Every result this practice has produced: verified, re-scoped, or retracted — with null tests, job IDs, and full falsification reports. A validation practice that has never falsified its own work is asking you to believe it would falsify yours. This is the proof it does.
What was claimed: A 14-bit toy-curve ECDLP solve using Regev's algorithm on IBM QPUs — one bit below the submission that ultimately won Project Eleven's Q-Day Prize. The submission was privately confirmed within the Prize, and a consolation prize was awarded.
What killed it: A Gidney-style null test — randomized data substituted into the classical post-processing — reproduced the "solve." The pipeline was finding the key; the QPU was not contributing the signal I had attributed to it. The retraction invalidated twenty months of post-processing methodology built on the same approach.
Publication: Retracted publicly with full mechanism disclosure. To the best of our knowledge, the first public falsification of a prize-confirmed ECDLP submission. The autopsy, including the specific test, the failure mechanism, and transferable rules, is in the Quantum Post-Mortem.
What survived: The null-test protocol itself — now mandatory preflight on every Firebringer result.
What was claimed: Hardware runs advancing the public ECDLP record from 6-bit to 12-bit on IBM QPUs.
What the null test showed: These runs survive randomized-input substitution — the correct key is not reproduced by classical post-processing alone, which distinguishes them from the 14-bit result. However, honest scoping is required: the correct key ranked outside the top 10 in the hardware-derived candidate pool in most runs (worst case approximately rank 25).
Current claim: Hardware runs produce candidate enrichment — the correct key rises toward the top of the pool relative to noise. Final key identification requires classical search over that enriched pool. "Solved" is withdrawn. "Enriched" is the defensible claim.
What was claimed: A full HHL linear-solver — not just the QPE stage — executed end-to-end on IBM Torino (Heron R2, 133 qubits). The blocker was an orphan-qubit problem: measuring a qubit the unitary never touched causes an IndexError in the stack. Solution: hang the eigenvalue rotation on a data qubit already inside every controlled-U, and pin the transpiler to only the qubits the problem needs.
Scope, stated plainly: 5 logical qubits, gate depth 210, 1,024 shots, 2-bit clock precision — a toy instance. Clock-register phases 1·2·3 dominate, matching the non-zero eigenvalues of the spectral operator built from elliptic curve 11a1. This demonstrates executability under measurement hygiene — not computational utility. Limits remain: depth near decoherence, coarse precision, hardware noise throughout.
Self-audit status: This result is currently being run through the five-category scorecard as Scorecard 002 — the same instrument used for market claims. The status entry above will be updated (green or amber or red) when the self-audit concludes. No result is awarded a green Verified badge before the instrument has run.
What was claimed: Register-specific reorganization of QPU output distributions around synchronized meditation windows, appearing to scale with participant count, with apparent negative controls from queue-delayed sessions.
What killed it — five independent artifacts:
Session 9-9: Two backends (IBM Torino, IBM Brisbane) merged into one timeline, sorted by submission time. The "organization event" was the backend mixture ratio changing, not anything on either chip. The "ordered" state celebrated was near-maximal entropy — the labels had been inverted from day one.
Decoder bug: A 2026 re-extraction decoded zlib-compressed payloads as raw bitstrings. Impossible shot counts (8,726–8,761 on a fixed 8,192 setting) proved the pipeline was describing its own compression, not the hardware. The correlation that "confirmed" the effect was correlations between noise series.
Session 5-25: Classified as a "failed synchronization" negative control but executed in a 2-minute burst — no time axis existed. The strongest evidence class was uninformative by construction.
Statistics: Independence-assuming t-tests on autocorrelated QPU drift series. At lag-1 autocorrelation r=0.65, these tests fire false positives 40–56% of the time. Under the valid test — exact circular-rotation permutation with FDR correction across 36 endpoints — zero significant results.
Metric naming: "Entanglement" was entropy divided by bit count — identical to entropy to machine precision on all 296 rows. The metric that appeared to show independent corroboration was the same quantity on a rescaled axis, counted as separate evidence for two years.
What survived: The 15-rule NISQ falsification checklist — one rule per artifact, published in the Post-Mortem.
The autopsy of both retracted experiments, artifact by artifact, with the transferable tell for each one. Timeline integrity (5 rules), pipeline integrity (3 rules), design integrity (3 rules), statistical integrity (4 rules), epistemic integrity (3 rules). Each rule names the specific artifact it kills. None require a PhD. All of them would have prevented a specific failure in this project.
Published as a standing resource for NISQ experimenters, quantum POC teams, and anyone evaluating hardware-backed quantum claims.
The most common claim in the post-quantum security market graded publicly, including the ten questions to ask before accepting it and the structural reason why a live quantum-resistance demo is a negative signal. Verdict: reject the phrase, fund the migration — cryptographic agility is the property that can actually be purchased.
Why publish retractions? Because a validation practice that has never falsified its own results is asking you to trust an instrument that hasn't been tested on anything it built. These entries are permanent — corrections are addenda with dates, not replacements. The instrument that kills a market claim first had to survive killing ours.