[BunsDev]

What changed

Adds docs/context-scoring-cost-model.md, the phase 4 cost model for per-chunk context scoring, and links it from docs/architecture.md (host seams / ContextScorer) and roadmap phase 4. It ticks only "Compare scoring/re-prefill costs with forfeited prefix-cache reuse". The egress review, shadow experiment and integration decision items stay open. No ContextScorer code is added.

Decision: no-go for ContextScorer integration code; conditional go for a shadow experiment on synthetic context only.

• Break-even drop fraction (with s'=0 and the scored block placed last): d* = s·(1 − r) + ρ·(1 + ε). Here r = 0.1 is the cache-read multiplier (OpenAI and Anthropic docs) and ρ = p_J/P is Jev's price relative to the proposer's: $0.042/Mtok against $1.75/Mtok for gpt-5.3-codex, so ρ = 0.024.
• A context block that stays stable across turns needs d ≥ 0.84 at 10 turns and 0.93 in the long-session limit. That is where scorer recall errors hurt most, so it is a no-go. A block that is reassembled every turn anyway needs d ≈ 0.03, which is cheap in money.
• Jev latency is a serial cost: 206 ms p50, 474 ms p95, 975 ms max over 132 recorded receipts from four live runs (the often-quoted 257 ms is the mean of run 1 alone).
• The only recorded token-usage workload (Codex CLI arena, ~29.6k input tokens, 57.6–91.1% cached) contains one 92-byte scorable file. Nothing there to score.

Every number cites a recorded artifact at a commit SHA (playground 2c6cac9, this repo 3b1704f) or an official doc URL checked 2026-09-22. Unknowns stay explicit parameters: context stability s, drop fraction d, chunk shape, turns per session, prefill time per token τ, the proposer model the arena used, Jev latency at 10–32k-token states, and scorer recall. The doc specifies the shadow experiment that would measure them.

Closes #3

Verdict impact

None. No decision-table row, question, fixture, threshold or model pin changes.

Verification

• pnpm typecheck (exit 0)
• pnpm test (exit 0; 84 pass, 0 fail)
• Commits are signed
• No provider call, credential, execution path, or new verdict added
• If the decision table or a question changed: not applicable

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
jev-harness	Ready	Preview	Sep 22, 2026 1:15pm UTC

[Copilot]

Copilot review overview

🟡 Changes recommended

The model incorrectly treats request count as a fixed overhead ratio and does not fully check Jev’s state-plus-question token limit.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1 · 1

Open (2)

• Validate complete serialized state and question against Jev token limit · New
• Break-even and cost tables rely on unmeasured overhead assumptions · New

What changed in this PR

Adds the phase 4 ContextScorer cost model and records a no-go for integration code, while keeping a synthetic shadow experiment conditional.

Changes:

• Adds pricing, latency, cache, break-even, egress, and experiment guidance.
• Links the decision from architecture documentation.
• Marks the cost-model roadmap item complete; no runtime or verdict changes.

File	Description
docs/​context-scoring-cost-model.md	New ContextScorer cost model and decision record
docs/​architecture.md	Links the cost model and decision
docs/​roadmap.md	Updates phase 4 progress

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.