---
title: "Context efficiency's lenses are analytic; output spend is scored by return · Agentic Atlas"
description: "Context efficiency's lenses are analytic; output spend is scored by return: The three lenses are analytic, not benchmarks."
canonical: "https://agentic-atlas.dev/decisions/0021-context-efficiency-analytic-lenses"
last-updated: "2026-09-14"
---

1. [Agentic Atlas](https://agentic-atlas.dev/)
2. [Decisions](https://agentic-atlas.dev/decisions)
3. Context efficiency's lenses are analytic; output spend is scored by return

Manifest-selected Decision

Canonical id

0021-context-efficiency-analytic-lenses

Status

accepted

Catalog order

10

Catalog revision `a9f60f26fd5ef6a1e0bbb39e8d0de32bf9c464b8c7b8ee3fc8c8c9fdb3a1021a`

# 0021: Context efficiency's lenses are analytic; output spend is scored by return

**Status:** accepted **Date:** 2026-09-10

## Context

`context-efficiency` sat at `seeded` from 2026-07-02 while its child `context-alignment` reached `fleshed` (2026-09-09). Parents-first blocks the child's promotion to `stable` until the parent is `fleshed`, and the parent carried one open fork: whether its three lenses — tokens per task, relevancy, prime-window completion — are measurements the node owes a procedure for, or analytic readings it owes only a definition for. The 2026-09-09 interference synthesis settled the evidence side: no turn-indexed decay curve exists, so prime-window completion has no curve to be read off, and no relevancy rubric has been validated against real sessions.

The node also carried a prescription — *spend output tokens on judgment, not formatting* — whose body restated the economy's formatting-hoop analysis and the mechanisms' remedies, in tension with ADR 0009's ruling that efficiency owns objective-derived guidance only.

## Decision

**The three lenses are analytic, not benchmarks.** Each could be measured; none is well-defined as a metric today, and the node says so rather than manufacturing a rubric. The cost lens counts. Relevancy carries a per-artifact inspection (*would the next turn's output change if this artifact were evicted?*) as its proxy. Prime-window completion carries one measurable proxy — effective-length crossing, the NoLiMa/RULER construct — and no curve. The node states, in *Evidence*, what result would promote each lens from analytic to measured.

**The output-spend prescription is reframed around the model's output economy.** Output splits into delivered tokens and throwaway tokens (reasoning, drafts, restatements, wrapper). Throwaway is scored by what it returns: reasoning tokens are the model buying judgment and are efficient when they move the outcome; re-emission returns nothing at any price. The priced failure mode (the formatting hoop) stays at the economy; its remedies stay with data-presentation separation and template-constrained generation; efficiency keeps the rule and the pointers.

**The node is promoted to `fleshed`** on the concept skeleton, with the stable gate a review pass — its empirical claims are grounded where owned (the economy's anchors; the 2026-09-09 synthesis; Anthropic's thinking documentation, retrieved 2026-09-10, for thinking-block preservation by model generation).

## Consequences

- `context-alignment`'s *Verification* residue points at the inspection proxy, not at a missing rubric.
- The economy's intermediary-tokens bullet gains the preservation-by-model fact as an additive anchor; the price list still depends on no sibling.
- `context-alignment` is un-blocked for `stable` once its four-arm bill lands.
- A validated relevancy rubric or a turn-indexed decay curve reopens the lens ruling; nothing else does.

[Browse all selected Decisions](https://agentic-atlas.dev/decisions)