# Three-milestone efficiency pilot

Preserve the complete game scope and required checks. Use small relevant reads,
coherent feature batches, selective subagents and compact verification artifacts.
Git remains mandatory: review, verify, commit, push and confirm remote success at
each finished implementation checkpoint. Never force-push or commit credentials.
The removed recurring automation stays removed.

| Milestone | Deliverable | Measurement |
| --- | --- | --- |
| 1: setup | Current status, decision index, reusable checks, saved-game snapshots/backups and HTTP seat restoration | Accepted checkpoint `6e26ba5`. Aggregate counters 57,957,554 → 58,124,416: 166,862 tokens including planning, coordination and the user-requested save reset. Two bounded helpers. Fifteen added regressions; 4,177 offline tests passed. Three broad attempts: two stopped before tests on type/lint errors; final check passed in 130.147s. One cancellation defect found and fixed in review. |
| 2: core readiness batch | Typed Advanced reserve-shipment commitments without Guild or optional modules | Accepted checkpoint `32791ed`. Aggregate counters 58,124,416 → 58,695,053: 570,637 tokens. Two bounded helpers: regression implementation and independent rules/privacy review. Fifteen added tests; 4,192 offline and 40 HTTP tests plus build passed. One stable broad run; focused checks resolved setup mistakes and two review findings before it. Two browser rooms verified distinct AI and human paths. Pending Guild settlement remains explicit. |
| 3: next dependency batch | Truthtrance total hand size and primary-role counts, controls, bots and recovery | Accepted checkpoint `7454456`. Aggregate counter 58,695,053 → 59,144,653 at completed verification: 449,600 tokens, before final Git/push overhead. Two workhorse-model helpers for regression implementation and independent review. Reused count panel and SQLite fixtures; eight added tests and expanded recovery cases. Two copy issues fixed before the single stable broad run: 4,200 offline tests, build and 40 HTTP tests passed. One browser room covered combined roles, total size, invalid input and phone refresh. |

The aggregate counter may update only between turns. Record a confirmed endpoint
when available rather than estimating missing usage. It includes coordination and
conversation overhead and is not a dollar-cost meter. Different capability sizes
limit direct comparisons; do not promise a savings percentage from this setup.
Do not count tests alone as accepted faction or mode completion. Scripts consolidate
existing checks; broad checks still run once source/review is stable and repeat if
subsequent changes or failures justify them. Browser fixtures remain feature-specific;
the reusable HTTP restoration helper does not automate browser setup or visual QA.

Milestone 2 cost more than the tooling baseline and cannot establish a saving.
Its larger rules/interaction scope, browser setup and context recovery contribute
to that total. Milestone 3 was smaller and reused more existing structure; its lower
count cannot isolate the effect of model choice or establish a savings rate.
Avoid repeatedly loading old audits. Three different deliverables alone cannot
support a reliable percentage estimate.

For initial planning, allow roughly **450,000–575,000 aggregate tokens for a
similarly sized reviewed gameplay batch**, based on milestones 2 and 3. This is a
low-confidence planning estimate, not a dollar quote, token-budget setting or a
reason to stop required work. Larger interaction sets need a revised estimate.
The tooling baseline was about 167,000 and is not a gameplay cost target. Reassess
after the next few comparable batches, including defects found after acceptance.

Next adjustment: use the owner plus one bounded helper for source research and
independent review when the existing implementation needs verification. Reuse
the same helper for its follow-up review, existing fixtures and applicable checks.
Avoid adding runtime code merely to make an audit appear to be implementation.
The full-plan interaction audit starts at aggregate counter 59,151,671, uses that
one-helper pattern and changes tests/documentation. It therefore requires the
full offline check, but can reuse the unchanged application's earlier build;
browser/restart acceptance supplies direct recovery evidence.

Check account usage after each verified checkpoint. The user explicitly confirmed
the stop point as **80% weekly usage consumed (20% remaining)**. Keep the goal
unfinished when stopping for usage, preserve Git and subagent requirements, and
leave the removed recurring automation absent. After `32791ed`, usage was 3% used.

Use [current status](CURRENT_STATUS.md), [rule decisions](RULE_DECISIONS.md) and
[verification workflow](VERIFICATION_WORKFLOW.md) as short entry points. Historical
logs remain references, not mandatory full-context input on every turn.