# 49 — Volume II: coverage-delta ledger, verdict delta, and corrections (https://jackin.tailrocks.com/research/token-optimization/49-extension-stacks-and-verdict/) # 49 — Volume II: coverage-delta ledger, verdict delta, and corrections [#49--volume-ii-coverage-delta-ledger-verdict-delta-and-corrections] The Volume II capstone: it collects the coverage-delta notes that prove novelty against files 00–32, states whether Volume II moves Volume I's headline verdict (with arithmetic), records cross-layer corrections and caveats, updates the composed stacks for the subscriber metric, and lists the Volume II graveyard of claims killed or downgraded this run. **TL;DR** * **The 10x dollar verdict is unchanged: ≈2.5× defensible / ≈5–6.2× with validated routing / no true 10× at equal quality.** No Volume II lever removes Volume I's two binding constraints (frontier-model thinking output; the cache-read floor of genuinely-used context). Multimodal adds a real but small dollar lever (vision-tier routing cuts 67% of *image* tokens, a minor share of most coding sessions); latency, governance, and portability change *which* choice is correct and *what is measurable*, not the dollar ceiling. * **But for this operator the metric is wrong.** The local credential is Max (file 41); a subscriber's binding constraint is the **usage cap**, not dollars, so Volume II presents a **second cost model alongside** Volume I's: optimize **tasks-per-cap**, where the lever order re-sorts (prefix stability, context-window size, and request-volume discipline rise; subagent fan-out partially inverts; style compression matters even less). Below the cap, dollars are sunk; the dollar model applies only to the overage tail (API-rate credits) and to the off-cap headless/SDK lane. * **A pricing reconciliation, not a multiplier change:** Volume I's dollar figures are Fable-5-priced ($10/$50); the operator's actual main model is **Opus 4.8** ($5/$25, identical tokenizer; measured 465/560 calls), and Fable 5 leaves the subscription. So Volume I's absolute daily dollars are **\~2× high** for this operator — but ratios are price-invariant, so the multipliers and tier list are unaffected. * **42 genuinely-new techniques** (files 41–47), each with a coverage-delta note proving absence from 00–32, plus 8 new frontier ideas (48) — all with the full Volume I §10 record schema and validation protocols. The coverage-delta ledger is below. * **Cross-layer corrections/caveats:** the server-cache-scope conflation was applied to `13`; the subagent-caching-default conflict (#29966 vs Volume I's measured subagent cache writes) remains version/path-dependent and must be audited in the operator's own JSONL before acting. *** ## Coverage-delta ledger (novelty proof vs 00–32) [#coverage-delta-ledger-novelty-proof-vs-0032] Every Volume II technique was checked against the named Volume I file before claiming novelty. | Technique | Vol I file checked | Why genuinely new | | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | | 41 Q1–Q7 (quota model, prefix-as-quota, window size, request-volume, 1h-in-allowance, overage decision, SDK-off-cap) | 13 (caching), 01 (econ), README | Vol I prices in $ only; "quota" named in 13, modeled 0× — the cap weighting, denominator, re-sort, and the 06-15 split are all new | | 42 M1–M6 (vision routing, downsample, text-over-screenshot, text-over-PDF, crop, lazy vision) | 12, 16, 03 | Multimodal = 0 substantive hits in 00–32 ; image/PDF token costs unmeasured | | 43 L1–L6 (fast-mode purchase, optimizer latency tax, parallel-time, cache-on-both-axes, batch-time, quota-edge decision) | 17:110, 18:132, 19:147, 01:52 | Latency mentioned in scatter; never built into a time-value decision model (v·t·s > Δ$) | | 44 F1–F6 (workspace fleet sharing, excludeDynamicSections sized, cold-start pre-warm, subagent-caching audit, multi-tenant boundary, jackin' fleet policy) | 13 tech 7, 19:116-187, 17 | Vol I covers self-host fleet (19) + the SDK flag (13); hosted cross-container sharing, the measured dynamic-section size, and #29966 are new | | 45 P1–P6 (portable core, keepalive borrow, 3-tier routing borrow, dedup borrow, per-tool budgets borrow, non-portable edge) | 14/15 availability labels, 18:196 | No portability matrix exists in 00–32; reverse-portability (borrow from other agents) is new | | 46 FL1–FL5 (CAG-via-caching, /cd, no-compressor-security, KV-family verdict, Fable→Opus re-decision) | 03, 19, 20 K1/K2/K15 | KV-eviction family (SnapKV/H2O/PyramidKV/KVQuant), CAG, CompressionAttack, /cd, the Fable promo = 0 hits / post-freeze | | 47 G1–G6 (real governors, max\_tokens replacement, online canary, guard tax, degrade-don't-die, recursion cap) | 15:123-135, 31, 32:13 | Vol I has offline harness + max\_tokens-as-rail only; runtime budget governance + online drift detection are new | | 48 V1–V8 (frontier) | 20 K1–K16 | Each maps to a Volume II blind spot; dedup notes vs every K-idea in 48 | **Count: 42 new techniques in 41–47** (Q7+M6+L6+F6+P6+FL5+G6) **+ 8 frontier (48) = 50**, against the brief's floors of ≥25 cataloged and ≥10 with the full record. **All 42 carry the full §10 record** (Name, Layer, Mechanism, Expected savings, Evidence tier, Quality risk, Availability, Effort, Composability, Validation protocol) plus a coverage-delta line. ## Verdict delta — does Volume II move the numbers? [#verdict-delta--does-volume-ii-move-the-numbers] **On dollars: no.** Carrying the arithmetic: * Volume I's binding constraints are structural: (1) frontier-model thinking bills as output and only effort/not-being-the-frontier-model touches it; (2) a cache-read floor of context the agent genuinely needs. **No Volume II lever removes either.** The KV-compression family that could attack the cache floor is self-host-only on hosted Claude (file 46) — same wall as Volume I's K1/K2. * The new dollar levers are real but small: vision-tier routing saves 67% of *image* tokens (file 42), but images are a minor share of a typical coding session; text-over-PDF saves \~50% on *document* tokens, only when PDFs are in play; CAG-via-caching (FL1) converts retrieval into 0.1× reads, a context-architecture win already in Volume I's family. None compounds into a new multiplier. * **So the headline stands after the independent correction: ≈2.5× defensible today, ≈5–6.2× if the Sonnet-main+advisor routing flip passes validation, no true 10× at provably equal quality.** Multimodal/latency/governance/portability change correctness and measurability, not the ceiling. **On the metric: yes — and this is Volume II's real contribution.** For a subscriber (this operator is Max), "$ per task" is the wrong denominator. The quota model (file 41): * Below the cap, **dollars are sunk**; the objective is **tasks per 5-hour/weekly window**. * The cap weights re-sort the levers: **prefix stability** and **context-window size** and **request-volume discipline** become the top levers (a cache *miss* is a 1.25–2× write against the cap; subagent fan-out is \~61% of calls); **style/register compression matters even less** than in the dollar model (visible output is a sliver of cap-weighted tokens). * The cap weighting itself is **officially unpublished**; community triangulation puts cache-read at \~0.1× (T3, file 41), and the **token denominator is opaque** (bounded INCOMPLETE). So "tasks per cap" is *directional* without the per-account prober (frontier V2). **A pricing reconciliation (not a multiplier change):** Volume I modeled the day at Fable 5 prices ($17–22/day). The operator runs **Opus 4.8** (half the per-token price, identical tokenizer family), and Fable 5 leaves the subscripti. So Volume I's *absolute* dollar figures are \~2× high for this operator going forward; **ratios, multipliers, and the tier list are unchanged** (they are price-invariant) and all of Volume I's tokenizer-arbitrage and the file-42 image math transfer directly (Opus 4.8 ≡ Fable 5 tokenizer). ## Tier-list delta [#tier-list-delta] Volume I's S/A/B/C/F tiers stand. Volume II adds, scored by the same `$saving × confidence ÷ effort` on the dollar axis, and re-weighted for the quota axis where noted: * **New S-tier (negative-cost or large, low-effort):** vision-tier routing for screenshot/PDF work (42 M1, −67% image tokens, trivial); text-over-PDF / text-over-screenshot (42 M3/M4, negative-cost); `/cd` cache preservation (46 FL2, zero-effort negative-cost); **prefix-stability-as-quota-lever** (41 Q2 — promoted to top on the quota axis); the real governors (47 G1, loss-avoidance). * **New A-tier:** CAG warm-repo cached prefix (46 FL1 / 48 V6); time-value fast-mode (43 L1/V5); fleet workspace cache sharing (44 F1–F3); online-canary-gated compression (48 V7); quota-window scheduling (48 V1) for capped subscribers. * **New B/C-tier:** portable-policy compiler (48 V8); the borrow-ins (45 P2–P5); cross-provider effort/cache levers (45) where the operator might switch agents. * **New F-tier (killed/blocked — see graveyard):** the hosted-Claude KV-compression family (46, $0); prompt-compressors in the hot path (now also a security kill, 46 FL3 / 47); `max_tokens` as a spend governor (re-killed, 47 G2); lossless at-rest compression as a "token" saver (46, category error). ## Composed-stack updates [#composed-stack-updates] **Conservative (do tomorrow, riskless) — adds:** vision-tier routing + text-over-pixels/PDF for any visual work; `/cd` instead of restart-in-new-dir; treat prefix stability as the #1 *quota* lever; set a Claude Code `/usage-credits` monthly cap and (for fleets) a workspace rate limit; re-decide Fable→ Opus 4.8 before 06-23. **Aggressive (with validation) — adds:** the CAG warm-repo cached prefix (preload the stable repo core once); fleet workspace cache sharing + cold-start pre-warm (jackin'); time-value-driven fast mode for interactive sessions; an online-quality canary so aggressive register/effort compression runs behind a live safety net. **Unbelievable (chase the ceiling) — adds:** the quota-window scheduler + per-account quota-denominator prober (close the cap INCOMPLETE); the cross-provider portable-policy compiler; the jackin'-baked fleet cache + governance pack. **Binding constraint, restated:** on dollars, still the frontier-thinking + cache-read floor (no 10×). On quota, the binding constraint is the **unpublished denominator** — you cannot prove "10× more tasks per cap" without measuring the cap, which only a header-reading proxy can do per-account. **For a subscriber specifically**, the stack re-orders: cap-headroom levers (prefix stability, smaller window, request-volume discipline, 1h-in-allowance, headless-off-cap placement) come first; dollar levers matter only on the overage tail. ## Cross-layer caveats [#cross-layer-caveats] Two cache-layer facts that span Volume I and the fleet analysis: 1. **Server cache scope vs local file cache (13 tech 7 + surprising-findings) — applied.** Original Volume I text stated "your git state is in the cache key… worktrees never share," citing the prompt-caching docs, implying the *server* prompt cache is keyed by machine/directory/git-snapshot. File 44's sources show the **server** prompt cache is **workspace-scoped** with **no** machine/dir/worktree key; the git-snapshot/worktree rules describe Claude Code's **local file cache** (a different layer, GitHub #17531). Practical effect is favorable: hosted fleets *can* share a prefix across machines/dirs. 2. **Subagent caching default (13 tech 2 vs GitHub #29966).** Volume I measured subagents *writing* 5m cache (1,128/1,128 calls). #29966 (T3, one session, Claude Code 2.1.63 / SDK 0.2.63) reports Agent-tool subagents with `enablePromptCaching` hardcoded **false** (caching off). These may both be true at different versions/spawn paths (cavecrew/Task-tool subagents vs SDK Agent-tool subagents). **Recorded as version/path-dependent; operator should audit their own subagent JSONL** (file 44 F4) rather than assume either. No other Volume I claim was contradicted; the min-cacheable-prefix values (13 tech 11), tokenizer-family equality, and cache multipliers were all re-verified live and **stand unchanged**. ## Volume II graveyard (killed or downgraded this run) [#volume-ii-graveyard-killed-or-downgraded-this-run] | Claim | Verdict | Where | | -------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------- | ----- | | "KV-cache compression (SnapKV/H2O/PyramidKV/KVQuant) will cut your hosted-Claude bill" | **KILLED** — all self-host-only; $0 on a hosted API; extends Vol I's K1/K2 block | 46 | | "cache\_read counts at 1× against the subscription cap" / "the cache double-counts" | **KILLED** — \~0.1× (T3); the 1× double-count was a LiteLLM/Bedrock proxy bug, fixed; direct Anthropic auth does not double-count | 41 | | "Anthropic's workspace dollar spend limit hard-stops spend" | **KILLED** — alert-only; hard block is via rate limits / `/usage-credits` / gateways | 47 | | "Set `max_tokens` low to cap spend" | **RE-KILLED** — truncates `tool_use`, bills the attempt + a higher-cap retry; use `model_context_window_exceeded` + a budget | 47 | | "The multi-agent 'up to 90% faster' result applies to coding agents" | **KILLED** — research-only; Anthropic says coding is a poor fit; naive parallelism risks pooled-limit backoff | 43 | | "Lossless prompt compression (LoPace) cuts your token bill" | **KILLED (category error)** — compresses bytes at rest, not tokens sent | 46 | | "CompactPrompt saves \~60% for coding agents" | **DOWNGRADED to T4 folklore** — single vendor guide, lossy NL; LLMLingua caveats apply | 46 | | "Keepalive pingers save the subscription quota on the Claude Code main loop" | **MOOT (re-confirmed)** — main loop is 1h-TTL in-allowance; Cherny: "a small win"; overage reverts to 5m | 41 | | "A screenshot is a cheap way to show the agent code/logs" | **KILLED** — a screenshot is 2–6× the text it shows and caps at one frame; text wins for textual content | 42 | | "Prompt-compression proxies are merely a cost trade" | **DOWNGRADED** — also a security attack surface (CompressionAttack, ≤80% ASR) | 46 | ## What survived the adversarial pass [#what-survived-the-adversarial-pass] Every Volume II headline number was either documented-and-locally-reproduced, honestly tiered, or explicitly INCOMPLETE. The two most novel/load-bearing numbers were re-attacked and held: * **Image-cap divergence is \~3.0–3.1×:** confirmed content-independent — a max-entropy noise image at 2560×1440 returned the same capped family split as the gradient. The formula and routing lever stand; exact capped counts vary by wrapper/envelope, so use the band rather than 4,784/1,568 as measured absolutes. * **The PDF tax \~2×:** confirmed across content — code-like text gave 1.90× (Opus) / 2.11× (Sonnet), bracketing the fox-text 1.98× / 2.30×. Holds. The community-sourced and unpublished items (cap weight \~0.1×, the cap denominator) are labeled T3 / INCOMPLETE and not promoted beyond their evidence. *** The Volume II self-audit against the definition of done is appended to [the research index](./) under `## Volume II — Extension`, alongside the index, headline numbers, blind-spot-map summary, and the Volume II Assumptions section.