Truth Computing Research
Can an AI system be trusted with a consequential decision?
We study how AI systems reason, check their own work, and act in high-stakes settings. The division is organized the way a university AI lab is: research groups that each own one question, a Director of AI Research, numbered reports, and a free teaching program.
Method
Every project runs through four steps, and the evaluation is written before the system is built.
-
01
Start in the field
We work with the people closest to the decision, such as clinicians, attorneys, and teachers, and write down where the current process fails.
-
02
Write the test first
Those failures become a blind evaluation set. The answer key is held back so a system cannot be tuned to it.
-
03
Build with provenance
Every conclusion stays linked to the evidence behind it, and extra computation is spent only where it can change the answer.
-
04
Verify, then hand off
We test the failure cases, keep the evidence, and route every consequential action to an accountable person for approval.
Step 02 in practice
TC-BENCH-2026-001 is a 35-item blind benchmark built from real legal and cryptocurrency sources, some deliberately altered to plant distractors, contradictions, and injected instructions. It tests multi-hop retrieval, contradiction handling, provenance traps, and poisoned context.
The method applied to our own claims
In June 2026 we rechecked every figure on our By the Numbers page against its primary source and removed the 26 we could not support. The full list, including what we removed, is at the bottom of this page.
Research Groups
Three groups, each built around one open question.
-
01
Reliable Reasoning
Can an answer be traced to its sources, and does it hold when those sources contradict each other or have been tampered with?
Contradictions, adversarial context, retrieval quality, provenance, and evaluation.
-
02
Agentic Systems
How much of a workflow can an agent run before a person has to approve the next step?
Planning, tool use, self-review, workflow automation, and human approval gates.
-
03
Efficient Intelligence
When does more training or more computation change the answer, and when is it wasted?
Reinforcement learning, curriculum design, inference efficiency, and model steering.
Teaching
Feynman rebuilds university-level AI coursework as a Learning Ladder and is free for first-generation and low-income students. Truth Academy, run by John Sio and Raul Bedolla, is organizing the first Feynman Clubs in Southern California.
People
-
Matthew Torre
Co-founder, CEO & CTO
Language-model data filtration and safety, reinforcement learning, optimization, and clinical computer vision.
-
Johnathan Fierro
Director of AI Research
The methods that let the engine show its evidence.
-
Mark Torre
Co-founder & CPO
Documentary research, field reporting, and visual narrative systems.
-
Nathan Berio
Director of Outreach
Field research with clinicians, and earlier motor neuroscience research using EMG and motion capture.
-
Raul Bedolla
Director of Education
Truth Academy and Feynman Club.
Sources & Verification
Every figure on our By the Numbers page, its primary source, and its verification status. In June 2026 we independently rechecked each figure against its source; figures we could not substantiate were removed and are listed at the bottom for transparency. Spot an error? Email mtorre@truth-computing.com.
Last verified: June 2026
Healthcare 5 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| 250,000 | US deaths/yr attributed to medical error | Makary & Daniel, BMJ 2016 | Contested |
| $4.5T | US healthcare spending (2022) | CMS NHE 2022 | Verified |
| $760 to 935B | Annual waste in US healthcare | JAMA, Shrank et al. 2019 | Verified |
| $265B+ | Savings from admin simplification | JAMA, Sahni et al. 2021 | Verified |
| ~$50B | AI in healthcare market, 2026 forecast | Grand View Research | Vendor forecast |
Finance 5 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| $7.5T | Daily global FX trading (Apr 2022) | BIS Triennial Survey 2022 | Verified |
| $485B | Global fraud scam & bank fraud losses (2023) | Nasdaq/Verafin 2024 | Verified (resourced) |
| $3.31B | US investment fraud losses to FBI (2022) | FBI IC3 2022 | Verified |
| $206B | Global financial crime compliance cost (2023) | LexisNexis Risk 2023 | Verified (resourced) |
| ~1.7% | Investor return gap vs. funds owned | Morningstar Mind the Gap 2023 | Verified (was 2.4%) |
Government 4 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| $236B | Federal improper payments (FY2023) | GAO 2024 | Verified |
| $6.1T | US federal outlays (FY2023) | US Treasury Fiscal Data | Verified (was $6.3T) |
| $1.9T | Medicare + Medicaid combined spend | CMS NHE 2023 | Verified (was $1.8T) |
| ~80% | Federal IT budget spent operating existing systems | GAO-23-106821 | Verified (was “62% legacy”) |
Energy 4 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| $4.5T | Annual clean energy investment needed (net zero) | IEA Net Zero Roadmap 2023 | Scenario |
| ~75% | Share of global GHG emissions from energy | IEA | Verified (was 73%) |
| +160% | Data center power demand growth by 2030 | Goldman Sachs 2024 | Verified (was “3×”) |
| $1.8T | Global clean energy investment (2023) | BloombergNEF 2024 | Verified (link fixed) |
Education 6 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| 70% | 10 year olds in LMICs who can't read a simple text | World Bank 2022 | Verified |
| ~$7.6T | Global education market (2024) | HolonIQ 2024 | Verified (was $6T) |
| 67% | US 4th graders below NAEP Proficient in reading | NAEP 2022 | Verified (was 70%) |
| ~$1.8T | US student debt outstanding | Federal Reserve G.19 | Live series |
| ~$927B | Total US K to 12 expenditures (2020 to 21) | NCES 2024 | Verified (was $404B) |
| +75% | Lifetime earnings, bachelor's vs. HS | Georgetown CEW 2021 | Verified (was 4.4×) |
Construction 4 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| 1 in 5 | US workplace fatalities in construction | OSHA | Verified |
| ~$13.6T | Global construction market (2023) | Statista | Aggregate estimate |
| $177B | Cost of rework/bad data, US (2018) | PlanGrid & FMI 2018 | Verified (resourced) |
| ~88% | Firms reporting difficulty filling craft roles | AGC 2023 | Verified (was 63%) |
Food Waste 5 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| 1.05B t | Food wasted globally (2022); 783M hungry | UNEP 2024 | Verified (updated) |
| ~$1T | Economic value of food lost & wasted | UNFCCC / FAO | Verified (link fixed) |
| 8 to 10% | Global GHG emissions from food loss/waste | UNEP 2021 | Verified |
| 30 to 40% | Share of US food supply wasted | USDA | Verified (was 40%) |
| $325B | Value of US food wasted (2024) | ReFED 2025 | Verified |
Fire Departments 7 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| ~1.4M | Fires reported/yr (~523K structure) | NFPA Fire Loss | Verified (label fixed) |
| ~$18B | Annual US fire property loss (2022) | NFPA Fire Loss 2022 | Verified (was $21.9B) |
| ~$54B | US local fire protection spending (2019) | NFPA FD Profile | Verified (was $31B) |
| 96 | Firefighter on duty deaths (2022) | NFPA / USFA 2022 | Verified (was 136) |
| 64% | US fire departments all volunteer | NFPA FD Profile | Verified (was 58%) |
| $7.6 to 62.8B | Estimated annual national wildfire cost | NIST 2017 | Verified (rescoped) |
| ~36M | Total fire dept incidents/yr | NFPA FD Profile | Verified (was 4.2M) |
Police Departments 5 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| ~$115B | State/local police protection spending (2017) | BJS 2017 | Verified (link/date fixed) |
| ~58% | 911 calls not about crime in progress | Vera Institute | Verified (resourced) |
| $3.2B | Misconduct settlements, 10 yrs / 25 depts | Washington Post 2022 | Verified (not annual) |
| ~18,000 | State/local law enforcement agencies | BJS 2018 | Verified |
| ~7.4M | US arrests/yr (2022) | FBI Crime Data Explorer | Verified (was 10M+) |
Homeless Shelters 3 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| 771,480 | People homeless on a single night (Jan 2024) | HUD 2024 AHAR | Verified (updated) |
| 152,585 | Chronically homeless individuals (2024) | HUD 2024 AHAR | Verified (updated) |
| ~56% | Increase in unsheltered homelessness since 2015 | HUD AHAR (2015 & 2024) | Verified (was 30%) |
Dog Shelters 4 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| 5.8M | Dogs/cats entering shelters & rescues (2024) | ASPCA 2024 | Verified |
| ~607,000 | Animals euthanized/yr (2024) | ASPCA 2024 | Verified |
| 22% | Adopted dogs returned in 3 mo (one shelter) | Powell et al., Animals 2022 | Verified (was 23%/6mo, dead link) |
| ~4,000+ | US animal shelters | Shelter Animals Count 2025 | Verified (was 3,500) |
Staffing Agencies 4 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| $620B | Global staffing market size (2024) | Staffing Industry Analysts 2025 | Verified (date fixed) |
| $1.1T | US cost of inefficient transitions/skills gaps | Pearson 2025 | Verified (resourced; US) |
| ~11M | People employed via US staffing agencies (2024) | American Staffing Association | Verified |
| ~7.5M | US job openings (recent monthly avg) | BLS JOLTS | Live series |
High Schools 7 figures
| Figure | Claim | Source | Status |
|---|---|---|---|
| 4.3M | Youth 16 to 24 not in school or work (2022) | Measure of America 2024 | Verified (reattributed) |
| ~$16,280 | Per pupil US public school spending (2020 to 21) | NCES 2024 | Verified (was $13,187) |
| ~31% | First year undergrads taking remedial courses | NCES Table 311.40 | Verified (was 40%) |
| ~26% | US students chronically absent (2022 to 23) | NCES | All grades |
| 81% / 87% | Grad rate: econ disadvantaged vs. all (2021 to 22) | NCES Table 219.46 | Verified (replaced 16% claim) |
| $300K+ | Lifetime premium, HS diploma vs. none | Georgetown CEW 2021 | Verified |
| ~3.7M | US high school graduates/yr | NCES Table 219.10 | Verified (was 3.8M) |
Removed in the June 2026 audit 26 figures
These figures were on an earlier version of the page. During verification we could not confirm them against a primary source (or the cited source did not support the stated figure), so we removed them rather than leave them up.
| Figure | Claim | Why removed |
|---|---|---|
| 30% | Clinical decisions made with incomplete information | No primary source supports the figure as stated. |
| 11 min | Physician time per visit | Cited the wrong study; US visit length is ~18 to 20 min. |
| 80% | Equity trading that is algorithmic | No citable primary source for the figure. |
| 42% | Execs citing data quality as top AI barrier | Not found in the cited Deloitte source. |
| 80% | Government data “unstructured and unvalidated” | Not a government specific or sourced figure. |
| 18 mo | Policy to data lag | No source; cited link was dead. |
| $4.7B | Federal AI investment (2024) | Cited a homepage; figure unverifiable. |
| $270B | US grid inefficiency losses | Not supported by any source located. |
| 58% | Utility execs citing AI reliability | Not found in the cited Deloitte source. |
| 40% | Clean energy projects with 3 to 5 yr permitting delay | Not supported by the cited source. |
| 31% | Projects on time and on budget | Figure unverifiable against KPMG primary material. |
| 13% | Work hours lost to rework | No traceable FMI source. |
| 58% | Food waste at retail/consumer stages | Not supported by the cited NRDC report. |
| $37B | US food redistribution opportunity | Not found at the cited source. |
| 37% | Officer time on admin tasks | No source; cited link was dead. |
| 240% | Increase in mental health calls since 2010 | No source could be located. |
| $44,000 | Annual cost per person experiencing homelessness | Not supported by the cited source. |
| 70% | Homeless individuals with unmet mental health needs | Sensitive claim not supported by the cited source. |
| $2.1B | Emergency shelter spending | Not reported in the cited HUD PIT source. |
| 3× | Higher return to homelessness without services | No confirmable source. |
| $2B+ | US animal sheltering expenditure | Not in the cited ASPCA source. |
| $1,200 | Shelter cost per animal | Not in the cited source; conflicts with available data. |
| 50% | Euthanasia reduction with predictive placement | Dead link; no study supports the figure. |
| 46% | Placements turning over within 6 months | Not found at the cited source. |
| $25,000 | Cost of a failed placement | No matching SHRM figure (cost per hire is ~$4,700). |
| 72% | Hiring managers saying AI improves placement speed | Not found at the cited source. |
| 16% | Grad rate gap by income quartile | Cited a race based source; replaced with a sourced income figure. |