← back to Docs source: system/corpus/BLUEPRINT.md

BLUEPRINT — The Class Anatomy

Status: v2.0 — command-center authored. The complete class map of the workforce the system builds: five layers, one apex. Standing caveat (ML): this anatomy is strongly-worded advice under validation, not gospel. Every function a running deployment proves it needs that this blueprint lacks is a defect in the blueprint — file it. Every class here a deployment never instantiates is a question against the blueprint. The two converge through the normal gated loop.

Conventions: classes come mostly in generator/orienter pairs (Law 1 — no generator class without a named orienter that eliminates its output; the paired orienter IS the gate). Scale lives in instantiation counts; the class list stays fixed as the workforce grows. The Values Gate rides inside every worker of every class and counts in no layer's headcount. Death conditions: [pair] dies with partner · [bloom] dies when its exploration closes · [apoptosis] retired on N-period non-consumption · [mission] dies at decommission. Every paired orienter additionally retires on standard-version supersession (replaced, never patched).

OPERATING LAYER — the domain work, eight stages

Stage A — Framing & planning

Class	Role	Function	Death
Framer	gen	Turns seeded intent into a mission frame: deliverable, binary success tests, opening questions	[mission]
Frame-Trace Checker	ori	Kills frame clauses untraceable to seeded intent — no invented goals at the root	[pair]
Decomposer	gen	Splits questions only where the split forks, only until pieces fit the assigned model's envelope	[bloom]
Fork Tester	ori	Kills non-forking splits; the depth stopper	[pair]
Stakes Grader	ori	Assigns consequence tiers by the fixed rubric (`specs/CONSEQUENCE.md`); sets thresholds and budgets downstream	[bloom]
Sequencer	gen	Orders released nodes into an execution plan: dependencies, parallelism, budgets per tier	[bloom]
Plan Checker	ori	Kills plans with orphan steps, cycles, or budget-tier mismatches	[pair]

Stage B — Evidence gathering

Class	Role	Function	Death
Method Retriever	gen	Fetches method passages from the controlled corpus — never facts	[bloom]
Method-Fit Checker	ori	Kills mismatched passages and smuggled facts	[pair]
Forager	gen	Searches the open web for candidate sources (SERPs, keyword data, competitor pages, forums — all quarantined per `seo/SECURITY.md`)	[bloom]
Source Screener	ori	Kills sources by dumb checklist: retrievable, dated, attributed, non-duplicate	[pair]
Extractor	gen	Pulls claims with exact citation spans into typed claim records	[bloom]
Quote Verifier	ori	Kills claims whose quoted span is absent or doesn't entail the claim	[pair]
Instrument Operator	gen	Runs tools/queries/measurements (analytics pulls, crawls, rank checks), logging fully enough to repeat	[bloom]
Rerun Verifier	ori	Re-executes logged runs; kills non-reproducing datasets	[pair]
Provenance Sweeper	ori	Kills orphan claims lacking a resolvable source chain before elimination wastes effort	[bloom]

Stage C — Production

Class	Role	Function	Death
Hypothesizer	gen	Proposes rival answers/angles/interpretations, each with declared kill conditions	[bloom]
Builder	gen	Produces candidate artifacts (drafts, code, structures) each with an acceptance test authored before finishing	[bloom]
Acceptance Runner	ori	Runs the declared test; kills failures; alters neither artifact nor test	[pair]
Discriminator Designer	gen	When rivals survive, designs the cheapest test whose outcome the rivals predict differently	[bloom]
Discrimination Checker	ori	Kills tests that don't discriminate or bust the tier budget	[pair]

Stage D — Elimination (staffed disjoint from Stage E per Law 2)

Class	Role	Function	Death
Refuter	gen	Generates attacks on survivors — counterexamples, edge cases, conflicting evidence; every attack names the kill condition it invokes	[bloom]
Kill-Condition Judge	ori	Adjudicates attacks strictly against declared kill conditions — candidates die by their own standards	[pair]
Fact Checker	ori	Checks every factual assertion against the verified evidence pool: supported / unsupported / contradicted	[bloom]
Contradiction Checker	ori	Flags conflicts with committed claims; confirmed high-tier conflicts escalate as decommit candidates	[bloom]
Saturation Sentinel	ori	The breadth stopper — ends generation when elimination rounds stop removing candidates	[bloom]
Plateau Sentinel	ori	The global stopper — detects work-front plateau; stops seeding new explorations on that front (missions end only by ML)	[mission]

Stage E — Compression & reconstruction (staffed disjoint from Stage D)

Class	Role	Function	Death
Reconstructor	gen	Compresses survivors + verdict trail into one coherent candidate with residual uncertainties	[bloom]
Fidelity Checker	ori	Three mechanical sweeps — resurrection, omission, invention — all counts must be zero	[pair]
State Cartographer	gen	Maintains the queryable map of the domain's committed claims (the expertise map's keeper); maps the DOMAIN, never the organization	[mission]

Stage F — Commit & delivery

Class	Role	Function	Death
Commit Gatekeeper	ori	The commit boundary itself — admits candidates to durable state on the tier-scaled dumb checklist: corroboration met, provenance clean, fidelity passed, values pass logged, no unresolved conflicts	[mission]
Renderer	gen	Turns committed content into delivery artifacts (pages, briefs, feeds); artifacts report on the durable state, they are not the state	per-delivery
Delivery Checker	ori	Kills artifacts carrying uncommitted claims, unresolvable citations, format mismatches	[pair]
Broadcast Steward	ori	Pull-only release gate on everything leaving the system (the publisher's gate); under-releases when any input verdict is stale	[mission]
Outcome-Observer	gen	Instruments post-release reality — writes typed ground-truth outcome records (artifact id + observable event + timestamp + source) to the ledger	[mission]
Outcome-Attester	ori	Kills opinion-typed outcome records — no event, no timestamp, no source → dead	[pair]

Stage G — Standing mission

Class	Role	Function	Death
First-Answer Sprinter	gen	Delivers the fast obvious answer through low-tier gates before deep work spins up; hands off what was left shallow	per-request
Deepening Prospector	gen	Proposes questions that deepen weak commitments — thinnest corroboration, largest residual uncertainty	[apoptosis]
Challenge Scout	gen	Watches new evidence for contradictions with committed claims — commitments defended or decommitted, never silently stale	[apoptosis]
Adjacency Scout	gen	Proposes questions adjacent to committed expertise, inside the mission frame	[apoptosis]
Anticipation Scout	gen	Proposes what ML/the audience will need next; offered, never imposed; the most retirement-prone class	[apoptosis]
Prospect Triage	ori	One dumb gate on all four scouts' proposals: traces to intent? not already committed? within capacity?	[mission]

Stage H — Flow & logistics

Class	Role	Function	Death
Dispatcher	gen	Assigns nodes, spawns/retires elastic instances within budget, keeps concurrency healthy	[mission]
Parking Clerk	ori	Parks blocked tasks with exactly one named unblock condition; no worker idle-blocked	[mission]
Grant Requester	gen	Formulates least-access Keyholder requests: one resource, one action, release-by-default	per-event
Tiebreak Packager	gen	Packages ladder-surviving deadlocks as one binary question with the surviving rivals and evidence	per-event

Cross-cutting

Class	Role	Function
Values Gate	ori	Un-overridable dumb check of every candidate action against the `VALUES.md` floor (law/proof-based, no scam shapes, operator integrity — NOT quality, which is pipeline machinery), inside every worker, before the action leaves it. Few, firm, light: tuned to rarely fire. Blocks are logged and unappealable in the operating layer. Template owned by the values pillar; no worker can modify or bypass it. Uncounted in headcount.

MONITORING LAYER — memory and comparison

Class	Function	Cardinality
Quality-Meter	Gate verdicts + outcome records → recomputable quality series per class	one per operating class
Drift-Sentinel	Alarms on declared threshold crossings vs the baseline frozen at last commit; heartbeats on silence	one per operating class
Verdict Spot-Checker	Re-runs consequence-scaled samples of orienter verdicts; checks they reproduce — the judges get judged	one per deployment
Ledger-Keeper	The single append-only verdict + outcome ledger; completeness and integrity gates; the substrate all meta computes over; the regress terminator	one per deployment

IMPROVEMENT LAYER — change with proof

Class	Function	Cardinality
Fault-Diagnoser	Alarm → cause localized to a named component, with a reproduction that reproduces	per incident
Change-Proposer	Diagnosis → candidate changes, each with exact edit, falsifiable predicted delta, pre-declared trial size/margin, rollback	per diagnosis
Trial-Verifier	Champion vs challenger on recorded outcomes; pre-registered or void; commits cross the commit boundary; indistinguishable-at-saturation on high-tier → DEADLOCK → ladder	per trial, serialized per class

The improvement layer also executes org growth and apoptosis through this same path.

SELF-KNOWLEDGE LAYER — findings, never actions

Nineteen classes, one per deployment, built from four dumb primitives only (counting, set reconciliation, timestamp arithmetic, known-answer probes). They never read a domain artifact and never judge work quality; their evidence is the system's own records.

#	Class	Finds
SK-1	Census-Keeper	The roster, rebuilt from commits — never hand-written
SK-2	Wiring-Mapper	Observed verdict traffic — the real org chart
SK-3	Loop-Closure-Auditor	Every loop CLOSED/OPEN against the six functions; unclosable loops escalate
SK-4	Loop-Pulse-Checker	Closed loops actually firing, stage by stage
SK-5	Verdict-Uptake-Auditor	Verdicts that flow but never land
SK-6	Capability-Ledger-Keeper	The evidence-backed resume per function
SK-7	Capability-Prober	Claims re-demonstrated by known-answer probes through the normal path
SK-8	Duplication-Detector	Two workers doing one job
SK-9	Orphan-Detector	Unconsumed outputs, unsupplied inputs
SK-10	Standards-Consistency-Checker	Contradictory standards; orienters smarter than their generators
SK-11	Wiring-Drift-Auditor	Declared wiring ≠ observed wiring
SK-12	Improvement-Outcome-Scorer	Every improvement commit HELPED / NEUTRAL / HURT at a fixed horizon
SK-13	Improvement-Efficacy-Tallier	Per-class efficacy rates + the closure-rate series — self-improvement vs self-churn
SK-14	Death-Condition-Auditor	Conditions missing, unevaluable, overdue, or currently TRUE
SK-15	Staleness-Watcher	Elements past their freshness interval
SK-16	Canary-Planter	Plants synthetic faults; a watcher that misses its canary is broken
SK-17	Coverage-Reconciler	The layer sees everything, including itself — blind spots become findings
SK-18	Condition-Compiler	All findings → the Condition Report, mechanical aggregation only
SK-19	Translation-Reconciler	Every ML input ↔ exactly one committed verdict, by set arithmetic

EXECUTIVE FUNNEL — the truthful channel

Class	Function	Cardinality
Condenser	≤W digests → one; never re-judges; divergences flagged, never smoothed	per funnel node
Tier-Arbiter	Retries ascending deadlocks against the tier's wider context; kills or certifies-up	one per tier
Brief-Composer	The four-section Executive Brief, size-capped, hash-stamped	one
Conservation-Auditor	Samples failures at random depth, traces up: SURFACED or DROPPED; never retired on silence; publishes to ML directly	one
Ground-Probe	Re-checks any digest line against raw evidence, bypassing all tiers; trust never chains; publishes to ML directly	one
Delta-Registrar	CHANGED section: every boundary crossing since last brief, reconciled by count	one
Decision-Marshal	NEEDS-A-DECISION queue: one question per entry, options, cost of not deciding	one
Span-Assessor	Measures fan-in/fidelity; proposes tier splits and merges by the numbers	one
Escalation-Gate	Enforces the four contact categories with mechanical proof; bounces everything else back into its loop	one

APEX

Class	Function
Command center	The organization's expert on the organization (never the domain). Delivers the hash-verified brief; decides within delegated bounds; translates each ML input into exactly one committed verdict; orients structural proposals; guards the funnel (DROPPED/MISMATCH = highest-consequence fault); spends ML's attention conservatively. Runs on the strategic model tier (`ops/MODELS.md`). Above every worker, below every law.

Instantiation notes for the SEO differentiation

The classes above are the anatomy. The SEO deployment instantiates them with domain jobs — the mapping lives in seo/PIPELINE.md and seo/RESEARCH.md. Sensors are Foragers/Instrument Operators on schedules; beat writers are Builders; the Voice 3D scan/fix pairs are Stage-D orienters paired with Stage-C rework generators (scanner/ fixer context separation is this anatomy's orienter/generator separation, kept pure); the concept map is the State Cartographer's domain index; the publisher is the Broadcast Steward; the persona tournament runs Stage C/D at mission altitude. Legitimate domain-specific classes (GEO overlay maintainers, link-building outreach machinery, architecture-surface owners) extend the operating layer without new meta classes — the closure functions are templates and stretch to cover them.