diff --git a/.agent-taskgraph/STATUS.md b/.agent-taskgraph/STATUS.md index de8b7c57..3d0e9958 100644 --- a/.agent-taskgraph/STATUS.md +++ b/.agent-taskgraph/STATUS.md @@ -7,7 +7,12 @@ - Baseline at formation: clean worktree; formation source `894c6ed4`; current local HEAD is the M102 hash-closure worktree (not pushed) - Completed handoffs: T-001 authority audit; T-002 QA/evidence baseline. - Completed tasks: T-006 current-build smoke at `HEAD` (PASS); T-007 current-HEAD owner handoff video (PASS). -- Current active gate: M104 P2 (full proxy recalibration) is in progress. +- Current active gate: M105 P1 (rank-authority alignment) is in progress. + The Owner selected the formal `docs/02` §8.1 `offer_level + risk` model. + P1 is limited to the enemy authoring boundary and a focused deterministic + proof; bridge/probe integration is M105 P2 and is not started yet. + Gameplay, economy, content, Save, UI, art, audio, and owner review gates + remain frozen. M104 P1 repaired the encounter class authority and passed both its focused marker and the M60 regression. The repaired density sweep reports `wall_late=7`; isolated enemy-count candidates reduce it to 4 and then 1, @@ -18,17 +23,14 @@ per cell) and still fails the late-clear target. Authority review confirms it is a standard-ladder, not economic-shape, proxy. The read-only formula audit is complete: 72/72 cells differ, with R12 ordinary rank 4 versus - current passed levels 10–24 plus region pressure. M104 is blocked at this - authority boundary; no global adjustment, economic proxy or production - change is accepted until the Owner/product authority selects the formal - model and a new narrow plan is opened. M60/P10.0 first-cycle behaviour - remains frozen. + current passed levels 10–24 plus region pressure. M104 remains historical + evidence, not an approved production balance change. M60/P10.0 first-cycle + behaviour remains frozen. - Current blocker: Owner playtest verdict, human listening, six-question comprehension and duplicate-portrait decision remain open downstream gates; do not claim MVP visual readiness from automation alone. - Runtime note: Product/System, UI, and QA are ordinary visible HAPI OpenCode sessions; this workflow does not use Team mode or native sub-agents. -- Next event: Owner/product authority selects §8.1 `offer_level + risk` or the - M80 regional model, after which a new narrow plan can be opened. Compress the remaining Owner - gate to the captured BAT-002 states and the six-question - comprehension/listening sheet. Do not turn proxy numbers into production - balance without the recorded decision. +- Next event: finish M105 P1, then open only its P2 bridge/probe step after the + focused marker is green. Compress the remaining Owner gate to the captured + BAT-002 states and the six-question comprehension/listening sheet. Do not + turn historical proxy numbers into production balance. diff --git a/CURRENT_CHECKPOINT.md b/CURRENT_CHECKPOINT.md index 9cc43b02..0c919c34 100644 --- a/CURRENT_CHECKPOINT.md +++ b/CURRENT_CHECKPOINT.md @@ -11,7 +11,16 @@ - Program Goal, active repair contract, implementation plan and ledger: `docs/06_DEVELOPMENT_AND_REVIEW.md` -## Active Plan (M104 Balance Curve Proxy Repair — P2 in progress) +## Active Plan (M105 Rank Authority Alignment — P1 in progress) + +- Durable plan: `docs/plans/M105_RANK_AUTHORITY_ALIGNMENT.md`. +- State (2026-09-20): **P1 IN_PROGRESS**. The Owner selected the formal §8.1 + `offer_level + risk` rank authority. P1 is limited to the enemy authoring + boundary and its focused deterministic proof; runtime bridge/probe wiring is + P2 and remains pending. Gameplay, economy, content, Save, UI, art, audio, + and owner review gates remain frozen. + +## Previous Plan (M104 Balance Curve Proxy Repair — authority boundary) - Durable plan: `docs/plans/M104_BALANCE_CURVE_PROXY_REPAIR.md`. - State (2026-09-20): **P2 IN_PROGRESS**. P1 repaired the M80 encounter-class diff --git a/docs/plans/M105_RANK_AUTHORITY_ALIGNMENT.md b/docs/plans/M105_RANK_AUTHORITY_ALIGNMENT.md new file mode 100644 index 00000000..1a59699b --- /dev/null +++ b/docs/plans/M105_RANK_AUTHORITY_ALIGNMENT.md @@ -0,0 +1,84 @@ +# M105 · Rank Authority Alignment + +## Player-visible stopping point + +The formal M80/entry battle simulation receives enemy rank from the gameplay +authority in `docs/02_GAMEPLAY_AND_BALANCE.md` §4.1 and §8.1: + +```text +offer_level(R) = min(10, 1 + floor((R - 1) / 3)) +encounter_rank = min(6, offer_level + Q + (boss ? 1 : 0)) +rank_multiplier = 1 + 0.08 * (encounter_rank - 1) +``` + +Rank is applied once to the authored enemy base budget. The old ladder-level +plus permanent region-step proxy remains available only as historical evidence; +it is not a second production authority. The player-visible stopping point is +one deterministic focused proof plus first-cycle regression evidence showing +that the formal entry route uses the same rank context at every region. + +## Authority decision + +- On 2026-09-20 the Owner replied “好” to the recommended choice A: §8.1 + `offer_level + risk` is the formal rank authority. +- `docs/02_GAMEPLAY_AND_BALANCE.md` has priority over historical M80 proxy + calibration notes. +- M104 evidence remains historical and is not silently reinterpreted as a + production balance decision. + +## Ordered steps + +### P1 · Authoring contract and focused proof — IN_PROGRESS + +- Add one explicit rank-context boundary at enemy authoring. +- Remove hidden permanent region-step scaling from the formal path. +- Preserve class identity and the existing ordinary/elite/boss base budgets. +- Add a focused deterministic marker covering R1/R12, risk 0/2, all classes, + and the negative region-only-scaling assertion. + +Acceptance: the focused source-bound attempt passes with one marker, strict +diagnostics empty, and exact expected rank/stat values. + +### P2 · Runtime bridge and probe integration — PENDING + +- Pass `market_index` and `risk` from the committed battle snapshot through + `battle_service.gd` into `battle_sim_bridge.gd`. +- Update M61/M80 probes to use the same context by default; retain any legacy + level model only behind an explicit historical switch. +- Do not change encounter counts, class schedules, economy, content, or UI. + +### P3 · First-cycle regression and handoff — PENDING + +- Re-run focused bridge/class coverage and the M60 first-cycle route. +- Compare committed battle/replay hashes and diagnostics with the pre-M105 + baseline. +- Record the owner gate as still open; this package does not claim MVP ready. + +## Frozen scope + +Enemy counts, elite scheduling, region mechanics, economy, XP, equipment, +content catalogs, Save semantics, UI/art/audio, P10.0 wave semantics, and all +owner comprehension/listening/feel gates remain frozen. Packaging and release +are out of scope. + +## Verification and evidence + +Every formal command runs through `tools/run_test_attempt.py` with an owned +source/HOME/cache/TMP/browser root and `--require-marker`. Retain the compact +attempt directory and append the marker, source commit, terminal, and strict +diagnostics to `docs/EVIDENCE_LEDGER.md` before advancing a step. + +## Issue register + +- M105-I01: M104 found that the old proxy passed ladder level plus region + pressure, while §8.1 expects `offer_level + risk`; this is the P1 boundary. +- M105-I02: the entry bridge currently has class metadata but not rank context; + P2 must carry the committed snapshot fields without changing the snapshot. +- M105-I03: owner playtest, listening, comprehension, duplicate portrait, and + visual-feel gates remain open after the technical proof. + +## Handoff + +At the stopping point, commit the narrow package, update +`CURRENT_CHECKPOINT.md`, `.agent-taskgraph/STATUS.md`, and the evidence ledger, +then stop for the next owner/product decision or a new plan. diff --git a/docs/reports/M104_CURVE_DECISION_MEMO_20260920.md b/docs/reports/M104_CURVE_DECISION_MEMO_20260920.md index 40910a65..163ebf96 100644 --- a/docs/reports/M104_CURVE_DECISION_MEMO_20260920.md +++ b/docs/reports/M104_CURVE_DECISION_MEMO_20260920.md @@ -2,7 +2,11 @@ 日期:2026-09-20 决策对象:是否把 M80 代理暴露出的后段墙转成正式战斗难度改动 -当前结论:**暂不改正式曲线;先解决敌方 rank 公式的权威冲突,再校准队伍形状。** +当前结论:**不改正式曲线;Owner 已选择 §8.1 `offer_level + risk` 作为正式 rank 权威,进入 M105 对齐实现。** + +Owner 决策记录(2026-09-20):对推荐选项 A 的“好”按项目上下文解释为 +采用 `docs/02_GAMEPLAY_AND_BALANCE.md` §4.1/§8.1。M104 的代理结果保留为 +历史诊断;M105 先修复 authoring/bridge 的权威边界,再决定是否需要经济形状实验。 ## 这份 memo 要做的唯一决策 @@ -38,7 +42,7 @@ | A. 直接降低全局等级/区域倍率 | **拒绝** | 已经出现高区过软、R3 仍为 0 的组合;会破坏区域压力与机制身份。 | | B. 现在就写分段曲线 | **暂缓** | 方向可能正确,但 F3 的代理输入尚未符合经济权威,决策证据不干净。 | | C. 先做经济形状代理,再决定曲线 | **后置** | 可逆、只改探针;但必须先确认敌方公式不是另一套历史模型。 | -| D. 先做权威公式对照审计 | **现在执行** | F6 可能是最早失败边界;若不先确认实现应服从哪套公式,经济代理和曲线实验都会在错误的敌方压力上运行。 | +| D. 先做权威公式对照审计 | **已完成** | F6 是最早失败边界;审计后 Owner 选择 §8.1,进入 M105 对齐实现。 | ## 下一次实验契约 @@ -46,7 +50,9 @@ 实验顺序必须先完成 D,再进行 C。D 已完成:72 个区域×回合×风险格全部出现公式差异,marker 为 `ABG_M104_RANK_AUTHORITY_AUDIT_OK`,且 `production_unchanged=true`。这证明当前 M80 authoring 不是可以继续盲调的产品权威,而是需要单独决策的历史/中间模型;在该决策前不再运行经济形状或分段曲线实验。 -如果 Owner/产品权威确认 §8.1 是正式规则,则必须先开一个新的窄计划,将 `battle_stat_authoring.gd`、正式桥和 M80 探针统一到 `offer_level + risk`,并保留首轮 P10.0 回归。若确认 M80 的区域纵向模型才是正式规则,则必须反向更新 `docs/02` 和相应验收公式,再重新定义 40–70% 目标。当前 M104 不替任何一方做这个决策。 +Owner 已确认 §8.1;M105 已开为新的窄计划,先将 `battle_stat_authoring.gd`、正式桥和 +M80 探针统一到 `offer_level + risk`,并保留首轮 P10.0 回归。M80 的区域纵向模型 +仅作为历史探针证据,不再作为生产权威,也不反向改写 `docs/02`。 经济形状代理(在 D 通过后)至少要记录并固定: @@ -70,4 +76,4 @@ ## 最短下一步 -**由 Owner/产品权威选择 §8.1 还是 M80 区域纵向模型;在选择和新计划建立前,不提交任何正式难度数值。** +**完成 M105 的 authoring focused proof,再进入桥接和探针整合;在此之前不提交任何正式难度数值。**