How Auction Edge reads the market.

You can only describe an auction if you have the auction data. Ours, in detail:

• NIFTY 1-minute price history — confirm exact bar count, date range, and whether NIFTY-I / spot / both. We currently say "5+ years."
• 100K+ historical rotations — derived from a fractal rotation engine that segments the price stream into directional moves and balances
• Options-chain history — IV smile snapshots, GEX, intraday flow. how far back, which instruments
• Tick data via ClickHouse — high-resolution storage for the high-frequency engines and back-testing. live-streamed or end-of-day reconstructed?

We are deliberate about scope. Currently the focus is NIFTY 50 (and the front-month NIFTY future, where applicable). Bank Nifty, Sensex, and single-stock work are on the roadmap — not yet shipped, not pretended.

The Narrator is an AI market analyst, currently powered by Anthropic's Claude. It runs after every session boundary — Post-IB, Lunch, EOD, and the Weekly Plan — and produces:

• A structured brief in the dashboard
• A long-form post on the headless WordPress blog at /blog
• A version cross-published to Substack

The Narrator reads from the regime state JSON, the structural-levels store, the options state, and the rotation library. It does not free-form a market view; it composes a brief from quantitative state. The output is the kind of brief a senior trader would write if they had time to write one every day — observational, structural, and committed to a "line in the sand" rather than a hedged non-answer.

The Narrator is not a signal generator. It does not say "buy" or "sell." It describes what the auction has done, what the structural setup is, and what the scenarios look like.

The forthcoming AI Trading Coach uses the same engines the Narrator uses, but applied retrospectively to your trades. It reconstructs the market state at the timestamp of every entry and exit, cross-references the Narrator's contemporaneous read, and rates the trade on multiple structural axes.

Full details: /coach.

We will not pretend the engineering is trivial. A few of the harder problems we have solved:

• Real-time profile recalculation at sub-second latency, including value area / POC, on a moving 1-minute stream — without UI stutter
• Regime classification with stable boundaries — naive bucketing flickers; we use trailing SD bands with hysteresis to avoid regime "flashing" between two adjacent buckets on a single tick
• Cache invalidation on settings change — time-based caches (e.g. 60s for daily OHLC) had to be reset at the top of recalc loops, otherwise lookback changes silently failed. We learned this the hard way.
• Single-print tracking across sessions — naive detection re-fires previously tested zones; persistence requires session-aware status tracking with intraday updates
• Narrator factuality — keeping AI commentary grounded in the underlying state JSON, with a structured prompt that surfaces all relevant numbers explicitly so Claude does not need to "remember" them
• Cross-engine consistency — the regime label, the structural levels, and the options state must all reflect the same moment in time. We use a single refresh tick that fans out to every engine, then a write barrier before the Narrator reads.

None of this is novel computer science. It is just careful engineering applied to a domain where shortcuts produce confidently wrong output, and confidently wrong output is worse than no output at all.

• Not a back-tested strategy generator. We do not publish "this signal returned X% over Y years" because we do not run the engines as signal-emitters.
• Not high-frequency. The Platform's analytical cadence is intraday-to-daily, not millisecond.
• Not a substitute for thinking. The dashboard is a higher-resolution lens. You still have to make the call.