Blueprint · careervector.blueprint-graph/v5

Graph explorer

Twenty composable top-level diagrams plus recursively nested 21×21 semantic layers. Nodes descend; boundary arrows cross to neighbours.

Nested layer 1 · runtime · control · feedback · evidence

Adaptive JobCache work routing

The orchestrator combines declared policy, observed feasibility, curiosity, and causal outcomes to distribute pull-based work across unreliable actor classes.

  • Declared policy to Addressed tensor field: shape
  • Due work to Rank + curiosity: prioritize
  • Addressed tensor field to Rank + curiosity: score options
  • Live eligibility to Rank + curiosity: current capability
  • Rank + curiosity to Lease + receipt: select
  • Lease + receipt to Normalized outcome: attempt
  • Normalized outcome to Pathwise learning: project
  • Pathwise learning to Addressed tensor field: update path
shapeprioritizescore optionscurrent capabilityselectattemptprojectupdate path
policyas-built

Declared policy

Budgets, source rules, quality floors, urgency, and curiosity enter at broad addresses.

JobCache operator
processas-built

Due work

Freshness, source turnover, user-triggered misses, and retries define current demand.

JobCache orchestrator
processas-built

Addressed tensor field

Each address prefix contributes preference and learned evidence for an actor-class option.

Sporewright
boundaryas-built

Live eligibility

Actors pull only work they can currently attempt; reachability remains empirical and changeable.

Participant and orchestrator
processas-built

Rank + curiosity

Expected utility, uncertainty, load, and exploration select an actor class without pinning one device.

Sporewright
processas-built

Lease + receipt

A bounded task carries the exact causal address and revision used to choose it.

JobCache orchestrator
storeas-built

Normalized outcome

Reachability, parse quality, latency, failure class, and cost become bounded signals.

JobCache ingestion
processas-built

Pathwise learning

Feedback changes the leaf and bounded ancestors; aging keeps the network responsive to drift.

Sporewright
  1. 01 · policy · as-built Declared policy

    Budgets, source rules, quality floors, urgency, and curiosity enter at broad addresses.

    JobCache operator
  2. 02 · process · as-built Due work

    Freshness, source turnover, user-triggered misses, and retries define current demand.

    JobCache orchestrator
  3. 03 · process · as-built Addressed tensor field

    Each address prefix contributes preference and learned evidence for an actor-class option.

    Sporewright
  4. 04 · boundary · as-built Live eligibility

    Actors pull only work they can currently attempt; reachability remains empirical and changeable.

    Participant and orchestrator
  5. 05 · process · as-built Rank + curiosity

    Expected utility, uncertainty, load, and exploration select an actor class without pinning one device.

    Sporewright
  6. 06 · process · as-built Lease + receipt

    A bounded task carries the exact causal address and revision used to choose it.

    JobCache orchestrator
  7. 07 · store · as-built Normalized outcome

    Reachability, parse quality, latency, failure class, and cost become bounded signals.

    JobCache ingestion
  8. 08 · process · as-built Pathwise learning

    Feedback changes the leaf and bounded ancestors; aging keeps the network responsive to drift.

    Sporewright

Flows

  • Declared policy Addressed tensor field shape
  • Due work Rank + curiosity prioritize
  • Addressed tensor field Rank + curiosity score options
  • Live eligibility Rank + curiosity current capability
  • Rank + curiosity Lease + receipt select
  • Lease + receipt Normalized outcome attempt
  • Normalized outcome Pathwise learning project
  • Pathwise learning Addressed tensor field update path

Canonical architecture note

Open note ↗

Ingestion

Outcome

Continuously acquire public job observations, preserve what was seen, extract and enrich it once per stable input, and publish a coherent result into JobCache Core without overwhelming the production database.

Pipeline

discover → lease → fetch → preserve raw → fingerprint → extract → enrich → validate
         → reconcile ad/role → publish batch → verify → age and refresh

Authorities

Concern Authority
Raw observations Corbet S3 with declared overflow/fallback storage
Unfinished work Small rebuildable pipeline state, not CrateDB truth
Routing and feedback JobCache's Sporewright application
Embeddings and selected analysis Cloudflare AI behind content-addressed caches
Published public facts CrateDB in JobCache Core

Invariants

  • Preserve raw evidence before expensive or lossy transformation.
  • Never perform the same expensive transformation twice for the same stable input and recipe.
  • Actors pull bounded leases; the orchestrator does not assume reachability or capability.
  • Browser, desktop, and container actors return the same receipt and feedback contract.
  • Failed enrichment cannot partially mutate the canonical role/ad projection.
  • Interactive workloads outrank asynchronous replenishment.

Current state

The adaptive extraction orchestrator, browser/Tauri/container participants, routing receipts, S3 preservation, CrateDB deposit path, evidence aging, and source adapters exist in varying degrees of maturity. The complete cold processing ledger and atomic publish pipeline remain target evolution.

Completion evidence

Load, drift, poison, retry, duplicate, outage, and heterogeneous-actor simulations plus a production trace from source discovery to a published role revision.

Source: architecture/modules/ingestion.md