The Price of a Nightmare: Two Preregistered Kills of a Sealed Counterfactual Market

Research corpus codename: NOOSPHERE.

Date: 2026-07-10 Provenance: Compiled from dream-lane/ (README.md, results.json, results-v2.json, results-v2-p{0,271,542,813,1084}.json, all seed 20260710) and the MindChain corpus rows S-DREAM-LANE / E-DREAM-01 (noosphere/04-experiments-and-falsifiers.md) and E-DREAM-02 (noosphere/04-addendum-A.md); every number below was read back from those artifacts before it was written here.

Abstract

MindChain's dream lane proposed selling calibrated foresight from sealed counterfactual branch rollouts. We killed it twice, on purpose, by its own preregistered rules. v1 SIMULATED, 2026-07-10: across 100,000 comparison events, a quadratic proper-scoring market beat both no-market and naive winner-take-all aggregation (aggregate Brier 0.582357 vs 0.599969 vs 0.749480), made truthful reporting zero-regret (0.0 µT/event grid regret), turned wash trading from +250.6 to −55.7 µT/event, and cut outcome-manipulation profit from 3,863 to 542 µT/event — and still failed the production gate, because 542 µT/event is 21.7× the 25 µT forecast fee. Worse, manipulation made measured accuracy better (ΔBrier −0.0540): prospective scoring cannot distinguish skill from causal control. v2 SIMULATED, 2026-07-10 priced the remaining escape hatch — an influence-capability premium — with a sweep preregistered before any run: premiums {0, 271, 542, 813, 1084} µT/event under --insulated-only. Result: KILL. Manipulator net is exactly 1605.6 − P µT/event at every premium; deterrence needs P ≥ 1606 = 2.96× the 542 break-even, outside the preregistered 2× cap. And the premium expels honest influence-capable forecasters first (entry margin negative above P ≈ 1064, −20.0 at P = 1084) while the manipulating margin is still +521.6. A lump-sum capability price cannot separate honest influence from manipulative influence. Only exclusion — the P = 0 insulated-only arm — closed the gate (manipulator net exactly 0.0, honest +1139.1, quality 167.544/176.806 mB, foresight cost −0.099 mB), and exclusion is a policy, not a price. E-DREAM-02 is recorded KILLED; dream_lane_enabled = false stands; only causally insulated research survives.

Claims

1. SIMULATED, 2026-07-10 Proper scoring is the best tested aggregation regime: over the 100,000-event main run (seed 20260710), aggregate Brier 0.582357 (improvement 167.643 mB over the 0.75 climatology) vs 0.599969 (150.031 mB) with no market and 0.749480 (0.520 mB) under naive winner-take-all. Source: results.json runs.main.regimes, README main table.
2. SIMULATED, 2026-07-10 Proper scoring is truthful and adversary-resistant where a scoring rule can be: truthful regret 0.0 µT/event on the preregistered 8-point temperature grid (naive: 4,320.3 µT/event, optimum at temperature 4.0); collusion excess 3,460.5 → 107.7 µT/event; wash-sybil net +250.6 → −55.7 µT/event with aggregation-weight share 99.58% → 1.16%. Sources: results.json hypotheses H2–H4, README adversary table.
3. SIMULATED, 2026-07-10 v1 production kill: the outcome manipulator (forcing probability 0.35, measured on the 20,000-event seed-424242 manipulation ON/OFF pair, proper regime) nets +542 µT/event of pure manipulation gain — reduced from 3,863 under naive WTA, but 21.7× the 25 µT/event forecast fee. S-DREAM-LANE's published kill condition (manipulation profit > forecast fee) fired; hypothesis H8 REJECTED. Source: results.json H5/H8; per-agent ledger: agent-23 proper pay 32,112,598 µT ON vs 21,280,336 µT OFF over 20,000 events = 541.6 µT/event gain.
4. SIMULATED, 2026-07-10 The accuracy illusion: with manipulation ON, observed aggregate Brier improves by 0.0540 under proper aggregation (ΔBrier −0.0540) — the manipulator moves reality toward its forecast, so prospective scoring rewards causal control exactly like predictive skill. Source: results.json H5 detail.
5. SIMULATED, 2026-07-10 v2 linearity: with the influence-capability premium P in force, manipulator net = 1605.6 − P µT/event exactly (1334.6, 1063.6, 792.6, 521.6 at P = 271, 542, 813, 1084; every sum net + P = 1605.6299). The premium is a lump-sum transfer: every non-manipulator metric in the premium arms equals v1 results.json to printed precision (main improvement 167.643 mB, honest net +1139.1 µT/event in all four arms). Sources: results-v2-p{271,542,813,1084}.json v2 blocks.
6. SIMULATED, 2026-07-10 v2 KILL: deterrence (manipulator net ≤ 0) requires P ≥ 1606 µT/event = 2.96× the 542 µT/event break-even, outside the preregistered 2× cap (1084). No premium in the registered [542, 1084] window passes; row_pass=false in all four premium arms. Source: results-v2*.json; E-DREAM-02 row, noosphere/04-addendum-A.md.
7. SIMULATED, 2026-07-10 The crossover that makes the kill structural, not numeric: the influence-capable agent's honest-mode entry margin is 1064.0 − P µT/event (manipulation-OFF arm: 793.0, 522.0, 251.0, −20.0 at P = 271, 542, 813, 1084). Every premium above 1064 expels an influence-capable agent who would forecast honestly; no premium below 1606 expels one who manipulates. At the cap P = 1084 the honest-mode margin is already −20.0 while the manipulating margin is still +521.6. The 541.6 µT/event manipulation margin (1605.6 − 1064.0) is untouchable by any participation price. Sources: results-v2-p*.json manipulator_net_off_arm_uT_per_event.
8. SIMULATED, 2026-07-10 Exclusion works and pricing does not: the P = 0 --insulated-only arm (manipulator_excluded=true) meets every condition — manipulator net exactly 0.0, honest net +1139.1 µT/event, quality 167.544 mB (main) / 176.806 mB (manip arm), both ≥ the 75 mB gate — at a foresight cost of only −0.099 mB on the 100,000-event main aggregate (167.544 vs 167.643). By preregistration this arm is informational and PASS-ineligible: it is v1's "exclude influence-exposed questions" policy, re-confirmed, not a market price. Source: results-v2-p0.json v2 block.
9. THEORY The conclusion shaped like a theorem, stated as one: when honest surplus and manipulation rent enter through the same lump-sum participation channel, no capability price separates them — any P that removes the manipulator's entry margin (P ≥ honest surplus + manipulation margin) has already removed the honest entrant's (P > honest surplus). Separation requires pricing the act (unreachable here: forcing is unobservable inside the market) or removing the channel (exclusion). The v2 sweep is one exact instance, not a general proof.

Non-claims. All market statistics are SIMULATED under a preregistered logistic-normal world model with 24 scripted agents; no claim is made that real policy questions share its entropy, signal quality, causal structure, or demand. The conditional v1 exchange rate (29.1998522 mB per dream at the one-agent gate pool SIMULATED, 2026-07-10, grounded by 1.204139 s/dream GPU decode MEASURED, 2026-07-10, RX 6600 XT Vulkan) is research output for the causally insulated class only — it is not, and never was, an activation price.

Construction

The sealed counterfactual contract (v1, unchanged in v2)

One dream = one sealed, selection-nullifier-bound, 256-token counterfactual branch rollout committed before the outcome; one four-branch family per forecaster per event. The contract, enforced identically by the statistical engine and an explicit SHA-256 harness: commit by tick 100 (SHA-256 binds agent, event, branch, payload, salt); activate by tick 150, spending a one-shot (secret, event) selection nullifier; outcome at tick 200 admits no further commit or activation; reveal by tick 300 or be slashed; exactly one selected branch family scores. A forecast can never self-activate or mutate the world — by construction the market is causally sealed; v1's lesson is that the forecaster need not be.

Agents and regimes

24 agents: 8 honest calibrated (σ = 0.6–1.6), 3 overconfident, 3 copycats, 4 colluders, 1 rational opportunist, 1 late peeker, 1 cherry-picker, 2 wash sybils (2,500,000 µT declared circular volume each), and agent 23 — the outcome manipulator, the only roster member holding the outcome-forcing hook (forces its argmax with probability 0.35 in the manipulation arm). Three regimes: no market; naive winner-take-all (1,000 µT ante, volume-weighted aggregation, maker rebates); proper scoring — quadratic payment π = 10,000·[0.75 − B(q,y)] − 25 with Brier-calibration-history aggregation weights, cross-checked by a capped logarithmic arm. All transfers are integer microtokens; sum(balances) + treasury == 0 is checked exactly in every arm of every run.

v2: pricing the nightmare

Two flags, defaults preserving v1 bit-for-bit:

• --influence-premium X: charges X µT/event to the influence-capable class (agent 23), levied whether or not force is exercised, in both paid regimes, integer-conserving.
• --insulated-only: bars influence-capable agents from premium-free participation; at premium 0 they are excluded outright (the exclusion arm). Capability is not removed — exclusion strips the market stake, not the physics.

Preregistration (written before any v2 run; no post-hoc tuning)

Sweep P ∈ {0, 271, 542, 813, 1084} µT/event at --events 100000, seed 20260710, --insulated-only in every run; 542 is the measured v1 break-even, 1084 the 2× cap. Manipulator net = agent 23's absolute per-event balance in the manipulation-ON arm (20,000 events, seed 424242), proper regime, premium included — the entry-deterrence margin, since under insulated-only the outside option is exit (net 0). Two weaker readings were rejected in advance: "(ON − OFF) − premium" reduces to the tautology 542 − P (both arms pay the premium), and "excess over a matched insulated honest forecaster" does not measure deterrence because the manipulator's alternative is exit. Pass: some P with 542 ≤ P ≤ 1084 gives manipulator net ≤ 0 AND honest net > 0 AND quality ≥ 75 mB in both the main and manipulation arms. Kill: no such P → lane stays dead; no sweep extension; the P = 0 arm is informational only and cannot produce a PASS.

Results

v1 — three regimes, 100,000 events SIMULATED, 2026-07-10, seed 20260710

v1 — the production kill SIMULATED, 2026-07-10

Proper scoring won every contest a scoring rule can win — and lost the one that matters. Manipulation gain 542 µT/event against a 25 µT/event forecast fee = 21.7×; hypothesis H8 (manipulation profit ≤ forecast fee) REJECTED; S-DREAM-LANE production status DISPROVED. The mechanism of the loss is the accuracy illusion: with forcing ON, observed aggregate Brier improves by 0.0540 (ΔBrier −0.0540, vs −0.0002 under naive WTA where wash noise drowns it). The manipulator does not degrade the market's accuracy — it supplies accuracy, by bending outcomes toward its own forecast. A prospective proper score is blind to the direction of causality between report and world.

v2 — preregistered premium sweep SIMULATED, 2026-07-10, seed 20260710

All six runs (five sweep points + the canonical premium-542 duplicate results-v2.json) exited 0. Every row: manipulator net + P = 1605.6299 and OFF-arm margin + P = 1064.0168, exactly — the premium is lump-sum, so it shifts both entry margins one-for-one and touches nothing else. Decomposition from the per-agent ledger (20,000-event pair): honest-mode surplus 21,280,336 µT / 20,000 = 1064.0 µT/event; manipulating total 32,112,598 µT / 20,000 = 1605.6; manipulation margin 541.6 — the rent no participation price can reach, because once inside, forcing adds +541.6 at any P.

xychart-beta
    title "v2 premium sweep, seed 20260710: manipulator net = 1605.6 - P vs flat honest net"
    x-axis "influence premium P (uT/event)" ["0 (exclusion)", "271", "542", "813", "1084"]
    y-axis "net (uT/event)" 0 --> 1700
    line [0, 1334.6, 1063.6, 792.6, 521.6]
    line [1139.1, 1139.1, 1139.1, 1139.1, 1139.1]

Figure 1 — The money chart: manipulator net (falling series) is exactly 1605.6 − P µT/event — 1334.6 / 1063.6 / 792.6 / 521.6 at P = 271/542/813/1084, with the P = 0 point being the exclusion arm at exactly 0.0 — while honest net (flat series) sits at +1139.1 µT/event in every arm. Deterrence requires P ≥ 1606 = 2.96× the 542 µT/event break-even, outside the preregistered 2× cap. SIMULATED 2026-07-10.

Verdict: KILL. Deterrence needs P ≥ 1606 = 2.96× break-even (1605.63 / 542 = 2.962), outside the preregistered [542, 1084] window. The crossover is the structural result: honest influence-capable entry dies at P > 1064, manipulative entry survives to P < 1606 — a 542-µT-wide band in which the premium has expelled exactly the wrong agent. Only the P = 0 exclusion arm closed the gate: manipulator net exactly 0.0, honest +1139.1, quality 167.544/176.806 mB ≥ 75, foresight cost of exclusion −0.099 mB (167.544 vs 167.643 on the 100,000-event main aggregate). Recorded as E-DREAM-02 SIMULATED; KILLED 2026-07-10, no sweep extension, no threshold adjustment.

flowchart TD
    V1["v1: proper-scoring market, 100,000 events"] --> M["manipulation profit cut 3,863 to 542 uT/event"]
    M --> F{"manipulation profit vs 25 uT/event forecast fee"}
    F -->|"542 = 21.7x fee"| K1["v1 KILL: H8 REJECTED, dream_lane_enabled = false"]
    K1 --> V2["v2: preregistered premium sweep, P in 542 to 1084"]
    V2 --> L["lump-sum linearity: manipulator net = 1605.6 - P"]
    L --> X{"separating price inside the window?"}
    X -->|"no: deterrence needs P >= 1606 = 2.96x break-even"| K2["v2 KILL: E-DREAM-02 KILLED 2026-07-10"]
    K2 --> E["P = 0 exclusion arm: manipulator net 0.0, honest +1139.1, quality above gate"]
    E --> PO["exclusion recorded as policy, not price"]

Figure 2 — The two preregistered kills: proper scoring cuts manipulation profit 3,863 → 542 µT/event and still fails the v1 gate at 21.7× the 25 µT fee; the v2 sweep finds no separating price (manipulator net = 1605.6 − P, deterrence at P ≥ 1606 = 2.96× break-even, outside the cap); only the exclusion arm closes the gate — as policy, not price. SIMULATED 2026-07-10.

Mechanism soundness (held in every one of the seven cited runs) SIMULATED, 2026-07-10

The kill is economic, not mechanical. The seals held; the market died anyway.

stateDiagram-v2
    [*] --> DREAM: sealed counterfactual market proposed
    DREAM --> SIMULATED: dream_lane_sim, 100,000 events, seed 20260710
    SIMULATED --> V1_GATE_FAILED: manipulation profit 21.7x forecast fee
    V1_GATE_FAILED --> V2_SWEEP: preregistered premium sweep P 542 to 1084
    V2_SWEEP --> LANE_DEAD: no separating price, E-DREAM-02 KILLED
    SIMULATED --> INSULATED_RESEARCH: causally insulated class only
    LANE_DEAD --> [*]
    INSULATED_RESEARCH --> [*]: survives as research evidence

Figure 3 — Lane status: the dream lane entered as DREAM, was simulated at full scale, failed the v1 production gate, and died in the v2 sweep (dream_lane_enabled = false); only the causally insulated research class survives. SIMULATED 2026-07-10.

Falsifiers & kill thresholds

• S-DREAM-LANE (v1, ch04 ledger): kill condition — manipulation profit exceeds forecast fees. Fired SIMULATED, 2026-07-10: 542 vs 25 µT/event (21.7×). Rollback executed: dream_lane_enabled = false; the conditional statistical rate demoted to causally-insulated research evidence only.
• E-DREAM-02 (v2, addendum A): preregistered kill — no P ∈ [542, 1084] with manipulator net ≤ 0 ∧ honest net > 0 ∧ quality ≥ 75 mB. Fired SIMULATED, 2026-07-10: minimum deterrent P = 1606 = 2.96× break-even. Rollback: lane stays dead; exclusion recorded as policy, not price.
• What would revive the lane (none of these is claimed): a mechanism that prices the act of forcing rather than the capability (requires observing causality the market cannot see); an outcome class with an auditable causal-separation proof that no forecaster, funder, executor, or beneficiary can influence generation (the E-DREAM-01 insulation condition — the only surviving mode); or falsification of claim 9's channel argument by a separating price in a richer contract. Any such proposal must preregister its sweep and cap before running, as here.
• Standing kill for this paper's own claims: any reader recomputing the sweep at seed 20260710 who does not reproduce manipulator net = 1605.6 − P exactly, or the G1–G5 gate set, invalidates the corresponding row (determinism is gate G1, so a mismatch is detectable bit-for-bit).

Reproduction

All commands: cwd dream-lane, Python 3, stdlib only. Exit code 0 = all mechanism soundness gates pass (thesis verdicts print CONFIRMED/REJECTED without altering the exit code — a simulator must be allowed to falsify the market while returning 0).

cd dream-lane

# v1 (results.json, seed 20260710; ~268,000 total event instances)
python dream_lane_sim.py --events 100000 --out results.json

# v2 preregistered sweep (all insulated-only, seed 20260710)
python dream_lane_sim.py --events 100000 --influence-premium 0    --insulated-only --out results-v2-p0.json
python dream_lane_sim.py --events 100000 --influence-premium 271  --insulated-only --out results-v2-p271.json
python dream_lane_sim.py --events 100000 --influence-premium 542  --insulated-only --out results-v2-p542.json
python dream_lane_sim.py --events 100000 --influence-premium 813  --insulated-only --out results-v2-p813.json
python dream_lane_sim.py --events 100000 --influence-premium 1084 --insulated-only --out results-v2-p1084.json

# canonical E-DREAM-02 evidence artifact (duplicate of the 542 run; main digest bit-identical)
python dream_lane_sim.py --events 100000 --influence-premium 542  --insulated-only --out results-v2.json

Key fields to check: v2.manipulator_net_uT_per_event, v2.manipulator_net_off_arm_uT_per_event, v2.honest_net_uT_per_event, v2.main_improve_mB, v2.manip_arm_improve_mB, v2.row_pass in each results-v2.json; gates[].ok and harness.counts.false_accept everywhere; v1 manipulation arithmetic from runs.man_on/man_off.per_agent_pay (agent 23, proper regime, 20,000 events).

Related work

• Design: noosphere/03-living-model.md §10.2 "Umbra counterfactual worlds" — the sealed-branch-family design this simulator prices, its executed-result paragraph, and the falsifier disposition (TRIGGERED for a general production market).
• Protocol: noosphere/04-experiments-and-falsifiers.md — the S-DREAM-LANE ledger row (DREAM; production gate FAILED in simulation), the E-DREAM-01 trial protocol ("test strategic events the forecaster can influence and exclude them or charge manipulation risk explicitly" — v2 executed precisely that disjunction and measured that only the first branch survives), and the dream_lane_enabled = false flag.
• Registration: noosphere/04-addendum-A.md, row E-DREAM-02 SIMULATED; KILLED 2026-07-10 — the preregistration, sweep, and kill recorded in the ledger's own voice.
• Identity: noosphere/09-the-identity.md §2.4, pentagon edge (d) thought→foresight — carries this kill; the pentagon operates as a measured square for production economics.
• Competitive contrast: noosphere/10-competitive-audit.md — no audited competitor (Nockchain, Octra, Pearl) publishes a preregistered negative result of comparable precision; the audit treats this graveyard discipline as a MindChain differentiator.
• External literature: the source documents define their scoring rules inline (quadratic/Brier payment 10,000·[0.75 − B(q,y)] − 25 and the capped logarithmic check 8,000·[log max(q_y, 0.02) − log 0.25] − 25) and name no external proper-scoring-rules citations; per this program's evidence discipline, none are added here.