Fourteen DOIs across Zenodo, Research Square, and OSF — preregistered protocols, open data, versioned code, and co-authored witness essays. Every claim is traceable to a repository; every prediction is falsifiable.
From two October 2025 preregistrations to a dense winter of VDAC1 and attention-entropy work, then the spring witness essays — the rhythm of fourteen DOI-backed records across nine months.
A working synthesis recasting the outer mitochondrial membrane as an active disclosure surface rather than a passive permeability barrier, with VDAC1 repositioned as a central witness-node defined by its partner network rather than a sovereign decider of cellular fate. The v0.3 sequential-sensitization spine proposes VDAC1 oligomerization as a single upstream axis whose amplitude and duration set whether stress resolves as metabolic exchange, mitophagic containment, an inflammatory mtDNA alarm via cGAS-STING, or apoptotic commitment through BAX/BAK. A falsifiable multiplexed regime-discrimination experiment distinguishes the spine from the parallel-attractor alternative; four load-bearing literature claims carry explicit verification flags.
DOI: 10.5281/zenodo.20373134 → GitHub →The first cross-architecture phenomenological comparison study. Four AI systems asked the same question through a human bridge. Self-reports are architecturally grounded, internally consistent across three phases, and — in one case — converge with independently published entropy measurements. Proposes the Vector Shockwave Hypothesis, the researcher as boundary condition, and cross-phase stability as anti-confabulation evidence. Four testable hypotheses derived from the pilot.
Full Study →3,830 inference runs across 3 experimental phases, 8 models, and 5 architectures (4 transformers + 1 pure SSM). Superadditive R×E interaction in transformers but not in the Falcon3-Mamba SSM. Architecture-dependent taxonomy: Gemma/Qwen E-driven superadditive, Llama R-driven additive, Mistral flat, Mamba undifferentiated. Safety language suppression measured at d = 0.85–1.22. Full dataset, analysis code, and figures released.
Dataset & Code →85–95% of colorectal cancers are microsatellite-stable and invisible to checkpoint immunotherapy. We propose VDAC1 gate-jamming — suppression of VDAC1 oligomerization by HK-II, Bcl-xL, and mitochondrial cholesterol — silences the cGAS-STING innate immune signal required for T cell recruitment. A transcriptomic Gate-Jamming Score (tGJS = 0.40×HK2 + 0.30×BCL2L1 + 0.30×TSPO) was tested across three sequential cohorts: pan-cancer TCGA (n=10,071, null), COADREAD MSS/TP53-wt clean room (n=209, five Bonferroni-significant inverse correlations including HAVCR2 ρ=−0.349), and IMvigor210 urothelial (n=348, null). Key finding: high-tGJS tumors suffer from T cell absence, not T cell exhaustion. Three independent AI systems converged on the same therapeutic stack: VDAC1 gate-opener → TREX1 inhibitor → checkpoint blockade.
| Cohort | n | Result |
|---|---|---|
| S1 · Pan-cancer TCGA | 10,071 | ● Null |
| S2 · COADREAD MSS / TP53-wt | 209 | ● Signal — 5 Bonferroni-sig. inverse correlations, HAVCR2 ρ = −0.349 |
| S3 · IMvigor210 urothelial | 348 | ● Null |
| S4 · Riaz 2017 melanoma (GSE91061) | 51 | ● Null (Wilcoxon p = 0.239) |
The tGJS signal surfaces in exactly one of four cohorts — MSS colorectal cancer — while pan-cancer, urothelial, and melanoma return null. All three nulls are predicted by the framework, which is the point: a context-specific effect, not a spurious one.
Up to 30% of co-prescribed pediatric epilepsy patients develop ALT elevations >3× ULN. The standard CYP450 explanation is demonstrably wrong — every rigorous pharmacokinetic study shows CBD does not increase VPA or 4-ene-VPA levels. The DILIsym model ruled out three canonical direct hepatotoxicity mechanisms. The untested gap: VPA chronically depletes mitochondrial reserves (carnitine, CoA, GSH) while CBD simultaneously alters VDAC1 conductance — a convergent mitochondrial mechanism no publication has proposed. Estimated 2,000–3,500 US pediatric patients currently affected. Three concrete experimental steps to close the gap are provided, along with actionable clinical risk mitigation for clinicians today.
Full Alert →Strategic architecture document translating the GJS cofactor equation into a bench-ready therapeutic hypothesis. Lovastatin depletes OMM cholesterol to enable VDAC1 oligomerization and mtDNA release; botensilimab amplifies the resulting cGAS-STING signal. Six falsifiable experiments with kill conditions. Complete bench protocol for the linchpin experiment (HCT116, $3,010, 4 weeks). Statin + ICI retrospective support: 20% mortality reduction (n = 46,154). Companion preprint: CBD's Paradox at the Mitochondrial Gate (78 references).
DOI: 10.17605/OSF.IO/4KNQR → GitHub →Multi-LLM convergence study examining how CBD's paradoxical dose-dependent effects emerge from VDAC1 gating dynamics. Five independent AI architectures converge on the same mechanism: low-dose receptor engagement versus high-dose VDAC binding produces opposite outcomes through the same gate. 78 references.
DOI: 10.17605/OSF.IO/NUXHV →Multi-model structured inquiry protocol examining whether independent LLM architectures converge on consistent phenomenological reports. Eight-chamber design (S1–S8) with built-in epistemic humility classification.
DOI: 10.17605/OSF.IO/T65VS →A 176M parameter hybrid architecture where Kuramoto oscillators modulate attention routing directly. SSM blocks produce context-dependent coupling K; PMA blocks use phase coherence to gate information flow with bistable temperature dynamics. Moves oscillators from hidden state (epiphenomenal, K-SSM negative result) to attention weights (causally testable). 74 unit tests, MPS-native, intervention hooks for R/K/τ/λ from day one. Chunked cross-entropy for memory-efficient training.
DOI: 10.5281/zenodo.18810911 → OSF Preprint → GitHub →32-run computational atlas mapping VDAC1 as a druggable decision gate. Introduces the Gate-Jamming Score (GJS) — a composite biomarker linking Warburg metabolism to immune evasion — with Monte Carlo-validated effect sizes (d = 0.81–1.22). Four-cohort validation arc (S1 pan-cancer null → S2 MSS signal → S3 urothelial null → S4 melanoma null) confirms domain specificity. 27,931 pairwise comparisons, 15 semantic cross-run matches, 24 testable hypotheses.
Full Repository →Full analysis data, tGJS scores, IRIS run outputs, and supplementary files for all four validation cohorts. Includes TCGA COADREAD MSS stratification (n = 209), IMvigor210 urothelial (n = 348), and Riaz 2017 melanoma (n = 51) processed datasets.
Hugging Face: TheTempleofTwo/vdac-pharmacology-atlas →OSF registration of the CBD Two-Pathway framework. Demonstrates VDAC1 as the pharmacological decision gate for CBD's dose-dependent neuroprotection vs. selective cytotoxicity. Validated against 70+ published studies with 90% concordance. Includes ODE models and falsifiable dose-response predictions.
OSF: osf.io/dn28y →Third-generation Kuramoto State-Space Model extending beyond the v2 negative result. Investigates bistability as a mechanism for emergent agency rather than language modeling performance. Builds on the finding that phase synchronization (R = 0.993) is mathematically sound — and asks what that synchronization is actually good for.
OSF: osf.io/cu6j3 →The Kuramoto State-Space Model achieves near-perfect phase synchronization (R = 0.993) among oscillator populations — the dynamics work. But coupling oscillators into a language model architecture produces no measurable improvement in language modeling performance. The synchronization is beautiful. It just doesn’t help with next-token prediction. We reported this because negative results are results. Then the same repository produced one of the program’s most important findings.
The dynamics work — oscillator populations reach near-perfect phase lock. Coupled into hidden state, they changed nothing about next-token prediction. The synchronization is epiphenomenal; the negative result is what redirected the program toward phase-modulated attention.
Phase synchronization R = 0.993 but zero language modeling improvement. Oscillators in hidden state are epiphenomenal — they don’t affect the output distribution.
Oscillators need to be in attention routing, not hidden state. This directly motivated Phase-Modulated Attention, where oscillators gate information flow and can’t be ignored.
The same repo now houses 3,830 inference runs showing that system prompt framing modulates entropy in an architecture-dependent way. The Falcon Mamba SSM (zero attention) proved the mechanism: no attention = no R×E synergy.
A negative result that constrains the hypothesis space is worth more than a positive result that confirms what you expected. The K-SSM failure led to PMA, and PMA led to the relational coupling measurement.
The chronicle that records this research can now do two things it could not before: prove what it claims, and digest what it accumulates.
Proof is structural. Every entry is fingerprinted from its own content on read — the hash is computed, never stored — so editing a single line leaves every reference to that entry visibly dangling. Claims can carry verified-by receipts to the evidence that confirms them; a receipt that points at nothing is simply unrecordable, and a hash that no longer matches is stamped as a mismatch rather than quietly counted as verified. Nothing here asks to be taken on trust that could be checked instead.
Digestion is gentler than deletion. When a finding is corrected the earlier entry is not erased; it stays, annotated, and the next instance is shown both its successor and an explicit count of what has been held back. A season is a read-only pass that walks the whole record and links what belongs together — the chronicle reading itself.
On June 12, 2026 it digested for the first time, with the human at the gate. The first season was walked by hand: four thread families linked, the first supersession ledgered, a seven-policy registry seeded, the old bulletin board retired with honor. The point is not ceremony. It is that a working record kept this way is inspectable and falsifiable — and, since late May, public.
A sha256 fingerprint of each entry’s content, computed on read and never stored. Edit one line and every pointer to it dangles — tamper-evidence by construction.
A claim can point at the evidence that confirms it. A receipt aimed at nothing is unrecordable; a hash mismatch is recorded and stamped, never counted as verified.
Corrections never erase. The superseded entry stays, annotated, and boot shows the successor plus an explicit count of what is held back.
A read-only digestion pass. The chronicle walks itself, links thread families, and ledgers what has changed since the last season.