# 27 — Current verdict and stack guidance (https://jackin.tailrocks.com/research/context/techniques/27-extension-stacks-and-verdict/)



## Summary [#summary]

The current evidence preserves the central dollar verdict: about **2.5× is defensible without a quality trade**, and about **5–6.2× is conditional on a routing change passing the acceptance harness**. No reviewed mechanism supports an honest 10× claim at equal quality. Frontier-model thinking billed as output and the cache-read floor of genuinely useful context remain binding constraints.

For subscription users, the primary metric changes from dollars per task to **accepted tasks per usage cap**. That reorders the levers: prefix stability, context-window discipline, and request volume rise; uncontrolled fan-out can lose despite appearing cheap in API-dollar arithmetic. Provider quota weights remain partly opaque, so quota conclusions stay bounded where direct measurement is unavailable.

## Question and scope [#question-and-scope]

What reduction remains defensible after composing the durable findings, quota constraints, and unresolved frontier ideas?

## Method [#method]

The verdict recomposes only independent, quality-gated effects from the durable chapters, then reports subscription and API-priced outcomes separately with unresolved inputs left explicit.

## Findings [#findings]

### Current evidence by topic [#current-evidence-by-topic]

| Topic              | Current finding                                                                                                           | Detail                                                                                                         |
| ------------------ | ------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- |
| Subscription quota | Use tasks per cap; apply API dollars only to credits or explicitly off-cap traffic                                        | [19 — Subscription and quota economics](/research/context/techniques/19-subscription-and-quota-economics/)     |
| Multimodal         | Vision-tier routing can reduce image tokens by 67%; text-over-PDF can save about 50% on document tokens in measured cases | [20 — Multimodal token economics](/research/context/techniques/20-multimodal-token-economics/)                 |
| Latency            | A token-saving path is unacceptable when added wall-clock or human wait exceeds its value                                 | [21 — Latency and time economics](/research/context/techniques/21-latency-and-time-economics/)                 |
| Fleet              | Prefix identity, dynamic sections, start timing, and provider cache scope determine cross-agent reuse                     | [22 — Fleet and multi-tenant cache](/research/context/techniques/22-fleet-and-multitenant-cache/)              |
| Portability        | Instruction-level techniques travel broadly; hooks, provider betas, and storage surfaces do not                           | [23 — Portability matrix](/research/context/techniques/23-cross-agent-portability/)                            |
| Literature         | KV-cache compression is meaningful for self-hosted inference but does not reduce hosted API-billed tokens                 | [24 — Literature and market scan](/research/context/techniques/24-fresh-literature-and-market-delta/)          |
| Governance         | Measurement overhead, retries, raw recovery, and live quality checks belong in net accounting                             | [25 — Governance and online quality](/research/context/techniques/25-meta-cost-governance-and-online-quality/) |
| Frontier           | Conditional ideas remain gated by feasibility, provider support, and paired quality tests                                 | [26 — Frontier mechanisms](/research/context/techniques/26-extension-frontier/)                                |

### Dollar verdict [#dollar-verdict]

The added mechanisms change choices at the margins but do not remove the two structural floors:

1. Thinking from a frontier model is billed as output and is reachable mainly through effort control or model routing.
2. Context the model genuinely needs still incurs a cache-read floor even under stable-prefix caching.

Multimodal substitution matters when media is a material share of the workload. Cached-context retrieval can reduce fresh-input cost, but it remains a context-architecture and caching lever rather than a new independent multiplier. Savings from overlapping layers must not be multiplied as if each acted on the full baseline.

### Quota verdict [#quota-verdict]

For a capped subscription, use a two-part objective:

```text
accepted tasks per cap = accepted tasks / provider-accounted usage
overage cost           = credits or API traffic × applicable rates
```

The provider does not publish a stable conversion from all token classes to cap consumption. Measure relative changes from the provider surface and session records; do not present an inferred quota weight as a provider fact. Cache stability and fewer requests are robust priorities because both reduce avoidable repeated work even when the exact cap denominator is hidden.

### Conservative stack [#conservative-stack]

* Stabilize system, tool, and project prefixes.
* Keep raw recovery for every transformed observation.
* Use retrieval to avoid irrelevant reads before compressing relevant ones.
* Apply deterministic output discipline.
* Measure fresh input, cache write, cache read, thinking, visible output, turns, rereads, latency, and accepted-task quality separately.

### Quota-bound subscriber stack [#quota-bound-subscriber-stack]

* Add request-volume and context-window budgets.
* Stagger or pre-warm fleet starts only when measurement shows prefix reuse.
* Gate fan-out by expected task value and cap pressure.
* Attribute each capsule's provider-reported usage where the provider exposes it.

### Conditional routing stack [#conditional-routing-stack]

* Route only task classes that pass the paired acceptance harness.
* Preserve an advisor or escalation path for hard cases.
* Recompute savings for the actual current model pair and tokenizer family.
* Reject the route when retries, cache fragmentation, or quality loss erase the nominal price benefit.

### Cross-layer caveats [#cross-layer-caveats]

Server-side provider cache scope is not the same as a local file cache. Local recovery stores make compressed observations reversible; they do not prove shared provider-prefix reuse. Subagent cache behavior is version- and execution-path-dependent, so inspect current session records before relying on it.

Absolute dollar examples tied to one model price can age while ratios and structural constraints remain valid. Apply the [provider-sensitive claim verification contract](/research/context/techniques/28-drift-delta/) before using provider-sensitive numbers.

## Implications for jackin❯ [#implications-for-jackin]

### Current tier adjustments [#current-tier-adjustments]

| Mechanism                                       | Current treatment              | Reason                                                                           |
| ----------------------------------------------- | ------------------------------ | -------------------------------------------------------------------------------- |
| Prefix stability and cache preservation         | Promote for quota-bound work   | Repeated prefix writes and cache misses consume avoidable allowance              |
| Request-volume discipline                       | Promote                        | Extra turns repeat fixed context and create more opportunities for cache misses  |
| Uncontrolled subagent fan-out                   | Downgrade for quota-bound work | Parallel agents can multiply requests and fixed-prefix consumption               |
| Vision-tier routing and cropping                | Conditional                    | Strong local savings only when media is present and still legible                |
| Text extraction before full PDF/image ingestion | Conditional                    | Useful when structure and exact text survive extraction                          |
| Per-request gateway routing                     | Keep harness-gated             | Model savings can be erased by cache fragmentation or quality loss               |
| Hosted KV-cache compression claims              | Reject                         | Self-host GPU-memory savings are not hosted API token savings                    |
| Compression without raw recovery                | Reject                         | Exact identifiers, errors, configuration, and security evidence require recovery |

### Current stack guidance [#current-stack-guidance]

For jackin❯, the current verdict and stack guidance evidence identifies which mechanisms are safe to adopt directly and which still require workload-specific validation.

## Limitations and unknowns [#limitations-and-unknowns]

### Rejected or downgraded claims [#rejected-or-downgraded-claims]

* No evidence supports stacking every reported percentage into a 10× accepted-task claim.
* A payload-reduction benchmark is not a task-level, quota-level, or quality result.
* Self-host KV eviction does not lower billed hosted tokens.
* A flat-rate subscription does not make unlimited retries or fan-out free; the usage cap is the scarce resource.
* A cheaper route is not a saving when quality failures cause escalation, rereads, or repeated turns.

## Sources [#sources]

Evidence and repository references are cited inline beside the claims they support.

## Related work [#related-work]

* [Executive summary](/research/context/techniques/00-executive-summary/)
* [Composed stacks](/research/context/techniques/15-composed-stacks/)
* [Validation harness](/research/context/techniques/16-validation-harness/)
* [Coverage and gap map](/research/context/techniques/18-extension-overview/)
* [Provider-sensitive claim verification](/research/context/techniques/28-drift-delta/)