ground-game/history/260807-attack-darvo-stress-design.md

# ATTACK / DARVO relevance — design challenge and H1

**Date:** 2026-08-07 (measurement 2026-08-08)  
**Status:** H1 **measured** — reject as baseline; direction kept for H2  
**Triggers:** [GROUND-RPT-0003](../reports/260807-clay-borg-attack-and-modes.md); inbox triage  
**Variant:** [`editions/experiments/h1-problem-stress/`](../editions/experiments/h1-problem-stress/)  
**Measurement:** [GROUND-RPT-0004](../reports/260808-clay-borg-h1-measured.md)  
**Catalog:** [`editions/CATALOG.md`](../editions/CATALOG.md)  
**Workplan:** [GROUND-WP-0006](../workplans/GROUND-WP-0006-h1-problem-stress-experiment.md)

---

## 0. The measured gap

clay-borg (RPT-0003): under competent play, **ATTACK never improves** outcomes in any mode, and in 500 competent games **DARVO armed zero times**. Modes themselves differentiate correctly; the driver bug that forced SHARED GROUND is fixed on their side.

So the namesake loop is reachable only by playing badly. That may be intentional (cautionary structure) or a missing Problem → Stress coupling. This note challenges design options **before** promoting any change into baseline `ground-darvo-r0`.

---

## 1. Stress ledger today (baseline r0)

| Source | Who’s stress changes? | Notes |
|--------|----------------------|--------|
| **ACT_ATTACK** | **Target** +1 or +2 | **Attacker: no relief.** Cancelled → no change |
| **SUPPORT** | **Target** −1 or −2 | Bond Support also readies Freedom, cancels DARVO |
| **GROUND—GR** | **Self** −2 | Readies Freedom; DARVO current stage still fires |
| **DARVO REVERSE** | **Self** −2; **Focus target** +1 | Blame on target; Protection for self |
| **Unsolved Problems** | **Nobody** | `unresolved_effect` is empty in `Problems.csv` |
| **Status / rank** | **Nobody** | No hierarchy stress |
| Start | All at **2** | |

At Stress 4–5 the rules force the *vocabulary* of ATTACK or GROUND (unless Freedom is spent), but ordinary ATTACK is not a stress valve—only GROUND—GR and receiving Support (and later REVERSE) are.

**Implication:** Stress is almost a **closed social subsystem**. If nobody attacks, stress drifts mid/down; DARVO never arms. clay-borg’s zero-DARVO result is the ledger working as printed, not a simulator bug.

INTENT principle (“DARVO must sometimes work **locally** — short-term advantage, delayed relational cost”) is therefore under-served: the only local advantage is REVERSE’s package, gated on already being at Stress 5 for three rounds of sequence.

---

## 2. Design directions (challenged)

### 2.1 Stress from stuck problems + outward projection

**Strong:** Restores Problem → Individual link in the four-layer model. Makes GROUND regulation under load; Support anti-cascade; DARVO “I can’t hold this.”

**Challenge:**

- Today **ATTACK and Blame do not reduce the projector’s stress**. Without a rules change, “forced projection” is flavor on a move that still does not help the projector.
- Passive problem ticks can flatten into always-DARVO or kill group thresholds (already ~75–83% of available points).
- Personal / Bond-attached problems are powerful and already listed as **extensions** (`EXT_PERSONAL`) — keep out of first core experiment.
- Even with self-soothe, **E04** still punishes Bond-breaking; mode-aware evaluation is required.

### 2.2 Hierarchy / competitive status stress

**Strong:** Injects stress without anyone choosing ATTACK; natural in semi-coop.

**Challenge:** Leader −stress amplifies Matthew effects; “leading” mid-game is noisy; co-op must not invent early winners; stacking with problem pressure overproduces Stress 5 at 2p.

**Parked** behind H1 (last-place-only semi-coop is the minimal follow-up if H1 under-delivers).

### 2.3 Competence score

**Strong as extension** (follow-through, self-regard). **Wrong tool** for RPT-0003: does not create ATTACK incentive unless it secretly becomes another stress track. **Out of H1.**

---

## 3. Playtest hypothesis H1

> If unsolved shared Problems raise Stress, and high-Stress ATTACK returns a small self-relief, then competent play will sometimes arm DARVO and sometimes select ATTACK, without making always-ATTACK optimal.

### 3.1 Two rule deltas only (relative to `ground-darvo-r0`)

| ID | Rule | Detail |
|----|------|--------|
| **H1-A** | Problem pressure | At **Round End**, after Stress clamp step preparation: if **any** Problem still in play is **unclaimed** (face-up, hidden, or Denied), **each player +1 Stress**, then clamp 0–5, then DARVO arm check as today. |
| **H1-B** | High-stress ATTACK self-soothe | When an **ACT_ATTACK** resolves and is **not cancelled**, if the **attacker’s Stress was ≥ 4 before this Attack’s effects**, the attacker takes **−1 Stress** (then clamp). Target / relation effects unchanged. DARVO-stage extra Attack uses the same Attack resolution (so it can self-soothe too if Stress ≥ 4). |

Unchanged on purpose: thresholds, deal, SOLVE legality, Support, GROUND modes, DARVO stage table, relation slots, starting Stress 2.

### 3.2 Success criteria (compare to RPT-0003 harness)

Under a competent multi-policy panel, same seat counts as RPT-0003 where practical:

1. **DARVO arm rate** non-trivial (target band to refine after first N; not 0/500).
2. **ATTACK selection** rises for some subpopulation (e.g. Stress ≥ 4) and **raising ATTACK rank** improves at least one objective in some mode/seat cell — not “always attack wins.”
3. **Group threshold success** does not collapse to “only perfect peaceful play.”
4. Bond Support and GROUND remain better than DARVO **when regulation holds** — DARVO stays the failure mode, not the dominant strategy.

### 3.3 Explicit non-goals for H1

- No personal / bond-group problem cards.
- No competence track.
- No leader/last-place status stress (Idea 2).
- No silent edit of baseline `point_value` / `required_solution` inside r0.
- No promotion of H1 into `ground-darvo-r0` until measured utility is accepted.

---

## 4. Rules variation versioning (how we compare experiments)

See full contract: [`editions/CATALOG.md`](../editions/CATALOG.md).

**Summary:**

| Mechanism | Role |
|-----------|------|
| **`editions/catalog.yaml` + `CATALOG.md`** | Authoritative **index of selectable packages** (baseline + experiments). Lives on `main`. clay-borg and humans read this to know what can be pinned. |
| **Directory path** | Selectable unit, e.g. `editions/ground-darvo-r0/`, `editions/experiments/h1-problem-stress/`. Matches clay-borg’s vendor-by-path practice (ADR-0011). |
| **`variant_id` + `base`** | Stable ids (`ground-darvo-r0`, `h1-problem-stress`); experiments declare inheritance and file-level deltas. |
| **Git** | Commits/tags pin content digests; optional feature branches for WIP **before** an experiment directory lands on `main`. Branches are **not** the primary selection API (sibling checkout + multi-ref is hostile to clay-borg CI). |
| **Utility fields** | Catalog rows carry `status`, `utility_estimate`, `decision` (keep / reject / merge-to-baseline / supersede). Updated when measurement lands — not when the idea is written. |

Baseline contract (GROUND-WP-0002 T01) still holds: changing frozen columns inside `ground-darvo-r0` requires a **new revision directory** (`-r1`), not an experiment. Experiments may **copy** and alter rules text/actions; only a later **promote** step rewrites baseline.

---

## 5. clay-borg selection contract

A package is **selectable for simulation** when it appears in `editions/catalog.yaml` with `selectable: true` and a `path`.

Minimum fields clay-borg should read (or we message explicitly):

```yaml
variant_id: h1-problem-stress
path: editions/experiments/h1-problem-stress
base: ground-darvo-r0
status: experimental
rules_engine_notes: rules_delta.yaml   # structured deltas if CSV alone is not executable
```

Engine note: today clay-borg **hardcodes** much of the rules kernel and **vendors** selected CSVs. H1-A/B are **rules-kernel changes**, not Problem table changes. Selecting H1 means:

1. Vendor or overlay the experiment’s changed CSVs for text/provenance, **and**
2. Implement `rules_delta.yaml` in the kernel (or a feature flag keyed by `variant_id`).

Until (2) exists, H1 is **specified here** and playable at a human table from the experiment package; bulk ATTACK-value re-runs wait on clay-borg.

---

## 6. Exploration order (after H1)

| Step | If… | Then… |
|------|-----|--------|
| H1 measured | criteria met | promote candidates into r1 design / keep as optional mode card |
| H1 measured | DARVO still ~0 | add Idea 2 last-place stress (semi only) as **H2** experiment |
| H1 measured | always-ATTACK optimal | reject H1-B or gate self-soothe tighter |
| Later | core loop works | competence / personal problems as extensions |

---

## 7. Decision log

| Date | Decision |
|------|----------|
| 2026-08-07 | Do not invent “DARVO-only-from-bad-play is intended” without measurement of H1-class pressure. |
| 2026-08-07 | First experiment is **H1** (problem pressure + high-stress ATTACK −1). |
| 2026-08-07 | Versioning = **catalog + directories on main**; git tags pin; branches optional for WIP. |
| 2026-08-08 | H1 measured (CB-EV-0030/0031 → RPT-0004). **Reject as baseline.** Direction (Problem→Stress) confirmed useful; flat +1/end is a solve-rate tax. H2 must change pressure *shape*, not only re-stack. Baseline finding: non-attacking policies never raise Stress above 2. |
| 2026-08-08 | **H2 packaged:** scoped stress (personal / bond / global). See [260808-h2-scoped-problem-stress.md](260808-h2-scoped-problem-stress.md). Bond network shares ticks → joint SOLVE incentive. |

## 8. Measurement summary (2026-08-08)

See RPT-0004 for full tables. One-line:

> H1 reaches the seat that will not regulate; under greedy regulation the table stops winning; H1-B is dead code for competent play.