---
source_hash: 2604ce12eef7b2252fd8e90014fa83ff9097fa6233d3ef62e2654853aa874ec7
post: /posts/tools-and-plugins/the-model-is-a-dial
---

# Context Pack: The Model Is a Dial

Reference for the July 2026 conceptual framework. Historical model labels and interactive scores are examples, not benchmarks, supported-provider configurations or a current buying recommendation. The page discloses AI-assisted drafting with human thesis/framing/final approval.

## Verbatim excerpts

> The contract does not replace context.

> The controls are conceptually separate; real hosts may constrain them together.

> Extra effort has a cost in <Term id="latency">latency</Term>, tokens, attention and throughput.

## Outcome contract

Precision belongs in the result, audience, constraints, evidence, failure conditions and review gate. A model may investigate and revise the route; it cannot reliably infer unstated people, conditions or authority. Explicit route steps remain useful when the procedure is fixed, deterministic or brittle. Do not turn the prompt ladder into a universal ban on instructions.

The exact template is retained below. Keep outcome_contract and the snake_case failure_conditions/review_gate keys; constraints, evidence and failure_conditions are lists.

```yaml
outcome_contract:
  result: "What must exist when the work is finished"
  audience: "Who or what must be able to use it"
  constraints:
    - "A boundary the route cannot cross"
  evidence:
    - "A test, source, measurement, or inspectable artefact"
  failure_conditions:
    - "A state that blocks completion"
  review_gate: "The independent check or human approval required"
```

## Allocation method

Four variables: path ambiguity; failure cost; time horizon; reviewability/reversal. A small edit may have large consequences. Long tasks need saved state as well as reasoning. Weak reviewability is a reason to improve the checking plan, not assume a stronger model supplies assurance.

Six steps: define acceptance including all six fields; separate destination uncertainty from route uncertainty; price failure including reversal/exposure/data loss/dependencies; inspect useful context/tools/tests/checkpoints; choose the smallest sufficient profile; budget review separately. Reasoning and independent review address different errors and neither guarantees completeness.

Four failure modes: maximum effort by default; model selection before contract; outcome-shaped wishes without evidence; asking one control to compensate for every missing part of the system.

Historical profiles: Haiku/light for bounded reversible short work; Sonnet/standard for moderate synthesis/familiar implementation with evidence; Opus/deep for ambiguous/high-consequence/long-chain decisions; Codex/high for tool-led implementation and verification. Check actual model/version/host capabilities rather than treating these names as equivalent across providers.

## Retained interactive examples

All component paths are under apps/web/components/blocks/:

- PromptLadderFidget.tsx: steps/TASK STEPS; brief/BRIEF; contract/OUTCOME CONTRACT. The first means a prescribed route; the second goal/constraints/acceptance; the third adds evidence/failure/review. Explicit procedures remain valid.
- EffortMatrixFidget.tsx: hand-set score=max(0, high-ambiguity 3 + high-failure-cost 3 + horizonScore - reviewable 1). Horizon is short 0, multi-step 1, long 2. Thresholds ≤0→Haiku/light,≤2→Sonnet/standard,≤4→Codex/high,otherwise Opus/deep. This is a scoring illustration, not a measured model ladder or sufficient assessment of whether work involves coding.
- ModelControlPanelFidget.tsx: selections change the displayed example, not a live run. Six controls and values: model haiku/sonnet/opus/codex; effort light/standard/high/deep; context 32k/128k/1m; tools off/read-only/full; persistence turn/session/durable; review self-check/fresh-audit/human-approval. Defaults haiku/light/32k/read-only/turn/self-check. Settings and combinations depend on the actual host/model; permissions are not granted by this panel.

The hero/Diagram remains /posts/tools-and-plugins/the-model-is-a-dial/01-hero-model-dial.png. Seven Term uses retain six IDs: effort-control twice, reasoning, agentic, context-window, token and latency. Glossary remains. Model effort, sources, tools, saved state and independent acceptance each have a distinct job.