverables and required explicit placeholders for unconfirmed information.
8 Measurement, and the study this implies
The section proposes four outcome-based, distributional instruments paired with anti-gaming counters, and argues that a single-author setting warrants a provenance-recorded lab study rather than a field study. Its core experimental instrument is the deterministic challenge trial, supplemented by artifact-level library analysis and severity-scaled review measures.
- Measurement: Four proposed instruments measure error-class recurrence, first-time-right rate, rework, and progression from new team member to independent delivery.The instruments are proposals whose validation, including publication of the required severity scale, belongs to the study rather than this paper’s accomplishments.
- Measurement: Each instrument has an anti-gaming counter: rule usage and cut lists, severity-weighted review catches, and escalation quality.Zero recurrence alone can be achieved by a library padded with dead rules, while escalation quality preserves junior learning rather than outsourcing depth.
- Measurement: The measures roll up into time saved and risk avoided, assessing whether the instruction grouping tightens rather than merely counting activity such as opening more chat windows.Usage measures the activity of aiming; the proposed instruments measure whether the grouping is tightening.
- The study this implies: The single-author validation programme is a lab study using a versioned reference implementation with public provenance for corrections, generalisations, and downstream recurrences.The study is not presented as a field study, and its recorded artifacts support analysis of instruction-library evolution.
- The study this implies: Challenge trials provide deterministic binary outcomes through independently constructed exposures, avoiding model-in-the-loop grading, deterrence, and self-grading circularity.The programme also reads rework and first-time-right rates against a published, versioned severity scale.
9 Limits
The paper’s mapping is a working analogy rather than an isomorphism, and its supporting evidence is limited by underexplored instruction reliability and single-practitioner experience. The proposed validation is correspondingly lab-based, with clean provenance and deterministic scoring but reduced ecological validity and several future benchmarks still missing.
- 9 Limits: The LLM-to-machine mapping is a working analogy, not an isomorphism: context windows are only cache-like, and instruction triggering remains underexplored.In-context retrieval degradation is empirically documented, but context windows are not literally RAM.
- 9 Limits: The evidence comes from one practitioner’s own environments, with no quantitative comparisons reported.The author omits personal quantitative comparisons deliberately because of this evidentiary limitation.
- 9 Limits: The lab-based validation program trades ecological validity for clean provenance and deterministic scoring, while production baselines, cross-team transfer, and independent expert grading remain future work.A single-author reference implementation enables the controlled evaluation but does not establish production or cross-team validity.
10 Conclusion
The conclusion frames human-AI engineering as an operating discipline: persist and govern capability-bearing state with provenance, adjust controls after misses, and manage drift rather than relying on memory.
- 10 Conclusion: After every miss, adjust the sights instead of memory, translating marksmen’s practice into a disciplined error-correction loop.The conclusion uses scattering instruments as an analogy for correcting configuration after failures.
- 10 Conclusion: Persist configuration and corrections with provenance so capability compounds under governance rather than disappearing from the system.The paper contrasts evicting state from plumbing with the need for state to live somewhere governed and traceable.
- 10 Conclusion: Systems engineers should persist configuration, prefer deterministic control, watch the tail, and govern drift in production machines.These practices are presented as established systems-engineering discipline applicable to the new machine.