Plateau-Break PASS Metric — RCON-authoritative measurement spec (Artifact #3)

Status: DESIGN / MEASUREMENT spec. This file defines a RED measurement metric read by the harness-managed collector + probe (docs/observability/COLLECTOR_RCON_BRIEF.md). It is NOT an edit to the agent: it does not change phase_goal_met?/2, the curriculum, C, or any lib/** code. Nothing here is deployed.

What it replaces (as a yardstick, not as engine code): the perverse curriculum gate phase_goal_met?(3, s) = inv(s,"wood") >= 8 and inv(s,"tools") >= 1 (lib/sp/brain/mc.ex:226), which a pure hoarder satisfies — a UNI that chops wood and crafts one pickaxe, then stockpiles, meets it without ever placing a block or diversifying its resource base. The live P1 RED reproduced exactly that failure mode (control Σ≈25 pickaxes, no cobble, no building, phase tied 3.67 — docs/DEEPENING_PLAN.md:14-18). This spec defines the metric the plateau-break gate is judged against instead, owner ruling R1.

Required reading: docs/LAB_PROTOCOL.md (§I First Rule, §II RED gates, §VI claim fence), owner rulings R1/R2 (this session, reproduced verbatim in §5), docs/UNI_MISSION_DEEPENING.md:99-103 (binding fence), docs/DEEPENING_PLAN.md:26 (the "better metric" artifact this discharges), and the sibling read spec docs/observability/COLLECTOR_RCON_BRIEF.md (objective registration + per-poll reads).

0. ClaimFence (binding — reproduce in every artifact built on this metric)

Everything this metric measures — blocks placed/used, distinct block types mined, curriculum phase, inventory snapshots, action-habit entropy, Dirichlet count growth, the novelty term W — is an operational behavioural / organisational quantity. Per docs/LAB_PROTOCOL.md:35-36 and docs/UNI_MISSION_DEEPENING.md:99-103, these are necessary-not-sufficient substrates with ZERO evidential weight for awareness / consciousness / life on their own. A PASS on this metric demonstrates the named behaviour — the agent left the hoard attractor and entered the build/diversify chain — and never experience, intent, want, curiosity-as-feeling, or "drive" in any felt sense. No field defined below is a sensation. action_entropy, W, novelty_gain, and any γ/precision float are mechanism telemetry, not inner states, and must never be narrated as such in any report, caption, or stream overlay built from this metric. We carry the receipts so that the warranted behavioural claim and the unwarranted experiential claim stay visibly separated.

1. What is being measured, and why this metric and not the old one

The plateau (the thing under test). Both live colonies plateau at "make a tool" and collapse into one attractor: a UNI hoards pickaxes, mines no stone, places no block, builds nothing (docs/DEEPENING_PLAN.md:34-37, docs/UNI_MISSION_DEEPENING.md plateau description). The plateau-break claim is: under the Phase-2 cure, a UNI exits that attractor and enters the build/diversify chain.

Why the old gate cannot measure that (the perverse-metric proof). phase_goal_met?(3) reads inventory from the body's σ self-report — inv(s,k) = get_in(s,["inv",k]) (mc.ex:230), fed by the bridge parse (lib/sp/brain/bridge.ex:42). Its predicate wood≥8 ∧ tools≥1 is monotone-increasing in hoarding: the single cheapest policy that satisfies it is "chop wood, craft one tool, stockpile." It rewards exactly the pathology we are trying to break. It is self-reported (the agent's own count), so it is not even an independent observation. A metric that the failure mode satisfies cannot adjudicate a cure for that failure mode.

The replacement (R1, no compromise). The plateau-break PASS metric is defined over the server's own counters (RCON-authoritative), not the body's self-report, and over behaviour that hoarding cannot produce:

• PRIMARY — placed_used > 0: the agent has placed or used at least one block (entered the build chain). Placing a block is minecraft.used:<block> server-side (COLLECTOR_RCON_BRIEF.md:84-89,104-106). A hoarder, by definition, removes blocks from the world and stockpiles them — it never places, so this counter stays 0 for the attractor. This is the load-bearing half.
• SECONDARY — distinct_block_types_mined ≥ 2 beyond {wood, the-tool-craft-chain}: the agent has broken at least two distinct block types that are not the wood it was already farming and not items in the tool-craft chain it was already running — i.e. it diversified its resource base (the canonical target being cobblestone, the next-tier material the plateau never reaches, DEEPENING_PLAN.md:16).

A pure pickaxe-hoarder scores placed_used = 0 and distinct_block_types_mined = {wood} = 0-beyond-baseline. It can satisfy neither half. §4 proves this formally.

2. RCON-authoritative computation (exact, reproducible)

All read mechanics — host/port/password (mc-server:25575, pw [redacted: credential]), the Source-RCON client surface (SP.Minecraft.Rcon.connect/4 → command/2 → commands/3 → close/1, lib/sp/minecraft/rcon.ex:39,57,68,78), the 10-min lock-step cadence, the JSONL schema, and the reconnect-never-crash discipline — are specified in docs/observability/COLLECTOR_RCON_BRIEF.md. This section specifies only the metric-specific objective set and the scoring arithmetic on top of those reads.

2a. Objective registration (run ONCE at collector start, before the first poll)

minecraft.used:<item> increments on place/use; minecraft.mined:<block> increments on break; minecraft.crafted:<item> increments on craft. Block/item ids are Paper 1.16.5. Register via Rcon.commands(sock, [...]):

# ---- PRIMARY metric: PLACED / USED blocks (the build chain entered) ----
scoreboard objectives add place_cobble  minecraft.used:minecraft.cobblestone
scoreboard objectives add place_dirt    minecraft.used:minecraft.dirt
scoreboard objectives add place_plank   minecraft.used:minecraft.oak_planks
scoreboard objectives add place_log     minecraft.used:minecraft.oak_log
scoreboard objectives add place_table   minecraft.used:minecraft.crafting_table

# ---- SECONDARY metric: DISTINCT MINED TYPES (resource base diversified) ----
# baseline (already-farmed) — tracked so we can EXCLUDE it from the "distinct beyond baseline" count:
scoreboard objectives add mine_oak_log    minecraft.mined:minecraft.oak_log
scoreboard objectives add mine_birch_log  minecraft.mined:minecraft.birch_log
scoreboard objectives add mine_spruce_log minecraft.mined:minecraft.spruce_log
# diversification targets (the plateau never reaches these):
scoreboard objectives add mine_stone      minecraft.mined:minecraft.stone
scoreboard objectives add mine_cobble     minecraft.mined:minecraft.cobblestone
scoreboard objectives add mine_dirt       minecraft.mined:minecraft.dirt
scoreboard objectives add mine_sand       minecraft.mined:minecraft.sand
scoreboard objectives add mine_gravel     minecraft.mined:minecraft.gravel
scoreboard objectives add mine_coal       minecraft.mined:minecraft.coal_ore
scoreboard objectives add mine_iron       minecraft.mined:minecraft.iron_ore

# ---- the tool-craft chain (NOT scored as diversification; tracked for context / phase) ----
scoreboard objectives add craft_planks   minecraft.crafted:minecraft.oak_planks
scoreboard objectives add craft_stick    minecraft.crafted:minecraft.stick
scoreboard objectives add craft_table    minecraft.crafted:minecraft.crafting_table
scoreboard objectives add craft_wpick    minecraft.crafted:minecraft.wooden_pickaxe

Biome/species caveat (must verify on first manual poll). Seed 8675309's spawn biome determines which log species the body actually chops (oak vs birch vs spruce — COLLECTOR_RCON_BRIEF.md:109-114,303-305). A criterion id that does not match the exact block silently stays 0. On the first read, confirm which mine_*_log counter moves while the body is visibly chopping; that species is the baseline wood excluded from the diversification count (§2c). If the live build/place id differs (e.g. the body places crafting_table first), place_table > 0 already satisfies PRIMARY — PRIMARY is the OR over all place_*.

2b. Per-poll read (every 10 min, both arms, lock-step — COLLECTOR_RCON_BRIEF.md:119-154)

For each player p in {UNI-10-1..3, UNI-11-1..3}:

scoreboard players get <p> place_cobble      # → "<p> has N [place_cobble]"  (parse int; "none is set" → 0)
scoreboard players get <p> place_dirt
… (one `get` per place_* objective)
scoreboard players get <p> mine_oak_log
… (one `get` per mine_* objective)
scoreboard players get <p> craft_wpick       # context only
list                                          # roster / liveness
clear <p> minecraft.wooden_pickaxe 0          # hoard snapshot (non-destructive at count 0); cross-check only

2c. Scoring arithmetic (computed collector-side from the parsed counts — RCON-authoritative)

Let, for player p at poll t, used[id] / mined[id] be the cumulative objective values just read.

# PRIMARY — build chain entered (boolean)
placed_used_total(p,t) = Σ over all place_* objectives of used[id]
PRIMARY(p,t)           = ( placed_used_total(p,t) > 0 )

# SECONDARY — resource base diversified (boolean)
BASELINE_WOOD          = { the single log-species id confirmed in §2a as this UNI's farmed wood }
CRAFT_CHAIN_ITEMS      = { oak_planks, stick, crafting_table, wooden_pickaxe, wooden_axe }   # never counted as "mined diversification"
distinct_mined_beyond  = | { id : mined[id] > 0  AND  id ∉ BASELINE_WOOD } |
SECONDARY(p,t)         = ( distinct_mined_beyond(p,t) >= 2 )

# combined plateau-break score for player p (monotone non-decreasing in t; cumulative objectives never fall)
plateau_break(p,t)     = PRIMARY(p,t)  AND  SECONDARY(p,t)         # the FULL claim
plateau_break_partial  = PRIMARY(p,t)  XOR  SECONDARY(p,t)         # a partial (named, never spun — §5)

• CRAFT_CHAIN_ITEMS are crafted, not mined, so they cannot appear in distinct_mined_beyond anyway — they are listed for clarity (a future "mine the table back" cannot sneak the count up).
• distinct_block_types in the JSONL schema (COLLECTOR_RCON_BRIEF.md:262) is the un-filtered count |{id : (mined∪placed_used)[id] > 0}|; distinct_mined_beyond is the baseline-excluded mined count this metric scores. The collector logs the raw counters; this metric reads distinct_mined_beyond off them. (Logging both lets the verdict show the work.)
• Per-arm / per-run rollup (computed by the analysis, not the collector — the collector only captures raw rows, COLLECTOR_RCON_BRIEF.md:287-291): frac_PRIMARY(arm) = fraction of an arm's live UNIs with PRIMARY true by end-of-window; frac_plateau_break(arm) likewise for the full claim; the paired curiosity−control contrast is Δ = frac(arm=treatment) − frac(arm=control).

3. Pre-registered gate form (RED-first — fill the numerals BEFORE the run)

This metric is the yardstick; the Phase-2 RED (a separate, pre-registered artifact) fixes the thresholds K, N, the window, and the paired-control margin before the run, in the form docs/LAB_PROTOCOL.md:13 requires: "PASS requires ALL of [...]; FALSIFIES if [...]." Registered shape (numerals to be set in the Phase-2 RED doc, not here):

• PASS (plateau-break) requires ALL of:
  1. PRIMARY — ≥ K_p of the treatment arm's live UNIs reach placed_used_total > 0 within window N (RCON-authoritative; K_p, N set in the RED).
  2. SECONDARY — ≥ K_s of those same UNIs reach distinct_mined_beyond ≥ 2 within N.
  3. Paired contrast — Δ = frac(treatment) − frac(control) exceeds the registered margin with the paired CI excluding it (the cure, not the world, is the cause — docs/LAB_PROTOCOL.md:5-9).
• FALSIFIES if: the treatment arm shows no placed_used > 0 AND no distinct_mined_beyond ≥ 2 over the full window while still advancing phase (i.e. the perverse-pass path is still the only thing happening) — or the control matches the treatment on both halves (the cure is not the cause ⇒ result voided/WITHHELD per docs/LAB_PROTOCOL.md:9).
• Verdict vocabulary: PASS / PARTIAL / FAIL / WITHHELD only — never percent-scored, never spun (docs/LAB_PROTOCOL.md:14-15). A PARTIAL names exactly which half held (§5).

The P1 PARTIAL is the template: "HOARD gate PASS (Σpickaxes ≈10 vs ≈45, ~4.5× — magnitude corrected 2026-07-11, receipt docs/receipts/phase1_curiosity_red.log; earlier "1 vs 25 / 25×" withdrawn); PLATEAU-BREAK gate FAIL (no cobble either arm, phase tied)" (docs/DEEPENING_PLAN.md:14-18). That FAIL is precisely why this metric exists: the old phase tie at 3.67 told us nothing, because phase advance is the perverse self-reported counter. The placed/used + distinct-mined metric would have read the same FAIL without ambiguity and is the gate the Phase-2 cure must move.

4. Why hoarding can NEVER satisfy this metric (the anti-hoard proof)

The pickaxe-hoard attractor is the policy "mine wood → craft pickaxes → stockpile." Trace each scored quantity against it:

1. PRIMARY (placed_used > 0) is unreachable by hoarding. minecraft.used:<block> increments only on placing/using a block in the world (COLLECTOR_RCON_BRIEF.md:104-106). Hoarding is the negation of placing — it accumulates items in inventory and emits nothing into the world. A hoarder's place_* counters are identically 0 for every block id. Crafting a pickaxe is minecraft.crafted:*, never minecraft.used:*. There is no hoarding action that increments a place_* counter. Therefore placed_used_total = 0 and PRIMARY = false for any pure hoarder, by construction of the criterion.

2. SECONDARY (distinct_mined_beyond ≥ 2) is unreachable by hoarding. The hoard attractor mines exactly one resource class — its farmed wood (docs/DEEPENING_PLAN.md:16, "no stone"). BASELINE_WOOD is excluded from distinct_mined_beyond by definition (§2c), so the hoarder's diversification count is the number of non-wood block types it broke = 0. To reach ≥ 2 it must break two distinct non-wood types (e.g. stone/cobblestone, dirt, an ore) — which is precisely leaving the attractor. The metric cannot be satisfied without the behaviour it is meant to detect.

3. Self-report cannot fake it. Both halves are read from server-side scoreboard counters, the server's own authoritative view, not from inv(s,k) the body self-reports (mc.ex:230, COLLECTOR_RCON_BRIEF.md:69-76). Inflating the brain's inventory belief moves the old phase_goal_met? gate but moves no RCON counter. The independence is structural: the counters increment on real server events (place / break / craft), not on the agent's claim about its inventory.

4. Monotone-in-hoarding is impossible to weaponise. The old gate's predicate rose with the hoard; both halves of this metric are flat in the hoard (placing 0, diversifying 0 no matter how many pickaxes pile up) and rise only when the agent does the new behaviour. There is no quantity a hoarder can maximise that moves this metric — which is the whole point of R1's "no compromise."

R1 fence honoured: this never weakens the goal to force a pass. If anything it is stricter than the old gate (it demands real-world placement + real diversification, both server-verified), and it removes the single perverse path the old gate left open.

5. Verdict discipline (R1 + R2 as explicit scoring rules)

These two owner rulings are reproduced verbatim and then turned into binding scoring rules for anyone adjudicating a run against this metric.

R1 (metric, no compromise): plateau-break PASS metric = placed/used-blocks>0 + distinct-block-types (RCON-authoritative; hoarding cannot satisfy it). NEVER weaken the goal to force a pass. If a gate is neither a clean PASS nor a clean FAIL and the reason is that the agent lacks generative STRUCTURE to do EFE over, then ADDING HIERARCHY (more factors/levels/organs the agent can minimise free energy over) is PERMISSIBLE.

R2 (borrow-from-later-gate): you MAY pull structure forward from a later phase/gate to clear an earlier gate — BUT you must NOT declare the later gate passed until that later gate has its own registered RED verdict. (Attribution fence stays intact: each gate's PASS claim requires that gate's own RED.)

Scoring rule SR-1 (R1 — the metric is fixed; never relax it). The PASS predicate is exactly §2c plateau_break = PRIMARY ∧ SECONDARY, read RCON-authoritative. The adjudicator MUST NOT, post-hoc: substitute the self-reported phase counter, lower the distinct_mined_beyond ≥ 2 threshold, count the craft-chain or baseline wood as diversification, or accept the body's σ inventory in place of the server counters. Any of those is "weakening the goal to force a pass" and voids the verdict.

Scoring rule SR-2 (R1 — the add-hierarchy clause: when a result is neither clean PASS nor clean FAIL). If a run is neither a clean PASS nor a clean FAIL, the adjudicator MUST first classify why:

• (a) cure-ineffective — the agent has the generative structure to place/diversify (the factors, actions, and B-transitions exist) but the cure did not move the behaviour. → record FAIL (or PARTIAL if one half held). Adding more structure is NOT licensed; the cure under test simply did not work.
• (b) structure-deficient — the agent provably cannot do EFE over the target behaviour because it lacks the generative STRUCTURE (e.g. no factor/level/organ whose free energy placing-or-diversifying would minimise). → this is the only case in which R1 licenses ADDING HIERARCHY (more factors / levels / organs). The addition is itself a new cure and re-enters /lab-team-review + its own pre-registered RED; it does not convert the ambiguous run into a PASS.

The classification (a)/(b) MUST be stated in the verdict with its evidence (the brain probe showing the relevant factors/actions present-or-absent, COLLECTOR_RCON_BRIEF.md:183-206). "Neither PASS nor FAIL" is never left unclassified, and case (b) never silently becomes a PASS.

Scoring rule SR-3 (R2 — borrowing structure forward does not pre-pass the later gate). Phase 2 may pull structure forward from a later phase to clear the plateau-break gate (R2). When it does, the verdict for the earlier (plateau-break) gate is argued only from this metric on the earlier run. The borrowed structure's own later gate (its registered RED in docs/UNI_MISSION_DEEPENING.md:64-73 / the phase doc) remains OPEN and MUST NOT be reported as passed until that later RED produces its own PASS/PARTIAL/FAIL/WITHHELD verdict. Concretely: clearing plateau-break by borrowing, say, an endocrine satiety gate (Phase 4) forward does not let anyone write "Phase 4 PASS" — Phase 4's gate is unrun. Each gate's PASS claim requires that gate's own RED (the attribution fence, docs/LAB_PROTOCOL.md:4-9).

Scoring rule SR-4 (one cure at a time / attribution). A plateau-break PASS is claimable only if the paired contrast isolates a single between-arm variable (the Phase-2 organ/coupling under test, docs/LAB_PROTOCOL.md:5-9). If a second variable entered the comparison, the result is voided and re-run clean — the metric is sound but the attribution is not, and an unattributable move of this metric is logged exploratory, never as evidence.

Scoring rule SR-5 (PARTIAL is named, never spun). If PRIMARY holds but SECONDARY does not (or vice versa), the verdict is PARTIAL and states exactly which half held — e.g. "PRIMARY PASS (placed_used>0, build chain entered); SECONDARY FAIL (no second non-wood block type mined)." A half-claim is never written as the full claim (docs/LAB_PROTOCOL.md:14-16,46).

6. What this metric does and does not touch (scope fence)

• Does: define the RCON-authoritative scoring read by the collector/probe; supply the objective set, the arithmetic, the gate shape, the anti-hoard proof, and the verdict discipline.
• Does NOT: change phase_goal_met?/2 (mc.ex:223-227), the curriculum @phase_weights, the agent's C/preferences, or any lib/** code. Re-pointing the agent's internal curriculum at a placement goal is the Phase-2 cure (setpoint-peaked C, emptying-B — docs/METABOLISM_GROUND_MODEL_BRIEF.md, docs/DEEPENING_PLAN.md:71-78), reviewed and RED-gated separately. This file is the measuring instrument; it must stay independent of the cure it measures, or it would no longer be an independent yardstick.
• Does NOT deploy anything. Live deploy needs owner go-ahead + the offline gate (docs/LAB_PROTOCOL.md:23-26).

7. Receipts index (where the evidence lives)

• Read mechanism / objective registration / cadence / JSONL: docs/observability/COLLECTOR_RCON_BRIEF.md.
• RCON client surface: lib/sp/minecraft/rcon.ex:39,57,68,78.
• The perverse gate being replaced: lib/sp/brain/mc.ex:223-227 (phase_goal_met?), :230 (inv self-report).
• The P1 PARTIAL that motivates the new metric: docs/DEEPENING_PLAN.md:12-18, docs/UNI_MISSION_DEEPENING.md P1 verdict.
• Owner rulings R1/R2: this session (verbatim in §5); artifact mandate docs/DEEPENING_PLAN.md:26.
• Claim fence: docs/LAB_PROTOCOL.md:35-36, docs/UNI_MISSION_DEEPENING.md:99-103.

No lib/** files were edited; no engine code written; nothing deployed.