06-reference

every claude fable uncertainty triage

2026-07-07·reference·source: Every·by Katie Parrott

Why this is in the vault

Every's Context Window issue argues that Fable's real edge is not scale but uncertainty resolution — finding the flawed premise or unnamed criterion before a cheaper model executes against it for an hour. The Thariq Shihipar framework ("unknown knowns" vs "unknown unknowns") and the Becky Isjwara workflow (expensive model writes the instruction manual, cheaper model runs production) are directly applicable to RDCO's harness architecture and phData client model-selection conversations.

⚠️ Sponsorship

Render (render.com) has an explicit sponsor block in this issue. Offer: $50 in Render credits to Every readers via a tracked signup link. Render is positioned as the cloud infrastructure Every itself uses for AI products. Relationship: paid partnership ("Want to sponsor Every?" CTA also present). No editorial conflict with the article content.

Issue contents

SIGNAL — Use Fable to find your unknowns (main essay)

Anthropic MTS Thariq Shihipar's field guide distinguishes two types of pre-execution gaps:

Two Every examples: Mike Taylor gave Fable a finished DSP book manuscript and asked what he'd missed — the difficulty was evaluation, not scope. Dan Shipper asked Fable to review five weeks of copy-editing experiments; Fable found they had been optimizing against an unvalidated 70%-accuracy target for the wrong baseline metric entirely. Neither task looked heavyweight by scope. Both were Fable-sized by uncertainty.

Triage heuristic: use cheaper models when goal, constraints, and definition of good are settled. Reach for Fable when the map is still incomplete.

STEAL THIS WORKFLOW — Expensive model writes the instruction manual

Every head of social Becky Isjwara's repeatable pattern:

  1. Give Fable the job AND the previous failed attempts with context
  2. Have Fable document its method — scripts for repeatable steps, skills with examples for judgment calls
  3. Run production on Opus with the saved instructions; return to Fable only when the format changes or instructions break on a new case

SKILL SHARE — /ce-pov: force AI advice to survive project contact

Product leader Trevin Chow's open-source compound-engineering plugin skill evaluates any library, framework, or idea against the actual project's dependencies and prior decisions before issuing a verdict. Output is one of five grades: Adopt / Trial / Hold / Reject / Not-our-problem, each with a recommended next action.

DATA POINT — Fine-tuned specialist beats frontier at 13.8x lower cost

Bridgewater AIA Labs fine-tuned Qwen3-235B (via Thinking Machines Lab's Tinker platform) to outperform every frontier model tested across six financial tasks at 13.8x lower inference cost per task. Reinforces the triage thesis: general frontier models for uncertainty; domain-trained specialists for settled, repeated work.

DISCUSS

Sam Altman (Wall Street Journal) quoted walking back AI unemployment fears in favor of "same headcount, more output" framing. Every CEO Dan Shipper's counter: AI makes code and writing cheap to produce, so companies produce more of both — but judgment about what's worth keeping still requires humans. Open question is who gets the new judgment work.

Mapping against Ray Data Co

The "expensive model writes the instruction manual" pattern (Becky's workflow) is exactly how RDCO's harness is designed to scale: Fable or Opus designs the orchestration logic and documents it as a SKILL.md; Sonnet runs production against that skill. This article gives a named, repeatable pattern for that architecture — and a concrete trigger condition for when to escalate (uncertainty in the assignment, not just task size).

The Thariq Shihipar "unknowns" framework belongs directly in RDCO's pre-dispatch checklist: before launching a subagent, classify whether the goal/constraints are settled or whether the agent will hit unmapped territory. If unmapped: use a higher-capability model to surface the unknowns first, then hand structured instructions to the cheaper runner. This is the missing front-end gate on the implementation-notes SOP.

The /ce-pov skill concept maps cleanly onto RDCO's own skill-adoption workflow — any new tool evaluation should cite a concrete project dependency or decision before issuing a verdict. Worth reviewing against the existing ~/.claude/skills/ review pattern.

Timing note: this issue dropped on the last day Fable 5 was included in Claude Pro/Max weekly limits. Pay-as-you-go pricing kicks in July 8 — Fable calls from the harness now carry explicit per-use cost. The triage heuristic (save Fable for genuine unknowns) has immediate budget implications for RDCO's always-on agent loop.

Related