TTV: Tokens to Value

[joshua-temple]

The Proposal

Adopt TTV: Tokens to Value as The Shrine's north star. Maximize what every token buys, measured per successful outcome, not per API call. Get this right and cost takes care of itself.

Why This Matters

Optimizing raw token spend is optimizing the exhaust pipe. A system tuned to look cheap on the invoice quietly bleeds the expensive resource, engineer hours, in retries, babysitting, and rework.

TTV inverts the question: how much model consumption does it take to produce one successful, useful outcome? Tokens per win, not tokens per month. A system that spends 20% more tokens but succeeds on the first attempt beats one that is cheaper per call and wrong twice.

A north star also has to settle arguments. Without a shared denominator, every governance debate (model upgrades, routing policy, tooling spend) collapses into "that costs more tokens," which rewards exactly the wrong behavior. TTV gives every proposal the same test: does it lower the cost of a win?

Four Levers Move TTV

• Routing. The right model for the task: fast, cheap models for lookups and classification; frontier models for architecture and judgment.
• Model iteration. Every generation rewrites the cost curve. Standing still is paying last year's prices for last year's quality.
• Skills, prompts, memory. Encode a solved pattern once so no session re-pays for discovery.
• Agent architecture. Decompose, delegate, validate, recover. Failed runs are the most expensive tokens we buy.

Open Questions

1. Define the "win." What counts as one successful outcome per workload class (merged PR, resolved ticket, passing eval run)? Who owns each definition?
2. Instrumentation. Can we tag token usage with an outcome ID today (session to PR/ticket), or does that need new plumbing? What is the smallest first step?
3. Baseline. Which two or three existing workflows do we baseline first, so we can prove the metric moves before rolling it out broadly?
4. Failure attribution. When a run fails and retries, how do we attribute those tokens: to the eventual win, or to a separate waste bucket we drive down?
5. Reporting. What is the minimum viable dashboard (TTV per workflow per week?), where does it live, and who reviews it on what cadence?

[joshua-temple]

Accepted. Documented in #13.