Dhan Astra is an Indian index-options intelligence engineered the way a Formula 1 team builds a car. Every part is designed, measured, raced on days it has never seen, and bolted on only if it survives. Nothing is believed. Everything is proven, or it is refused.
In the Itihāsas, an astra is not a sword in the hand. It is a force invoked by mantra, given only to one who has earned it through discipline, and it answers only to true knowledge.
Karṇa learned the Brahmāstra from Paraśurāma under a false identity. When the deception was found, the curse was precise: the knowledge would desert him at the moment he needed it most. It did.
That is overfitting, described in the Mahābhārata long before statistics had a name. A model that has seen its own answers performs in every rehearsal and fails on the battlefield. Dhan Astra is built so that no knowledge enters it by deception.
Our first hard rule: never tune toward a return target. Measure, and report what was measured.
No invented constant, no synthetic row, no plausible default. A value is measured and its source cited,
or the engine returns Refused with the reason. A zero that flows into a P&L is a lie
that looks like arithmetic.
सत्यमेव जयते “Truth alone triumphs.” · Muṇḍaka Upaniṣad 3.1.6
Every result is split by time, purged and embargoed, charged every statutory cost, and raced against a shuffled-label control and its own majority-class baseline. A number fitted on its own test is not a result.
योगः कर्मसु कौशलम् “Yoga is skill in action.” · Bhagavad Gītā 2.50
When the system falters it stops reaching for new risk, but it always lets open inventory out. Blocking an exit strands a position and turns a bounded loss into an unbounded one.
समत्वं योग उच्यते “Evenness of mind is called yoga.” · Bhagavad Gītā 2.48
Not the look of F1. Its discipline. Every Dhan Astra component exists because a part of a racing car taught us what it must do, how it must be tested, and when it must be thrown away.
Every surface of a Formula 1 car is simulated, tunnel-tested and track-tested. It stays on the car only if the stopwatch agrees.
Every Dhan Astra component has a job, a benchmark that could prove it wrong, and a place in the parts bin with its measured reason if it fails. Scroll to take the car apart.
Pressures, temperatures and loads stream from the car to the pit wall on every lap, through every corner.
Every tick enters through one decoder that mirrors the Upstox V3 protobuf schema field for field: touch, five-level depth, chain, OI, IV and greeks, each stamped with its source and venue time. Missing depth is reported dark, never filled with a zero.
A hybrid power unit harvests, stores and deploys energy, and every joule of it is metered so it is spent at exactly the right moment.
Raw prints become features without look-ahead: microprice, touch imbalance, Lee-Ready trade signs, VPIN and depth imbalance at levels 1, 2, 3 and 5. Each estimator is O(1) per observation, allocates nothing on the hot path, and refuses below its 385-sample floor.
Computed<T>: Measured or RefusedThe driver's every input runs through a single electronic control unit. When a subsystem reports a fault, the wheel knows, and the strategy changes.
21 specialists across 16 capability layers: Greeks, volatility, liquidity, order flow, microstructure, regime, expiry, risk, data quality and more. Each must pass its own falsification benchmark when it is registered. One that fails is quarantined, and a faulted taskforce closes the decision gate.
Air pressed onto the car keeps it planted when it is pushed hardest. Without it, speed is only a faster way to leave the track.
Size is clamped down to whole lots and never rounded up. Orders are sliced under the exchange's quantity freeze. One losing round trip ends new entries for the day. The book stays planted when the market turns.
Everything the car does reaches the track through four contact patches. Grip, wear and temperature are where theory meets asphalt.
Everything a strategy does reaches the market through its fills. Market orders walk the visible depth and the rest is cancelled. Resting limits fill only when the opposite touch reaches them or a print trades through. Every statutory charge and brokerage is deducted before expectancy is judged.
The rear-wing flap could open only in a designated zone, only within one second of the car ahead, and only when race control enabled it.
Capital is unlocked the same way. A strategy earns authority only when all ten gates pass together: data integrity, causality, purged walk-forward, held-out stability, shuffled-label control, inverse controls, execution realism, regime coverage, paper confirmation and risk approval.
The halo protects the driver when all else has failed. Under a red flag the race stops, but every car is still brought back to the pit lane.
A Dhan Astra halt is a red flag: it stops new entries at once and never stops an exit. The cell is the architecture itself. There is no broker order path in the crate, a source scan fails the build if one appears, and every fill is simulated and labelled paper.
Then everything that was measured is tested at once, on a track the car has never seen.
Enter the race weekendA part is not fast because its designer believes it is. It is fast when it wins on a track it has never seen, under the same rules as everyone else. Every Dhan Astra candidate races the same weekend, in the same order, with no way to skip a session.
A candidate is assembled from explicit primitives, entry, exit, sizing and context, and stamped with an immutable version, parameter hash, feature list and lineage.
Fitting happens only on the training window. The test window stays sealed, and look-ahead is refused at the type level.
Expanding chronological folds with an embargo. Every fold must clear its cost-adjusted bar and beat its own majority-class baseline.
Once qualifying starts nobody touches the setup. The shuffled-label, inverse and no-signal controls race the frozen car.
₹10,000, one position, every charge, and the first losing round trip ends entries for the day, on days the strategy never saw.
All ten gates and twenty counted sessions, or the part is archived with the measured reason it failed.
A yantra is read from the outside in. Its bhupura admits you only through four gates, and every enclosure refines you until you reach the bindu. ALTRON is built the same way. Evidence enters through a gate, is refined plane by plane, and converges on a single decision, or on a refusal.
expiry_clockPerceptionoi_build_directionOptions chaintheta_dragGreeksgamma_theta_balanceGreeksvolatility_premiumVolatilityexpiry_itm_probabilityProbabilitybook_pressureLiquidityexecutable_liquidityLiquidityorder_flow_imbalanceOrder flowtouch_imbalanceMicrostructuremicroprice_displacementMicrostructuremid_momentumPrice structurenet_edge_after_feesQuantitativechain_oi_skewRegimestrike_gravityExpiryspread_costExecutionsingle_tick_ruin_riskRiskobservation_freshnessData qualityiceberg_detectionOrder flow · darkspoof_detectionOrder flow · darkevent_proximityEvent · darkDark means registered and visible but unable to measure: iceberg and spoof detection need per-order book detail no Upstox tier provides, and event proximity needs a licensed calendar. A dark capability is reported, never faked.
Not illustrations. Each is shown with the exact line of Rust that computes it and the file it lives in. Each returns a measured value or a refusal that carries its reason.
Fair value inside the spread. Each price is weighted by the opposite side's size, so a heavy offer pulls fair value down toward the bid. A crossed or one-sided touch refuses.
(bid * ask_qty as f64 + ask * bid_qty as f64) / totalOrder-flow pressure at the touch. ±1 are real extremes. Both sides empty is an absent book, not a balanced one, so it refuses instead of returning zero.
if total == 0 { return Computed::refuse(..) }W and L are the ledger multipliers after every charge. Below p* full-capital compounding shrinks the ledger however many cycles run. At ₹0.20, one tick each way, p* is 56.79%.
let p = -ln_l / denom;Below 385 observations a worst-case proportion's 95% Wilson half-width exceeds five points, which is too vague to size on. Every estimator below its floor refuses.
pub const MIN_FLOW_SAMPLES: usize = MIN_CALIBRATION_SAMPLES;Volume-synchronised probability of informed trading over the last n complete buckets. A partial bucket is an unfinished measurement, so it refuses until the window is full.
Computed::Measured(sum / self.window as f64)Who crossed the spread. An exact-midpoint print is broken by the price path. A zero tick is Indeterminate and adds nothing to any sum, rather than leaning it in a direction.
let mid = 0.5 * (bid + ask);T is buy plus sell turnover. GST is charged on services only, never on STT or stamp duty, because those are taxes. Rates are dated: these apply from 1 April 2026.
let gst = (exchange_txn + sebi + ipft + brokerage) * self.gst_on_services;Passing every day of a finite sample does not prove a perfect daily pass rate. The verdict carries the one-sided 95% Clopper-Pearson lower bound for n of n.
pub fn assess( .. ) // every counted day, or nothingEvery candidate below raced on held-out days, under real costs, against its controls. None has
taken the chequered flag yet, and the engine says so: its default state is still DISARMED. A result
nobody could attack was never a result.
| Pos | Candidate | Held-out | Bar | Result |
|---|---|---|---|---|
| — | Queue-imbalance predictor71,666 labelled samples; slope positive in every fold | 54.98% | 56.79% | DNFbelow breakeven |
| — | Passive maker, real fill model10–25 band, ~91% of resting bids filled | 49.28% | 53.50% | DNF−4.21 pts, the closest gap measured |
| — | Depth feature sweepsix causal features; best: ask-touch concentration | 66.39% | 78.67% baseline | DSQlost to the majority class |
| — | Queue priority sweepback of queue to a front no retail account can have | 9.80–33.53% | 53–63% | DNFadverse selection at every level |
| — | 16 pre-commitment conditionspassive entry filters | 2 inverted | held-out edge | DSQpositive in-sample, inverted out |
| — | Oscillation veto (0.97 → 0.95 → 0.97)in-sample z −3.50, held-out z +6.69 | inverted | held-out edge | DSQthe sign flipped out of sample |
| — | DeepLOB-style network2 conv layers + GRU, walk-forward, label-shuffle placebo | 0 of 9 folds | 55–89% per slice | DNFit read real structure, and still lost money |
| — | Open-time selector, 975 dayspassed L0, then attacked | 1.444 payback | no-model control 1.449 | DSQa control did as well |
| — | 4,250-node consensus meshpinned by its own tests | always BUY | a real decision | DSQa constant function, removed |
| — | Daily +500% gate, shipped makerone loss ends the day; every order billed | 0 of 6 | +500% net | DNFmeasured, never tuned toward |
Three candidates looked brilliant in-sample and inverted the moment they met unseen days. That is Karṇa's curse in a spreadsheet: knowledge gained by seeing the answers deserts you when it matters. The gates exist so that it deserts us in research, not with capital.
The live path has no shortcuts and no side doors. Every stage can refuse, and a refusal travels forward with its reason instead of becoming a zero.
The credential is read from Windows Credential Manager, the venue clock is synced, and contract masters load for NIFTY, BANKNIFTY, FINNIFTY and MIDCPNIFTY, each contract with its own lot and freeze. Any failure stops the start, and says why.
One decoder, one book per contract, every update stamped with venue time. A stale quote is flagged, never traded on.
Market orders walk the visible depth and cancel the rest. Resting limits fill only when the opposite touch reaches them or a print trades through.
Kill switch, daily loss limit, stale feed or stale quote: entries stop at once. An exit that cannot be priced waits for the first fresh book, then fills.
The ledger is sealed. Every record says simulated: true, and the schema cannot say otherwise.
Replay, folds, controls and attacks. Settled questions are written down so nobody relitigates them, and open ones are argued in the room.
Five instruments run the engine's own arithmetic in your browser, transcribed line for line from the Rust and checked against the engine's pinned results: the 56.79% breakeven, the 53.45% minimum, 49,985 units and ₹1,368.80 of brokerage all reproduce exactly. When the engine would refuse, so do they.
Set a two-sided touch. Cross it, or empty a side, and watch the engine refuse.
The engine's own round trip: whole lots of 65, slices under the 1,755 freeze, every charge, log compounding.
Statutory rates in force from 1 April 2026, itemised exactly as the engine books them.
The derivation behind the engine's 385 floor. Tighten the margin and watch the requirement climb.
A simulated tape, for illustration only: no market data. Each print is signed by Lee-Ready and fed to the engine's VPIN.