Haga · Executive Deck
Independent verification for physical AI
Derived directly from published Lab test fixtures.
Haga · Overview
01 / 14
Independent verification for physical AI
Stress-test. Break. Report.
We stress-test robot policies in simulation and audit generated worlds frame by frame — returning a physics verdict with every failure shown.
Evidence figures from Lab artifacts. Financial figures from plan assumptions.
The problem
02 / 14
Physical AI is funded at scale, but the hard tasks remain unsolved.
- Robotics VC (2025)
- $27.6B
- $27.6B across 1,009 deals in 2025 (PitchBook Q4 2025 preview).
- RLBench sim success
- 89.4%
- EquAct 2D/100 setting (arXiv 2505.21351, 2025).
- BEHAVIOR-1K sim success
- 12.4%
- BEHAVIOR Challenge winner, full-task success on BEHAVIOR-1K in simulation (AI Index 2026, Ch. 2).
Note: Sources: PitchBook Q4 2025 Robotics & Physical AI VC Trends; EquAct (arXiv 2505.21351, 2025); AI Index 2026 Ch. 2 for BEHAVIOR-1K.
ProofCapital is real ($27.6B), difficulty is real (89.4% RLBench 2D/100, 12.4% BEHAVIOR-1K sim success), checking is still done by builders.
TakeawayThe gap between funding and verified capability is the opportunity.
How it works
03 / 14
Your policy goes in. A verdict with evidence comes out.
ProofThree-state gate (FLAG/CLEAN/ABSTAIN) with per-clip measured support plane.
TakeawayThe checker refuses on thin evidence and does not assume the floor.
Proof · policy stress
04 / 14
The harness applies and measures the perturbation correctly. A learned-policy run needs a partner's checkpoint and has not happened.
| Task | nominal | mild | moderate | severe |
|---|---|---|---|---|
| Stack | 0.900 | 0.840 | 0.840 | 0.820 |
| Door (non-monotonic) | 0.800 | 0.760 | 0.660 | 0.760 |
Note: Signal: Stack and Door. Lift and PickPlaceCan are insensitive controls — in the run tables, not on this slide.
ProofScripted OSC_POSE controller, four tasks, four tiers, 50 episodes per cell. Lift and PickPlaceCan hold at 1.000 across every tier. Stack moves 0.900 to 0.820. Door moves 0.800 to 0.760 with the moderate tier at 0.660, which is non-monotonic and is reported as such.
TakeawayWhat this establishes: the ladder applies, the scoring measures it, and the harness reproduces. What it does not establish: policy robustness. That run needs a learned checkpoint, and it runs on a partner's own. The relative gate passes a weak policy at 0.8x its own weak baseline; an absolute floor is not yet defined — the gate tier is the only threshold.
Proof · real footage
05 / 14
On real robot footage the gate flags, and shows its working.
| Dataset | Tracks | Flag rate (clean) | 95% CI | Target | Verdict |
|---|---|---|---|---|---|
| towel | 67 | 0.045 | [0.0153, 0.1236] | 0.020 | FAIL |
| candy | 83 | 0.012 | [0.0021, 0.0651] | 0.020 | FAIL |
| ziploc | 57 | 0.193 | [0.1113, 0.3134] | 0.020 | FAIL |
Note: Run verdict: FAIL under SAP v2's Wilson 95% upper bound ≤ 0.02 rule. Towel FAIL (0.045 [0.0153, 0.1236], upper 0.124), Candy FAIL (0.012 [0.0021, 0.0651], upper 0.065), Ziploc FAIL (0.193 [0.1113, 0.3134], upper 0.313). The contact penetration threshold (0.0080 m) sits below the measured tracking noise floor (0.0160 m); this is the attributed cause of the false positives. Under the frozen rule a PASS at a 0.020 target needs at least 189 clean tracks with zero flags (280 with one); each dataset here has fewer (67, 83, 57), so none could pass.
Prooftowel 3/67 flagged, candy 1/83 flagged, ziploc 11/57 flagged; towel detection 1.000 on 134 injected violations from 67 clips.
TakeawayThe gate flags clean footage on all three datasets under the frozen rule.
Proof · generative worlds
06 / 14
Generative world-model cohort
- cogvideox flagged
- 0.889
- 8 of 9 held-out clips flagged
- cosmos flagged
- 1.000
- 9 of 9 held-out clips flagged
- wan flagged
- 0.889
- 8 of 9 held-out clips flagged
- Real-control specificity
- 0.000
- 15 real controls, 15 flagged
ProofP2-E2 published: 3 models × 9 clips + 15 real controls. Real-control specificity 0.000 (15/15 flagged). Verdict FAIL. Superseded checker (binary FLAG/CLEAN, no ABSTAIN). Re-scoring this cohort under SAP v2 is outstanding; until then its flag rates carry no information (real-control specificity 0.000, 15/15 flagged).
TakeawayThe generative cohort is measured and published. The gate failed on real controls; re-scoring under SAP v2 is outstanding.
How an engagement works
07 / 14
Two weeks. Criterion agreed in advance. One private report.
ProofObjection: unreleased policy is exposure. Answer: criterion agreed first, report private, method public.
TakeawayYou commission one independent check of a claim you framed, before committing hardware.
Pricing
08 / 14
Three ways to buy: an audit, a pilot, a run. Prices indicative; design-partner validation in progress.
| Offer | Scope | Price |
|---|---|---|
| Design-partner pilot | Two weeks: scope and criterion frozen, the run, one private report | Set on the published ladder and agreed on a call |
| Fixed-scope audit | Foundation: 6 clips, 5 days, short report • Standard: 6 clips, 5 days, full report | Foundation $1,900 • Standard $3,500 |
| Per engagement | Policy stress or world-model physics run, criterion frozen before run | Fixed fee, quoted per engagement |
ProofGross margin 78–84% (one-off), 70% (continuous) — plan assumption; variable costs not published.
TakeawayMargin is a plan assumption, not a measured figure.
Market
09 / 14
A thin verification layer under a large physical-AI build-out.
- TAM
- $39M
- Global robotics + world-model QA budget holders: 1,080 orgs (PitchBook robotics cap table + AI Index 2026 lab survey) × $36K ASP (blended audit+pilot+continuous)
- SAM
- $7.7M
- Reachable ICP (sim-first, NA/EU): 430 orgs (filtered by HQ, public cloud spend >$1M, robotics/AI team >5) × $18K ASP (single engagement)
- SOM (yr 3)
- $0.8M
- 5-person capacity at 6-7 engagements/yr × $25K avg (design-partner pilot → standard scope)
ProofTAM: 1080 companies × $36,000 = $39,000,000. SAM: 430 reachable ICP × $18,000 = $7,740,000. SOM: 25–35 design partners × $25,000 = $600,000–900,000. Steady burn $78,000/mo = $936,000/yr; SOM ($600,000–900,000) does not cover burn in year 3.
TakeawayBreak-even on SOM in year 3 is not supported; TAM/SAM represent upside if independent verification scales.
Competition
10 / 14
Robocurve publishes benchmarks. Titus consults. Prismatical is a personal site. Haga referees.
| Player | What they sell | What Haga does that they don't |
|---|---|---|
| Robocurve (robocurve.org, $10M seed, Initialized Capital) | Open-source benchmarks and public result pages: CIs on charts, no verdicts, no price. | Freezes the criterion before the run, refuses on thin evidence, keeps the run private. |
| Titus (titus.io) | Evaluation as a service for industrial operators; pre-release site, no public product. | Ships a product with published artifacts — failures included. |
| Prismatical (prismatical.com) | Solo thesis: sim policy evaluation plus synthetic data; personal site. | Pre-revenue, self-funded, published runs with failures. |
ProofRobocurve $10M seed verified (Initialized Capital, Dealroom). Titus and Prismatical assessed from public sites. Haga: pre-revenue, self-funded, published runs.
TakeawayCriterion-first, private, and willing to refuse to certify.
Roadmap
11 / 14
2 design-partner reports by month 4. World-model depth by month 6.
ProofPlan spends $997,000 of $1,000,000 raised over 14 months. Remaining $3,000 unspent.
TakeawayRound buys the milestones above and stops. Last months need larger close or smaller scope.
Team
12 / 14
One founder, self-funded, every published run carries its failures.
- Applied ML/AI eng
- 7 years
- Fraud detection, multi-agent pipelines, retrieval, evaluation harness to production. Cut inference latency 2.1s→180ms (91.4%); fraud FPR 14%→3.5% at 1.2M tx/day; built 1,200+ case eval harness. BS SE (Hons.), GIFT University 2019.
- Built
- 4 tasks · 3 world models
- P2-E2 real-control specificity 0.000 — gate fails on real footage; generative rates reported, not certified.
- Funded by
- Self-funded
- First commit 2026-07-13 (haga-core). No outside capital.
Note: Planned in North America (Delaware C-Corp default, Canada alternative). Depends on lead investor requirements. No final decision made yet. Incorporation triggered on first committed investment (SAFE execution).
ProofTeam budget 14 months: $503,500. Headcount: 3 (founder + 2 contractors) → 4 (month 4 hire) → 5 (month 9 co-founder).
TakeawayMoney buys named people; headcount rises on published dates.
The ask
13 / 14
Raising $1.0M — burn, runway, and where it goes.
- Avg burn (mo 1–12)
- $70,083
- Mo 1 $53,500 · steady $78,000 (mo 10)
- Runway (months)
- 14
- Recomputed from burn schedule.
- Team funded
- 3 → 5 people
- Founder + 2 contractors · hire mo 4 · co-founder mo 9
- Design-partner pilot
- Priced (indicative)
- Set on published ladder, agreed on call.
Where the money goes
| Bucket | Share | Amount |
|---|---|---|
| Team (Founder + Hire + Contractors) | 50.4% | $503,500 |
| Compute (GPU) | 19.3% | $193,000 |
| Product | 9.4% | $93,800 |
| GTM | 12.1% | $121,300 |
| Legal + Buffer | 8.5% | $85,400 |
What the money buys
| # | Milestone | When |
|---|---|---|
| 1 | 2 live design-partner evaluation reports | ≤4 mo |
| 2 | World-model scoring at depth (n≥30 held out) | ≤6 mo |
| 3 | Continuous-scoring product design | ≤6 mo |
Note: Full 18-mo plan costs $1,309,000; last 4 months need larger close or smaller scope.
ProofPlan spends $997,000 of $1,000,000 raised over 14 months. Remaining $3,000 unspent.
TakeawayThe allocation above is what the plan commits over the funded window; the rest is headroom.
The ask
14 / 14
To earn the right to raise a Series A.
- Instrument
- Post-money SAFE (YC standard)
- No discount, no MFN
- Valuation cap
- $8,000,000
- Standard YC side letter only; offered to lead investor or single check >= $250K
Primary Contact
Mushood Hanif · Founder