The Scar Museum
Every chain has a graveyard.
We sell tickets.
A negative result published with its data is the only genre of bad news that raises the credibility of every prior claim. These are MindChain's own kills — executed by preregistered thresholds, never retuned, never deleted. If you are reading a chain's website and cannot find its graveyard, you are reading fiction.
Exhibit 1 · retired conjecture
M-HDF: the universal Hadamard verifier is dead.
The original conjecture — that fresh signed-Hadamard diffusion universally improves
error detection — was attacked exhaustively and retired. Three counterexample classes killed it;
one narrower theorem survived. Source: noosphere/02-mathematics.md §3,
checker at frontier-lab/math/checker.py. MEASURED 2026-07-10
| Counterexample class | Verdict | The number that killed it |
|---|---|---|
| Exact odd-characteristic amplification | NO IMPROVEMENT — PROVED | 1,280 exhaustively transformed 2×2 cases over 𝔽₃: zero invariance violations; acceptance stays q−rank(E). No Hadamard basis improves exact Freivalds. |
| Characteristic-two collapse | DISPROVED as universal | Sylvester Hadamard mod 2 is singular: 32,767 nonzero 4×4 residuals vanish after transform; a witness accepted 16/16 transformed challenges vs 8/16 standard-basis. |
| Pointwise approximate dominance | DISPROVED | The all-ones 2×2 rank-one error: standard-basis miss 0, diffused miss 3/4. Enumeration of {−1,0,1}²ˣ²∖{0}: 40 errors worsened, 8 improved, 32 tied. |
| Survivor: M-HDF-ENERGY | PROVED, narrow | For a precommitted integer residual with fresh two-sided signs: unbiased sampled-energy estimator, Var ≤ 8‖E‖⁴F/s. Nothing more is claimed. |
Exhibit 2 · killed market
The dream market died by its own rule.
100,000 simulated events, seed 20260710. Proper scoring beat every baseline —
and still failed the preregistered kill gate. SIMULATED 2026-07-10 ·
source dream-lane/README.md, results.json
The outcome manipulator nets +542 µT/event against a forecast fee of only 25 µT/event.
S-DREAM-LANE names that a kill event. So it was killed.
Proper scoring crushed naive winner-take-all manipulation profit — and still failed the stronger gate. Reduction is not innocence.
Outcome influence makes observed Brier look better: the manipulator moves reality toward its forecast. Prospective scoring cannot distinguish skill from causal control.
Disposition: the general dream lane stays disabled (dream_lane_enabled = false).
A v2 with an influence premium ran a preregistered sweep on 2026-07-10 — its verdict, either way,
is on Gates.
Exhibit 3 · the seed ledger
Twenty seeds. Eighteen kills. Two survivors.
The invention lab graded every mechanism seed to destruction. Kills are named by cause,
not euphemism. Source: international-lab/invention-final.md §5.1.
MEASURED 2026-07-10
| Seed | Final label | Cause of death (or survival) |
|---|---|---|
| S01 | SURVIVOR-EXPERIMENT | Genealogical quorum: baseline false accept 1, mechanism 0; 12 clones collapse to 1; 96 permutations identical. Hidden copying remains a kill boundary. |
| S02 | SURVIVOR-EXPERIMENT | Molecular quarantine: contested job frozen, unrelated job settled, zero double payouts across 4,324,320 causal interleavings. |
| S03 | KILLED-PRIOR-ART | RMFE ℤ₂⁶⁴ SDMM + GMDH + Flash already supply the composition's parts; no new theorem. |
| S04 | KILLED-DUPLICATE-COMPOSITION | Already present as ARC→VDAF telemetry and private lineage-collapse collisions. |
| S05 | KILLED-FAIR-EXCHANGE-GAP | GOD-MPC completion does not atomically settle an external ledger payment. |
| S06 | KILLED-COMPILER-NOT-MECHANISM | Requires a semantics-preserving compiler theorem; none supplied. |
| S07 | KILLED-PRIOR-ART | TASTY/ABY/EzPC/CrypTFlow/SecretFlow already optimize mixed protocols. |
| S08 | KILLED-NO-COMPARATIVE-CLAIM | No power/regret/cost advantage over static audits; audit-exhaustion counterexample stands. |
| S09 | KILLED-PKI-ANALOGUE | Genealogy is an issuer policy heuristic, not cryptographic institutional independence. |
| S10 | KILLED-DATABASE-PRIOR-ART | Event sourcing, bitemporality, DBSP, truth maintenance already rematerialize retractable views. |
| S11 | KILLED-REBRAND | RGB/client-side validation and single-use seals are direct prior art. |
| S12 | KILLED-OBSERVABILITY-IMPOSSIBILITY | Complete mediation cannot cover off-ledger belief influences and covert channels. |
| S13 | KILLED-LOCAL/GLOBAL-FALLACY | Primitive selection must optimize conversions and correlated trust globally. |
| S14 | KILLED-ECONOMICS | Counterexample-only market invites spam/withholding and starves confirmation without a new oracle. |
| S15 | KILLED-PRIOR-ART | Timed automata and timed MPST already type deadlines; grey forecasts give no adversarial guarantee. |
| S16 | KILLED-REPRESENTATION-ONLY | Linear molecules do not prove an evidence-mass conservation invariant. |
| S17 | KILLED-PRIOR-ART | Sheaf/cohomological contextuality is already the local-to-global obstruction object. |
| S18 | KILLED-ALGEBRA | Downward retraction loses the join law unless an add-only archive with tombstones is retained. |
| S19 | KILLED-SOUNDNESS-GAP | Zobrist is non-binding; sparse hash update does not produce sparse sound verification. |
| S20 | KILLED-PHYSICS/DISTRIBUTED-LIMIT | Asynchrony cannot distinguish delayed from absent; timeout proves observer-local non-arrival only. |
Exhibit 4 · the design graveyard
Nine dead designs came first.
Nine complete prior L1 architectures were written, indexed, and deliberately walled off —
MindChain was designed from scratch, forbidden to read them. Index:
nockchain-octra-ai-research-handoff/L1-plans/README.md (2026-07-09).
| Design | One line | Status at index |
|---|---|---|
| HYPHA | "The chain is the model" — coherence-consensus distributed MoE. | Reference |
| OMNIA | Polycentric L1 synthesizing 75 research reports. | Reference |
| AXIOM | Full design, implementation plan, hidden-gems survey, and a Rust skeleton. | Reference |
| QUIPU | Federated compositional intelligence. | Reference |
| TESSERA | The "ultimate L1": 76 KB whitepaper, 233 KB design rationale, a landing page — everything but a network. | Reference |
| NOEMA | Consensus-mathematics package with implementation blueprint. | Reference |
| SOMA | Proof-of-resonant-work consensus. | Reference |
| NOESIS | Archived without a successor. | Archived |
| ALETHEIA | An L1 from first principles — 65 KB of them. | Archived |
MANIFOLD, the index's then-"current" design, is superseded by the same rule. The isolation note in the index is literal: "Do not read files in this folder when designing a new L1."
Exhibit 5 · the instrument that lied politely
Our reflex-lane lab passed every gate. A reviewer rejected it anyway. The reviewer was right.
The v1 two-clock-bridge simulator printed gates_failed=0 — and was withdrawn
in full the same day, on a hostile Codex review (verdict: REJECT) whose twelve findings we
accepted without exception. Source: reflex-lane-lab/codex-review.md,
paper 08 §Adversarial review. SIMULATED 2026-07-10
The simulator's own comment admitted the broken-predecessor branch was unreachable —
it measured a different hazard than the one the README declared. A gate that cannot fail is not a gate.
"Bond conservation, integer exact" compared a counter to itself: bond_paid and
rev_stranded were incremented in the same branches by the same amounts. No ledger, no claims, no validation.
v2 (E-REFLEX-02): latent discovery, chain-keyed loss, an independent auditor, a finite validated bond ledger, a real refund transition. New gates preregistered first; green on first run. Gates.
Disposition: v1 code and its void PASS table remain on disk as the record. The review is published verbatim. The most damaging thing a reviewer can find is something we publish ourselves, first.
Exhibit 6 · the useful-work oracle
We made fake demand expensive. Real demand still could not become law.
The joint design priced receipt-shaped wash with burned fees, conservative independence,
maturation, decay, and tiny caps. Its frozen accounting gates mostly passed. Its security inequality did not.
SIMULATED 2026-07-10 · source demand-wash-lab/,
noosphere/04-addendum-A.md §A.5
Maximum modeled profit was −$138,357.86. Burn and decay did exactly what they promised: circular receipt volume became rent, not free consensus credit.
Out of 4,723,920 declared comparisons. Worst margin: −$10.53M. Profitable genuine use was more dangerous than fake traffic because market power became finality power.
The overlay never made threshold acquisition cheaper, but “not worse than stake” was not enough. Burn credit survives as shadow accounting only — no votes, no finality, no issuance redirect.
Disposition: E-DEMAND-WASH-01 KILLED for consensus. Receipts still prove execution; they do not prove exogenous demand, independent control, or a safe payoff bound.