Drafting Contract

This file is the compact contract for a future drafting agent. It assumes the scaffold is already stable.

Output Target

Create a draft file only when explicitly asked, probably:

ai_systems/ai_views_blog.qmd

Do not overwrite ai_systems/ai_views.md.

Required Inputs

Read before drafting:

  • style_contract.md
  • article_thesis.md
  • claims_ledger.md
  • source_map.md
  • article_spine.md
  • section_claim_ladders.md
  • section_paragraph_plans.md
  • ../concepts/reference_inventory.md

Use concept files only as supporting notes. Do not copy their prose mechanically.

Target Shape

The draft should be a blog post, not an academic paper.

Target:

  • One title.
  • Short opening.
  • 8 to 10 titled sections.
  • One substantial paragraph per body section by default.
  • No abstract.
  • No numbered theorem-like structure.
  • No long literature-review section.
  • References cited where claims need support.

Approximate length:

  • First draft: 3500 to 5500 words.
  • Shorter is better if the argument remains complete.

Primary Argument

The post should make this argument and no larger one:

  1. AI progress is shaped by feedback cost.
  2. Proof search is a clean example because correctness can sometimes be checked cheaply.
  3. Mathematical work also requires question selection, digestion, training, and community knowledge.
  4. Institutions may reward the part of science that AI makes easiest to measure.
  5. Broad questions about plateau, AGI, and replacement should be split by feedback regime.

Section Order

Use the order in article_spine.md unless the author changes it.

If shortening is needed, merge:

  • Cheap Feedback + Feedback Cost Map.
  • Importance Oracle + Conjecture Economy.
  • Synthesis Debt + Community Layer.
  • Plateau + Amplification To Replacement.

Do not drop:

  • Cheap feedback.
  • Proof generation / verification / digestion.
  • Importance oracle.
  • Synthesis debt.
  • Acquisition asymmetry.
  • Measurability distortion.

Citation Rules

  • Cite only verified sources in public prose.
  • Unverified sources can appear only as TODO comments in scaffold, not in the final draft.
  • Use source_map.md to place citations.
  • Do not turn a paragraph into a citation catalog.
  • If the draft uses Erdos, decide whether to keep ASCII Erdos or use the accented form consistently with site style.

Tone Rules

  • No manifesto voice.
  • No staccato chain of short declarative sentences.
  • No mechanical conversion of scaffold bullets into prose.
  • No “crisis” language unless the paragraph gives a concrete mechanism.
  • No generic AI adjectives: transformative, revolutionary, unprecedented.
  • No performative humility: “I may be wrong, but…”
  • No hype-deflation theater: “This is not X, it is Y” unless X is an actual misconception.
  • No long throat-clearing about online discourse.

Final Paragraph

End with the feedback-loop test. Do not end with vague warnings about humanity, civilization, or the future of work.

Candidate ending function:

  • State that AI already does parts of mathematics and science.
  • State that the useful question is which feedback loops it can close.
  • State that the remaining bottlenecks are not automatically less important because they are harder to measure.