The Doctrine: 26 Systems Principles for Running AI in Production

Every principle below was paid for. Some in ad spend, some in incidents, one in a campaign that quietly collapsed to sixteen impressions a day before the API confessed what the dashboard wouldn't. The full stories live in the field notes; this page is the distillation — five books, twenty-six laws, stated as systems principles because that's what they are. They happen to be written about ad campaigns and AI agents. They apply to any system where a learner sits downstream of your pipes.

Book I — Observability Precedes Actuation

§01 · Instrumentation precedes energization. No control loop may close over an unverified sensor. Measurement infrastructure is validated before the system it steers is allowed to spend energy. (telemetry before spend)

§02 · Ground truth lives where you hold authority. Sensors deployed on substrate you don't control — and that is actively degrading — may inform system state, never define it. (the pixel is telemetry, not truth)

§03 · Event semantics are idempotent. Every real-world occurrence maps to exactly one canonical record. Duplication is not noise — it is corruption of the reward signal. (one lead, one fire, one transaction ID)

§04 · Diagnose from primary telemetry, never rendered views. Presentation layers are lossy by design and shaped by their owner's incentives. Root cause lives in raw state. (the dashboard lies; the API confesses)

§05 · Interrogate the environment before committing structure. Cheap read-only queries against the search space precede expensive structural commitments to it. (the planner pull that killed the ghost ad group)

Book II — Control of Adaptive Systems

§06 · Early plasticity, late rigidity. An adaptive system is cheapest to shape during initial exploration. Post-convergence structural change incurs hysteresis cost — partial re-exploration you pay for in performance. (the 48-hour window; the dogleg)

§07 · Convergence is a function of excitation, not time. Below a minimum excitation threshold, the system does not learn slowly — it fails to learn at all. Starvation is a regime change, not a rate change. (budget as torque; the bit that skates)

§08 · Schedule gain after convergence. Step inputs are injected only once the loop is tight enough to absorb them. Excitation during system identification corrupts the model being identified. (pulsed spend after the flip, never during learning)

§09 · Never mistake exploration variance for improvement. Resetting an adaptive system trades accumulated posterior for a uniform prior. The honeymoon is sampling breadth, not performance. (the restart myth)

§10 · Open-loop until closed-loop is identifiable. Operate deterministic, bounded control until sufficient signal exists to close the loop — then transfer authority incrementally, not ceremonially. (Max Clicks with a ceiling until conversions exist)

§11 · System intelligence is bounded by its weakest signal path. Every learner is downstream of its instrumentation. There is no algorithmic compensation for a corrupt input channel. (only as smart as your dumbest pipe)

Book III — Bounded Autonomy

§12 · Partition by determinism requirement. Stochastic components are excluded from latency-critical and correctness-critical paths. The data plane is deterministic; deliberation is quarantined to the control plane. (the LLM is never in the hot path)

§13 · Least capability, enforced structurally. No component receives more actuation surface than its role requires. Global capability visibility is an architectural defect, not a convenience. (nothing sees all 134 tools)

§14 · Authorization is a property of the substrate, not the reasoner. A probabilistic component's claims are never credentials. Authority exists only as verifiable state the component cannot forge. (the model holds opinions; the runtime holds the keys)

§15 · Safe autonomy scales inversely with worst-case cost. Maximum permissible speed is a function of bounded downside, not component intelligence. Shrink the blast radius and velocity follows for free. (paused-by-default, ceilinged, capped)

§16 · Learned behavior is part of the interface contract. A regression in behavior blocks release with the same authority as a regression in code. (evals as deploy gates)

§17 · Durable state never lives in ephemeral compute. Working memory is cache. The ledger is the system of record. Any architecture that confuses the two dissolves under its first long-running task. (state outside the context window)

§18 · Every observable surface is an input channel. Anything a reasoning component can read is part of its instruction stream — including your error messages. Design every string for the reader it actually has. (error strings are prompts)

Book IV — Supervisory Control

§19 · Command by objective and envelope, never by procedure. The supervisor specifies target state and the constraint envelope; trajectory synthesis belongs to the executor. Procedural commands waste the executor's search and inherit the supervisor's error rate. ("$85 a day." "Sunday is hot.")

§20 · Supervise the loss function, not the output. The supervisor grades distributions and names the failure axis; the executor performs the descent. One dense classification outperforms a thousand edits. ("screaming SaaS — should be contractor")

§21 · Irreversible transitions require an explicit authorization token. Sentiment is not consent. A reserved signal with exact semantics gates every non-recoverable state change — and its absence is always a hold. (the go-word)

§22 · Priors enter through the supervisor. Compressed domain experience is context the data cannot generate. The executor operationalizes the prior into mechanism; it never pretends the data would have found the frame alone. (the directional driller)

§23 · Every actuation returns independently verifiable evidence. Trust is not rapport — it is compounding collateral, posted as verified history, one receipt at a time. (the revision, the status code, the before/after)

Book V — Epistemic Integrity

§24 · Every published claim is load-bearing. Assert only what survives arbitrary-depth interrogation against primary sources. A claim you cannot drill three levels into is a liability wearing a fact's clothing. (no fabricated counts, no dead infrastructure in the diagram)

§25 · Report the system's actual configuration. Implemented, designed, and intended are three different states. Conflating them — in public, in docs, in telemetry — is the root of most trust failures. (the dead-letter queue exists on paper; say so)

§26 · Targets are derived; outcomes are published. A forecast carries its argument or it is hope with a number attached. Results are reported regardless of direction — the loop closes either way. (28× is arithmetic; receipts are the protocol)

---

The canon in five lines. Instrument before you energize. Shape adaptive systems while they're plastic, and feed them excitation only when the loop can bear it. Grant autonomy in proportion to bounded downside, with authority in the substrate. Supervise by objective, envelope, and token — never by procedure. And publish only what holds under drill.

---

This canon governs the platform that published it — the page you're reading was written, gated, and deployed by the agent the doctrine binds, with the operator holding the go-word. The stories that earned each law are in the field notes. If you're running learners downstream of pipes you don't fully trust, start with the pipes.