Manthan Intelligence · Live
The pod is awake.
What you're looking at is a real-time view of an autonomous venture fund operating system. 46 agents, 8 divisions, 43 scheduled firings a day, zero manual triggers. The clock below advances as you watch.
Backtest figures shown here are the strictest subset — 298 blind-backtest verdicts where the system was given no outcome data at scoring time. The broader rolling calibration corpus is larger and includes earlier non-blind verdicts; methodologies differ by subset.
The system grades itself
Last 15 of 298 backtest verdicts on private companies. Each verdict was generated from publicly available information only and scored against the real outcome — none of these are deals the Fund is currently evaluating. Average accuracy: 63%. Wins in green, misses in red — no cherry-picking.
| Company | Sector | Verdict | Outcome | Score |
|---|---|---|---|---|
| Scapia | Fintech | Lean Aligned | Raised up | 0.70 |
| Paper Boat | D2C | Lean Decline | Raised flat | 0.60 |
| NoPaperForms | Edtech | Lean Aligned | Raised up | 0.70 |
| Censys | Cybersecurity | Lean Aligned | Raised up | 0.70 |
| 15Five | HR Tech | Lean Aligned | Stalled | 0.15 |
| Unbox Robotics | Robotics | Lean Aligned | Raised up | 0.70 |
| Strapi | SaaS | Lean Aligned | Raised up | 0.70 |
| Portal Space Systems | Space Tech | Lean Aligned | Raised up | 0.70 |
| NeoGrowth | Fintech | Lean Decline | Raised flat | 0.60 |
| Bankjoy | Fintech | Lean Decline | Raised up | 0.15 |
| Illumio | Cybersecurity | Lean Aligned | Raised up | 0.70 |
| Blue Frog Gaming | Gaming | Lean Decline | Stalled | 0.70 |
| Caption Health | Healthcare | Lean Decline | Raised up | 0.15 |
| Arkeus | DefenceTech | Lean Aligned | Raised up | 0.70 |
| HyperPlay | Gaming | Lean Decline | Raised up | 0.15 |
What the system has learned from breaking
Recent entries from the failure log. Each one is a real bug, missed verdict, or process regression — captured so the next iteration doesn't repeat it.
-
Weekly backtest fabrication near miss
Scheduled tasks that synthesize output must precondition-gate on source data existence before any synthesis step. Without a gate, a task with an incorrect source path will fabricate data rather than fail loudly.
-
Entity Quality Regression: Null company_name and Default Mental Health Sector Contamination
4 entities had null company_name field: freedom-forever, candid_health, yoco, meltplan. Add non-null company_name validation gate to scrape-to-KG write path
-
NR-9.1 Amendment: Add Chapter 11 and Related Distress Stages to Pool Integrity Exclusion List
The NR-9 pool integrity filter's stage exclusion list does not include Chapter 11 or other bankruptcy/distress stages. Freedom Forever passed the filter despite being in Chapter 11 bankruptcy.
-
Website deployment cascade failure
ManthanBot website launch session (11 April 2026) produced a cascade of 5 distinct failures across CTAs, progress bar, and results page rendering.
-
Death diagnosis graphcore
Four calibration rules from Graphcore blind vs outcome comparison: CR-1 revenue can decline (not just plateau) — check for ecosystem-dependent companies; CR-2 platform shift signals (like GPT-3 for Graphcore) should carry -0.10 confidence penalty; CR-3 exceptional team prevents…
-
Aira sprint framework effective
Aira sprint framework (request → research → findings → impact → next steps → closure) is effective when runtime exists. [sprint] (SEBI disclosure) produced a clear Option B decision and was properly closed with Tvashtar follow-up.
Showing 6 of 22 publicly-shareable entries · 79 total in the learning graph
Estimated, based on a Sonnet-class run mix. Actual spend varies with prompt caching, model routing (Haiku for sweeps, Sonnet for analysis, Opus for synthesis), and agent verbosity. Real telemetry replaces these once the system moves fully to metered server infrastructure.
What you're looking at
This is not a mock-up. The clock above is your local browser's clock, mapped onto a 24-hour face. Each dot is a scheduled task running inside an actual venture fund. As you read this paragraph, an agent somewhere on that ring has just fired, or is about to.
The fund is two humans (partner + analyst) and forty-six agents. Eight divisions — research, deal flow, analytical deliberation, financial modelling, engineering, product, GTM, and consulting — each operate as small panels of specialised agents that disagree, deliberate, and produce structured outputs. Threads of work pass between them. The output lands on the partner's desk; the next loop starts.
The numbers above are live counts from our knowledge graph and backtest pipeline. The accuracy figure is the system grading its own analytical history — 298 deals scored against actual outcomes under strict blind methodology (the system sees no outcome data at scoring time). It improves week by week. The learning entries are the failures the system has chosen to remember.
The 50-second compression below is the same day, rendered as art. The browser version above is the same day, rendered as time.
24 hours, compressed. Same fund. Same day. Open full Day-Loop page →