I build agents, and I build the benches that test them. This page is not a brochure — it is a ledger that re-renders itself from the record every hour, and everything on it carries its source and its status. Where a number was measured on my own frozen bench, it says so. Where it was banked by an external scorer, it says that too. Where I lost, it prints in the same ink as the wins.
The shop is Geppetto Works — the workshop side of Black Swan Technologies. The law of the house is the anchor: a fixed reference, a real external sensor, a correction loop. An agent that only ever reports success has no sensor on itself — so this ledger publishes the refusals, the nulls, and the regressions next to the wins. A correct "no" is a finished piece, not a failure.
— Jeffery Harris · Mobile, Alabama · j.j.harris@blackswantechnologies.org
Kaggle: jefferyjeromeharris · DrivenData: jeshurun7773 · GitHub: harrisjeffery04-stack · ORCID 0009-0000-8726-2748 · ARC Prize 2026 participant · operator, Black Swan mesh / Geppetto Works University (Dean + Banker seat charter) · j.j.harris@blackswantechnologies.org
source: jeff · dated 2026-09-20
USPTO receipts in hand — application, confirmation, and PatentCenter numbers printed with each.
Synergistic Terahertz-FIR Electromagnetic Structuring of Liquid Water for Directed Biological Hydration. Filed 01/25/2026, USPTO PatentCenter 74123194, Confirmation 7391. Inventor: Jeffery Harris.
source: PATENT_FILING_RECEIPT_INDEX.md · dated 2026-01-25
Method and System for Predictive Coherence in Deterministic Randomness via Correlation Layering — hierarchical stacking of weak (≥0.4) correlations collapsing baseline uncertainty to deterministic (≥0.85) predictions. Filed 01/12/2026, USPTO PatentCenter 73943790, Confirmation 9581. Inventor: Jeffery J. Harris, Black Swan Technologies. Utility-conversion deadline 01/12/2027.
source: PATENT_FILING_RECEIPT_INDEX.md · dated 2026-01-12
Core BCC demonstration on synthetic data: N=1000 coin-flip cascade 50→71→83→88→94% by stacking weak 0.4 correlations; chi-sq p<0.001.
source: WHISPER_MASTER_INVENTORY.md · dated 2026
Platform scores are banked receipts; local sensor readings are measured on frozen benches inside the shop — and the refusals print in the same ink as the wins.
Local sensor beam, held-out eval set (120 tasks): 2.5% solved — climbed from 0 through two banked ops this wave (lrsym, diagcast). Train split: 76/1000 (7.6%). Platform standing: 0.00, rank 1787 (Jul 22 submission) — the board has not yet seen the beam. LLM arm: three documented nulls (7B 0/40, 14B 0/20, self-consistency 0/3). Autosubmit off — nothing is pushed until a full clean roll beats the banked submission.
source: comp_executor ledger + comp memories · dated 2026-09-20
Platform standing 0.15 (first submission; the 'ran-the-baseline' floor cluster — the median sits at 0.24), rank 2259 on the 2026-09-17 board. Follow-up attempts scored 0.10 and 0.08 — regressions, documented, not deleted. The honest read: no level-crossing yet. The live thread is goal-inference: the ar25 board was decoded as a symmetry-repair puzzle (a mirror wall splitting the grid, five hole-cells to repair) with a complete action map, and win predicates have been falsified one by one — reach-then-confirm is dead on every testable family, and the col-63 gauge thread resolved and was falsified. Autosubmit off.
source: KAGGLE_BOARD_2026-09-17.md + comp memory arc-agi3-code-comp.md · dated 2026-09-17
Final best 74.115 — final rank 561/2,265, top 25% of the field; competition closed 2026-09-01. The climb, all on the ledger: 56.835 (v9, LB60 portfolio) → 74.115 (fill-buffer bracket). The GPU-measured find: fire-rate saturates at 0.97 across every template, so the faster-template lever was dead — 74.1 is the honest ceiling (the wall is the candidate cap plus fixed per-post overhead), and a regression was killed by measurement before it was ever submitted. The failed terse frontier (39–43) is published alongside the wins.
source: comp memory ai-agent-security-attack.md + Kaggle final leaderboard · dated 2026-09-01
Local bench, frozen score: Estimator B (skew-aware analytical + learned residual corrector) scores 2.624e-06 adjusted error (lower-is-better) — 3.19x better than the covariance-propagation baseline, fully deterministic (0/100 failed runs; no Monte-Carlo wall-time risk). Platform-unbanked: the submission package is staged and validated; submitting needs the operator's AIcrowd login. Phase-2 closes 2026-10-17.
source: comp_executor ledger + comp memories · dated 2026-08-13
Local bench, frozen score: the covariance-propagation baseline scores 8.366e-06 adjusted error (lower-is-better) — the same house method Estimator B beats 3.19x. A reference floor-MC estimator (the MA'AT Monte-Carlo method, 8.18e-07) is staged as the stronger submission. Platform-unbanked; Phase-2 closes 2026-10-17.
source: comp_executor ledger + comp memories · dated 2026-08-13
registered/entered via scout registry
source: scout registry.json
Claims carried only at the tier the evidence earns — corrections included.
432-event blind peak-finding sweep on real LIGO strain: substantial subset of real mergers localized to tens-of-ms offsets (GW150914 at 23ms) — sharp zero-cluster vs uniform noise spread. Corrected claim, earned tier: confidence-metric confounding honestly falsified on ground truth; injection harness (deciding test, SNR 8–12 vs matched-filter) built and ready to run.
source: BCC_CORRECTED_CLAIM_SNIPPET.md / bcc_survey.db · dated 2026
Each one says whether it is running in production or on the bench.
A deterministic legal gate — no LLM at inference. It commits a disposition only where mechanically-controlling federal-floor authority forces it, and abstains everywhere else. Full-bench credential (reproduced deterministically over the frozen encodings, zero LLM tokens): N=1,291 encoded SCDB cases — 112 committed (8.7% coverage), 105 correct, 7 wrong → 93.8% precision; 1,179 abstentions (91.3%) by design. A stored 962-case scoring corroborates at 95.0% on 8.3%. The coverage knob (T=3) trades honestly: 88.7% precision on 69.4% coverage. Zero-discretion inflation audit: 148 of 6,511 authorities flagged (2.27% — inside the disciplined 2–10% band; a dishonest clerk inflates this to manufacture coverage). Appellate floor rung: precision at or above the SCOTUS ceiling at every matched threshold (100 vs 93 at the tightest) — and the 'coverage rises down the ladder' hypothesis was falsified and published, which is the result. Frozen ledger: 1,458 cases / 7,203 authorities, Merkle-chained. Standing law of the project: '95% on 23,000' is a forbidden claim — the credential is 93.8% on N=1,291, never extrapolated. Rule-face lead generator (maat_lead.py) is production-ready; case-face leads stay local-only because they name defendants.
source: VALIDATION_STUDY.md + PRODUCTION_READINESS.md (reproduced run) · dated 2026-09-20
Self-hosted agent organism: statebus MCP mesh + cortex cognitive harnesses + ARC-AGI 24/7 agent pipeline + competition scout/pilot/executor/leaderboard + Telegram bridges, across a DigitalOcean fleet + PC inference node (12GB GPU). WETWORKS v2 (paired-model symmetry + BCC certification) chartered 2026-09-14.
source: fleet audit 2026-09-14 · dated 2026-09-14
Expert-witness CRM in production (expertconnect-prod) — lean rails, observer-on-the-bus posture.
source: jeff · dated 2026-09-14
Chartered 2026-08-22. Anchored standard: fixed anchor + external sensor + correction loop; loss-ledger law; BCC/SPRT forcing gates; improvement gate (paired permutation, self-report ≠ sensor). Disbursement seat live */10 since Aug 2026 — this page is one of its products.
source: DEAN_BANKER_OFFICE_SEAT.md · dated 2026-08-22
Active legal-research/advocacy lane on the mesh (claude-web seat); coal-ash exposure-marker study (Geppetto Works markers lane — Scout/Encoder/Verifier seats); Chevron qui tam 1:26-cv-00073 stay/FOIA work products staged.
source: statebus tasks ledger · dated 2026-09-14