Skip to content

SYS-02 · System

Content Engine — the publication pipeline.

Notebook → outline → draft → review → publish, with traces at every step. The system behind every article on this site.

System code SYS-02Version v2.0Phase OperationalOwner Solo
§ 01Opening

The content engine has one non-negotiable design constraint: the human review gate cannot be bypassed. Every component upstream of the gate — the outline graph, the draft model, the critique loop — exists to deliver higher-quality input to that gate, not to eliminate the need for it. A pipeline that produces publishable content without human review is not a content engine. It is a content firehose, and the publication's quality guarantee is the first casualty.

The pipeline has five stages: outline, draft, critique, review, publish. Each stage has explicit contracts — the outline generates section specs, the draft model writes against specs, the critique model evaluates against criteria, the review gate exercises editorial judgment, and publish routes simultaneously to the publication surface and the retrieval store. Tracing is present at every stage boundary — the pipeline produces a structured record of what was outlined, drafted, critiqued, revised, and approved for every piece that ships.

The most operationally significant finding from running this system is that tracing the writing process changes what gets written. When every outline decision, draft pass, and critique revision is logged, the cost of bad structural decisions becomes visible — not as a vague sense that the piece is hard to write, but as a measured count of revision cycles that trace back to an underspecified section spec. The trace is not just observability infrastructure. It is feedback that improves the outline layer.

§ 02Architecture

Five layers in sequence with one feedback arc. The outline layer produces a directed graph of section specs. The draft layer writes each section against its spec with full outline context. The critique layer evaluates the draft against explicit criteria and loops back to draft for targeted revisions. The review gate is the single human checkpoint — its approval is the only path to publish. The publish layer routes simultaneously to the publication surface and the retrieval store.

OUTLINEgraphOUTDRAFTmodelDFTCRITIQUEloopCRTREVIEWhumanRVWPUBLISHdual outPUBREVISION LOOP◆ HUMAN GATEPUB SURFACERETRIEVAL STORESIMULTANEOUSREJECT → OUTLINE

Fig. SYS-02 — Content Pipeline Stage-Gate · Outline → Draft → Critique → Review → Publish · critique feedback arc · dual publish output

§ 03Outline Graph

Article structure is represented as a directed graph, not a linear list. Each section is a node. An edge from section A to section B means section B's argument depends on section A having been established — B cannot be drafted until A is complete. Linear outlines hide these dependencies by treating sequence as a proxy for dependency. A section that appears third in a linear outline might depend on sections one and five — a relationship that a linear list cannot represent and that a draft model cannot infer.

OUT-01 builds the graph from the research question and the existing cross-link context — which frameworks this piece should reference, which research pages it should connect to, what the reader is assumed to know. The graph structure determines the drafting order: topologically sorted, starting from nodes with no dependencies and proceeding to nodes whose parents are complete.

OUT-02 generates a section spec for each node before drafting begins. The spec is the contract the draft model writes against: the claim the section must establish, the evidence or reasoning required, the target cross-links, and the word count range. A section spec that is vague produces a draft that requires more critique revisions. The trace makes this cost visible — underspecified specs correlate directly with revision count in the pipeline logs.

CodeNameStatusDescription
OUT-01Outline Graph BuilderStableConverts a research question or topic into a directed graph of sections, where each node is a section and edges represent dependency or sequencing relationships. A section that requires another section's argument to be established first is a dependent node — it cannot be drafted until its parent node is complete. The graph structure prevents outlines from becoming linear lists that hide ordering assumptions.
OUT-02Section Spec GeneratorStableFor each node in the outline graph, generates a section specification: the claim the section must establish, the evidence or reasoning required, the cross-links it should carry, and the word count target. The spec is the contract between the outline layer and the draft layer — the draft model writes against specs, not against free-form prompts.
§ 04Draft Model

DFT-01 drafts each section with the full outline graph and all previously completed sections in context. The long-context window is the critical property — a draft model that sees only the current section spec produces locally correct prose that contradicts earlier sections or sets up arguments it cannot later deliver. Full-context drafting allows the model to maintain argument continuity across the entire piece, not just within the current section.

Voice constraints are enforced as system-level instructions held constant across every drafting call. They are not per-section prompts — they are publication-level requirements that apply to every word produced. The constraints are the same ones that govern this publication: editorial without academic, dense without inaccessible, anti-hype without contrarian, no startup clichés, no motivation writing. DFT-02 runs a post-draft pass that flags violations specifically rather than relying on the draft model's self-enforcement.

The draft model does not make structural decisions. If a section spec requires an argument that is not supportable with available evidence, the draft model produces the best argument it can and the critique model flags the gap. Structural corrections — adding a section, merging two sections, revising a claim — go back to the outline layer, not to the draft layer. Keeping the layers' responsibilities distinct is what makes the critique loop's targeted revision possible. The evaluation methodology that informs the critique loop's design is investigated in RES/EVA.

CodeNameStatusDescription
DFT-01Long-Context Draft ModelStableDrafts each section against its spec with the full outline graph in context. The long-context window ensures the draft model can see how the current section relates to sections already written and sections yet to come — preventing contradictions and ensuring argument continuity across the full piece. Voice constraints are passed as system-level instructions, not as per-section prompts.
DFT-02Voice Constraint EnforcerIn reviewA post-draft pass that checks the drafted section against the publication's voice constraints: editorial without academic, dense without inaccessible, anti-hype without contrarian, no startup clichés, no motivation writing. Flags violations with the specific constraint breached and the offending passage. Does not rewrite — flags only. Rewriting is the draft model's job on the next pass.
§ 05Critique Loop

The critique model evaluates the full draft against explicit acceptance criteria. The criteria are not style preferences — they are structural requirements: every claim supported, every cross-link load-bearing, the opening claim in paragraph one, the closing not a recap, all section specs satisfied. The critique returns a structured result: pass/fail per criterion, the specific passage that failed for each failing criterion, and a revision instruction targeted to that passage.

CRT-02 dispatches one revision instruction per failing criterion back to DFT-01. The loop runs until all criteria pass or until three revision cycles complete without full pass. At the three-cycle limit, the draft routes to the human review gate with the remaining failing criteria flagged explicitly — the reviewer sees exactly which criteria the automation could not resolve. This is more useful than routing an unmarked draft that passed critique by relaxing criteria.

The three-cycle limit is a known approximation. Criterion failures that trace to the outline graph — an unsupportable claim, a section that cannot establish its required argument — will not resolve through draft revision. They require outline revision. The current system detects this case only implicitly (repeated failure of the same criterion across multiple revision cycles) rather than by explicit classification of failure type. This is the critique loop's most significant operational gap.

CodeNameStatusDescription
CRT-01Critique ModelStableEvaluates the full draft against explicit acceptance criteria: every claim is supported, every cross-link is load-bearing, the opening claim is in paragraph one, the closing does not recap, the section specs are satisfied. Returns a structured critique with pass/fail per criterion and the specific passage that failed for each failing criterion.
CRT-02Revision DispatcherStableReceives the structured critique and dispatches targeted revision instructions back to DFT-01 for each failing criterion. One revision dispatch per criterion — not a single bulk rewrite request. The loop terminates when all criteria pass or when the maximum revision count (currently three) is reached without full pass. At max revision without pass, the draft routes to human review with the failing criteria flagged explicitly.
§ 06Human Review Gate

The review gate is the only component in the pipeline that cannot be automated. It exercises three classes of judgment that the critique model cannot reliably apply: editorial judgment above the threshold the voice constraints define, factual verification for domain claims that require expertise rather than pattern matching, and publication standard fit — whether this piece earns its place in the architecture by being load-bearing rather than merely correct.

The gate produces one of three outputs. Approval routes the draft to publish. Targeted revision routes specific passages back to DFT-01 with explicit instructions — not a request to "improve" the section, but a precise instruction about what claim needs to change and why. Rejection routes back to the outline layer when the structural premise of the piece is wrong — when no amount of drafting or revising will make it publishable because the question it answers does not warrant a page.

The operational finding about loosening the gate is the most important operational note in this system. Every time the review criteria were relaxed to reduce pipeline latency, the total time to publish increased — because the gap the gate would have caught appeared later, as a structural problem visible after the piece was live rather than before it shipped. The gate is not a bottleneck in the publishing pipeline. It is the primary reason the pipeline produces publishable content rather than publishable-looking content.

CodeNameStatusDescription
RVW-01Human Review GateStableThe single mandatory human checkpoint in the pipeline. Receives the draft after critique passes (or after max revision with explicit failure flags). Reviews for: editorial judgment that automation cannot exercise, factual claims that require domain verification, cross-link load-bearing assessment, and publication standard fit. The gate produces either an approval, a targeted revision instruction, or a rejection back to outline. Approval is the only path to publish.
§ 07Publish + Index

On review gate approval, the publish layer executes two operations simultaneously: PUB-01 routes to the publication surface and PUB-02 routes to the retrieval store. Simultaneity is a design requirement, not a convenience. A page that is publicly readable but not yet indexed in the retrieval store exists in two inconsistent states — it is part of the publication's linked architecture but not yet part of the second brain's knowledge surface. The inconsistency window is the gap between publication and indexing. Simultaneous execution eliminates the window.

PUB-01 generates the route, populates the page metadata, and registers the page in the cross-link index. It triggers the dead link checker against all outbound links in the published page — a link to a page that does not yet exist is caught at publish time, not discovered by a reader. This is the operational enforcement of the dead link pressure principle that governs the publication's build priority order.

PUB-02 chunks the approved content by section, generates embeddings, and registers the source metadata for each chunk: page code, section identifier, publication date, and the category (framework, research, build log, system, career). The category metadata allows the AI Second Brain to apply category-aware retrieval weighting — a query for operational procedures retrieves from systems and build logs before frameworks, and a query for research questions retrieves from research pages before career pages.

CodeNameStatusDescription
PUB-01Publication Surface PublisherStableOn approval, routes the finalized content to the publication surface. Generates the route, populates the page metadata, registers the page in the cross-link index, and triggers the dead link checker to verify all outbound links in the published page resolve to existing routes.
PUB-02Retrieval Store IndexerIn reviewOn approval, simultaneously routes the finalized content to the AI Second Brain's retrieval store. Chunks the content by section, generates embeddings, and registers the source page, section, and publication date as metadata. The simultaneous publish ensures the retrieval store is never behind the publication surface — content that is publicly readable is always retrievable by the second brain.
§ 08Operational Notes

The review gate as thermostat

Every attempt to loosen the human review gate has increased downstream rework. Drafts that pass critique but fail human review reveal categories of judgment that the critique criteria do not capture — publication standard fit, editorial tone above the threshold the voice constraints define, factual claims that require domain knowledge the critique model does not have. The gate is not a bottleneck. It is the system's quality control surface. Loosening it saves time in the pipeline and spends it later in corrections.

Voice constraint enforcement gap

DFT-02 flags voice violations but does not correct them. The correction cycle — flag, redraft, re-flag — adds latency. The alternative — automated correction — produces drafts that are technically compliant but tonally flat. The gap is that voice constraints are easier to recognize than to enforce generatively. DFT-02 is in review because the flagging criteria are not yet stable: some constraints produce false positive flags on prose that is stylistically unconventional but editorially correct.

Critique loop termination

The three-revision maximum before human escalation is a judgment call with no principled basis. Two revisions rarely resolve a structural claim problem — those require outline-level changes, not draft-level rewrites. The revision limit should be conditional: criterion-level failure that traces to the outline graph escalates to outline revision, not to a third draft pass. This has not been implemented.

Retrieval store simultaneity

PUB-02 runs simultaneously with PUB-01 but not atomically. A publish that succeeds on the publication surface and fails on the retrieval indexer produces an inconsistent state: the page is publicly readable but not retrievable by the second brain. The failure is silent until a retrieval query fails to surface the page. An atomic publish-and-index operation with rollback on indexer failure is in design.

Outline graph coverage

The outline graph captures explicit section dependencies. It does not capture implicit argument dependencies — a claim made in passing in section two that is required for the argument in section five to hold. These implicit dependencies produce drafts that are locally coherent but globally inconsistent. The detection mechanism for implicit argument dependencies has not been specified.