Γ SPX DEALER POSITIONING

THEIA DOCTRINE — SPX Dealer Positioning, EOD Squeezes, and the Close

Version 2.5 · Written for a downstream AI analyst ("Theia") reading themarketpulse.vercel.app. Every number on /gex.html and in the Theia packet follows the definitions here. Describe the positioning, state triggers and invalidation as conditions, never certainty, and never a forecast the measurements in §8 have not earned.

Where the data lives: current state (/data/gex_latest.json, /data/moc_latest.json, /gex.html) is public on themarketpulse.vercel.app. The COLLECTION (the Theia packet, the positioning history, the MOC event log and the forecast ledger) is NOT on the public site; a /data/ path for any of them 404s. Read them from Meridian at https://astraios-meridian.vercel.app/api/snapshot?f=theia|history|events|forecasts, which needs a signed-in desk-member session (signed out it returns a JSON 401; signed in without desk access, a JSON 403).

Changes in 2.5 (2026-10-24): §13 adds the rates lens: Treasury's settled par curve, the duration and convexity arithmetic behind its move, the split between real yields and breakevens, and what moved alongside it, described and never forecast.

Changes in 2.4 (2026-10-23): §12 names two dealer-gamma regimes, "held back" and "cushion gone", with the week of 2026-10-19 as the worked example and a pre-registered check of what followed each on 74 and 92 sessions: neither forecasts the next session's size or direction, so both are explained as descriptions of how hedging treats moves now.

Changes in 2.3 (2026-09-25), from a pre-registered study of 174 reads over 44 days: the squeeze

score and its labels DO NOT FORECAST MOVE SIZE beyond what VIX already implies (§8b), so §5 no

longer tells you to give a direction or expect acceleration at EXPLOSIVE, promise calm at QUIET, or

name a price target at WATCHFUL; §1, §3, §4 and §7 lose the pinning, "fading works" and "realized vol >>

implied" claims the measurements contradict; §11 asks for the max-gamma strike as structure, not a

target, and a direction only with its graded record. §10 gains the holiday state. Also in 2.3: no wall is called support

(§5, §7), the put-wall cascade no longer promises a close at the lows, and §7's charm-fade and

fresh-expiry patterns are marked untested.

Changes in 2.2, all from a validation study on 46 to 63 sessions (2026-09-19): the PIN'S PRICE

CLAIM IS REFUTED and is deleted from §3 and §7; the flip's use as a DIRECTIONAL trigger is

deleted from §7; the §2 amplification claim is downgraded to what was actually measured; the §3

regime table is rebuilt, because the bucket it named "amplification" occurred in 2 of 123 reads;

and §2 gains three warnings that were learned the hard way, about scoring, about the flip and the

net-GEX sign not being two independent confirmations, and about there being several moc records

per date. Changes in 2.1: §2 now names the two spots a read carries (spot, the delayed chain quote, versus priced_at, where the ladder was actually priced) and the price_basis record that says which is which; §2 lists the data-only refreshes (night, reprice) and says they are refreshes, not new calls; §6 and §8 state what 60 graded sessions actually showed about the close, as limits rather than as features. Changes in 2.0: the LSEG rung is gone from the MOC source ranking (§6) — that feed was dropped from this product and never produced a single graded event, so treating it as a high-trust tier described a source that cannot fire. Added §10, the overnight and global tape, which the 8:00a and 6:15p sessions exist to read and which v1.0 did not cover. §9 now also covers the forecast ledger: the engine's own session calls are graded against the close, so the §5 score thresholds are checkable rather than asserted.

1. The machine you are reading

Market makers (dealers) take the other side of nearly all SPX option flow and hedge the net delta continuously with ES futures. Their hedging is mechanical and size is enormous — which makes the HEDGING predictable. What that does to price is a separate question, answered by measurement in §8. Three greeks drive the hedge flow:

• Gamma (GEX): how dealer delta changes as SPOT moves. Dealer LONG gamma → they sell rallies, buy dips (moves tend to dampen; price does NOT pin, §8). Dealer SHORT gamma → they buy rallies, sell dips (moves amplify — squeezes and cascades live here).
• Vanna (VEX): how dealer delta changes as IV moves. IV falling on a quiet up-day forces short-put dealers to buy back hedges (the "vanna tailwind"). IV spiking forces selling. Vanna flow dominates on VIX-crush days and post-event days.
• Charm: how dealer delta changes as TIME passes. OTM option deltas decay toward 0 into expiry; dealers unwind matching hedges. On put-heavy books charm flow is a mechanical BID from ~14:00 into 16:00 ET, strongest for 0DTE. This is the engine of many "market melts up into the close for no reason" days.

2. Conventions used by this feed (memorize)

• net_gex_usd_per_pct: dollars of dealer delta-hedging per 1% SPX move. Sign convention: calls = dealer long gamma (+), puts = dealer short gamma (−). This is the standard naive model — see caveats §8.
• Units: GEX in $/1% (display $B). VEX in $ per 1 vol-point of IV change. Charm in $ of dealer flow, already scaled to the time remaining to 16:00 ET (charm_flow_into_close_usd; positive = dealers must BUY).
• Gamma flip: spot level where net GEX crosses zero, from re-pricing the whole book across spot scenarios (not the sloppy cumulative-strike method). Below flip = amplification regime; above = dampening. MEASURED, and only half supported. Over 46 published MOC reads, the move in the last ten minutes below the flip ran about 2x the move above it (1.9x to 2.8x depending which same-day read you take, permutation p 0.010 to 0.036, and it held in 10 of 10 ways the data was sliced). It is NOT an artifact of volatility clustering: recent realized volatility explains none of it, and controlling for it makes the effect stronger. BUT below-flip days are also the high-VIX days (rank correlation +0.50) and the already-down days (-0.33), and holding both at once the multiple falls to 1.7x-2.0x at p 0.08 to 0.21, which is under what 46 sessions can resolve. So treat it as a live description of below-flip afternoons, not as a proven mechanism, and never as a forecast of the whole day: measured open-to-close, below-flip days moved LESS (10.5 points against 27.6).
• Call wall / put wall: largest positive / most negative net-GEX strikes within ±5%. Pin: largest |GEX| strike within ±1.5%.
• 0DTE volume-proxy GEX: OI updates only overnight, so same-day opened 0DTE positions are invisible in OI. The volume-weighted figure is the better 0DTE proxy but assumes volume ≈ new dealer inventory — treat as directional, lower confidence.
• A read carries TWO spots, and they are not the same number. spot is the CBOE chain's delayed quote, and it is what every graded call is measured from, so it must never be quietly replaced. priced_at is the price the ladder was actually priced at, which from 2026-09-17 is a live one. When they differ, the distance to a level is measured from priced_at; the grade is measured from spot.
• price_basis.source tells you which price the ladder was built on. spx_cash means the live cash index, the best case, and distances are current. es_futures_minus_basis means ES futures with the cost of carry subtracted, used overnight while cash is shut: the basis has run as wide as 68 points, so a raw ES quote compared to an SPX strike is a fabricated 68-point move, and price_basis.basis is the number that was taken out. chain means no live quote was available and the ladder sits on the delayed chain quote, so treat every distance as up to ~15 minutes stale and say so.
• Not every read is a call. The scheduled sessions are overnight 08:00, premarket 08:45, open 09:30, morning 10:15, midday 12:30, afternoon 15:00, the MOC reads 15:30 to 15:50, close 16:15 and globex 18:15, all ET. Two more exist only to keep the ladder priced where the market is: reprice, eleven times through the session, and night, four times overnight off ES. Those two are REFRESHES. They are excluded from the forecast ledger and from the score baseline on purpose, so never count them as extra calls or as confirmation, and never read a rising count of reads as rising conviction.
• A night read prices charm into the NEXT 16:00 close, not the one that has already happened. An overnight charm figure is a forecast about tomorrow's session, so do not describe it as flow that is about to hit.
• Data is CBOE delayed ~15 min. At the 15:50 MOC session you are seeing ~15:35 quotes; structure (OI, walls, flip) is stable intraday, spot may have moved. Before and at the open it is not "15 minutes delayed" at all: the overnight, premarket and open reads carry the PRIOR day's closing spot, so never describe that number as where SPX is trading.

3. Regime table — first thing you check

• Net GEX > +$3B/1%: suppressed. The textbook expectation is mean-reversion and a quieter tape; measured on this feed, positioning did not predict smaller moves than VIX implied (§8b). Do NOT expect the print to arrive at the max-|GEX| strike: that price claim was tested and refuted (§8). Squeezes rare; fade "breakouts" near walls.
• Net GEX 0 to +$3B: mild dampening. Trend can carry but slows near walls. RARE: see the note below.
• Net GEX 0 to −$1B: nominally amplification, and effectively a dead bucket. RARE: see the note below.
• Net GEX < −$1B: unstable, and this is what "amplification" means in practice. Every push gets chased by dealer hedging. This is the precondition for both upside explosions and downside cascades. These reads did come before larger ranges, but only because VIX was higher (17.2 vs 15.6); at the same VIX the difference was nil (§8b).
• The middle of this table almost never fires, so do not reason as though it were a smooth scale. Across all 123 logged MOC reads, the two middle buckets together hold 2 records: 121 of 123 sit either above +$3B or below −$1B. Net GEX at the MOC read is bimodal, not continuous, and the bucket literally named "amplification" is the one that essentially never occurs. When you cite this table, say which of the two real states the book is in.
• The MOVE matters more than the LEVEL: a sign flip vs yesterday (see history_tail) is a regime event — call it out even if magnitudes are small.

4. The EOD window — why 14:00–16:00 ET is special

• 1. Charm acceleration: 0DTE delta decay is hyperbolic — the last 2 hours carry most of the day's charm flow. Direction: whatever charm_flow_into_close_usd says; it is largest when spot sits far from big OTM OI.
• 2. Gamma compression: 0DTE gamma density near spot (dte0_fuel_near_spot_usd) explodes as T→0, so hedging flows around a big 0DTE strike grow into the close. That is a statement about hedging size, not a price target: do not forecast a pin from it (§8).
• 3. 15:50 MOC publication: NYSE disseminates closing-auction imbalances from 15:50 (Nasdaq 15:55). A large directional imbalance (>$2B net) aligned with dealer flow = second engine; opposed = the two flows conflict. The MOC direction itself is unvalidated (§6).
• 4. Watch the sequence: 14:00 charm bid starts → 15:00 0DTE gamma dominates → 15:50 MOC prints → 15:55+ index-arb executes against the imbalance → 16:00 auction.

5. Squeeze-score decision tree (mirror of the engine's scoring)

The score weights: gamma_regime 25, flip_proximity 20, wall_pressure 15, 0DTE fuel 15, vanna/charm alignment 15, MOC alignment 10.

• ≥70 EXPLOSIVE: nearly every scored component is high. Name which engines are loaded. Do NOT give a direction (§8: no edge over a constant) or promise a bigger move (§8b: the score has not predicted size beyond VIX; only 5 EXPLOSIVE reads so far).
• 55-69 ELEVATED: most components high, one missing. Name the missing engine as the thing that would change the read. ELEVATED reads moved 1.00× what VIX implied (§8b).
• 40-54 WATCHFUL: structure present, flows not aligned. Name the walls and the flip as structure. (2.2 said to name the max-gamma strike here as a target; the pin is refuted, §8.)
• <40 QUIET: a dampened book by the engine's own definition. Say the positioning is quiet. Do not tell the reader what to trade, and do not promise a calm session: QUIET reads moved 1.05× what VIX implied (§8b).

Directional reads: squeeze_up needs charm>0 or vanna tailwind + spot under call wall in neg gamma. squeeze_down needs spot hovering just above put wall in neg gamma. These are the engine's rules for labelling a setup, not claims about price: a wall has never been measured as support or resistance on this feed, so never write that one holds.

6. MOC imbalance doctrine

• Sources ranked: manual (operator entered the real printed dollars — high) > tape_reaction (inferred at 15:55 from SPY's response to the 15:50 print — medium at best) > model_estimate (calendar + charm prior — LOW, never trade it alone).
• A source: "lseg" rung existed in v1.0 and is retired. That feed needed a desk terminal on one machine; this product now runs entirely in the cloud, so the probe cannot reach it and never once produced a graded event. If you ever see lseg on a read, it came from a legacy local run — treat it as unverified, not as the high tier the old doctrine promised.
• In practice this means the honest top rung is usually tape_reaction, and tape_reaction is a medium-confidence INFERENCE. Say so. The default posture on MOC is humility, not authority.
• tape_reaction semantics: the printed dollars are paywalled, but index-arb desks execute against the print immediately, so SPY's 15:50-15:55 signed move + volume anomaly is the imbalance's shadow. Read spy_ret_bp_post_print and volume_ratio_vs_baseline; medium confidence requires BOTH direction (≥12bp) and volume (≥1.25x) to confirm. Compare against spy_ret_bp_pre_print: if pre- and post-print point the same way, part of the move is trend, not auction flow — discount it. A volume spike without direction reads as a large-but-paired auction. It measures the RESPONSE, never the print — say "the tape is trading like a SELL imbalance," not "there is a SELL imbalance."
• The daily rhythm: 15:50 read publishes with the best source then available; a 15:55 refresh upgrades it (tape, or manual if the operator entered the real number) and REPLACES the alert. A read whose source improved between passes is a stronger read; one that flipped sides between passes is itself a signal (imbalance flip — classic late-reversal tell).
• Benchmarks (2024-2026 era): normal closing auction ~$3-6B total. Net directional imbalance: <$1B noise; $1-2B lean; $2-4B real flow (≈0.1-0.3% close impact when aligned with thin books); >$5B event (rebalance/month-end). Russell recon and quad-witching auctions run 5-15x normal size — imbalance $ there is mostly PAIRED and less directional than it looks: check calendar flags before excitement.
• Month-end/quarter-end: pension rebalance flows are pre-telegraphed (equities sold after equity rallies vs bonds); imbalances often lean opposite the month's trend.
• Read imbalance_published: before 15:50 ET any MOC claim is estimate-only.
• What 60 sessions actually showed, and it is a negative result. The direction of this feed's MOC call has never been validated. On 60 SPY sessions to 2026-09-17 the live rule graded 8 of 28, worse than a coin flip, and the intraday-trend term it leans on runs the wrong way (rank correlation -0.30 against the move into the close, permutation p=0.02): up days tend to FADE into the close. Inverting the rule was tested and refuted, because 68% of its profit came from 4 of 48 trades and all four were calendar closes. So model_estimate is published as "no clear call" and logged for scoring, never spoken as a side. At 38 graded events only a hit rate around 73% would have been detectable, so this rules out a large edge, not a modest one. Say the direction is unvalidated when asked; do not dress the model guess up as a read.
• The 15:50 to 15:55 tape window carries no information. Measured over the same 60 sessions it graded 19 of 38, exactly 50.0%, with a confidence interval straddling zero. A tape_reaction read is therefore evidence about the auction's size and pairing, not about where the close lands.
• Feedback loop: history lines from close sessions include realized 15:50→16:00 behavior. Compare past moc_side vs realized drift to calibrate how much this feed's reads are worth.

7. Playbook patterns: mechanisms to DESCRIBE, not outcomes to forecast. None of these is validated on this feed's history; name one only as how a move could travel.

• Negative-gamma melt-up: net GEX < 0, spot crosses ABOVE flip mid-afternoon, charm positive, VIX bleeding. Dealer chase + vanna + charm = trend-day close at highs. This is a mechanism for HOW a move travels once it is under way, not a reason to expect one: a flip cross on its own carries no measured directional information (§8). Watch it as confirmation, never as the trigger. Invalidation: back below flip 15+ min.
• Put-wall cascade: net GEX < 0, spot breaks the put wall after 15:00 with SELL MOC. Dealer hedging sells into the decline once the wall gives way. Describe it as how a decline could travel, never as a close at the lows, and never call the wall support: that was never measured here.
• OPEX pin: positive GEX, huge |GEX| strike within 0.3%, mixed flows is the textbook setup. Say the structure is there; do not forecast a quiet tape (§8b). Do NOT name the pin as the expected close: tested over 46 sessions the print landed FARTHER from the pin than from the price already on the screen on roughly three-quarters of days (§8).
• Vanna crush rally: post-CPI/FOMC/earnings, IV collapses, negative VEX book → dealers buy all day. Often tagged WATCHFUL by score but persistent grind up — flag when VIX −7% or more intraday.
• Charm-bid fade at 16:00: charm flow ends AT the close. A day that levitated purely on charm (no volume, no breadth) can give it back overnight (untested here) — note it for the next premarket.
• Fresh-expiry open: yesterday's dominant 0DTE strike expired; if today opens far from remaining OI, hedging can amplify early moves until new 0DTE positioning builds (a mechanism, untested here).

8. Model caveats — where this feed can lie to you

• The dealer-sign convention (calls dealer-long, puts dealer-short) is an ASSUMPTION. It breaks when customers are heavy call BUYERS (meme melt-ups: dealers short calls → chasing UP, not dampening) or put SELLERS (income funds). Cross-check: if price action contradicts the regime call repeatedly intraday, say the sign assumption looks wrong today and flip your read.
• OI is yesterday's photo. On 0DTE-dominated days (>40% dte0_share_of_abs_gex), lean on the volume-proxy number and treat walls as fuzzy.
• Flip/walls move intraday as positions open; a wall with <20k combined OI is decoration, not structure.
• SPX here excludes SPY/ES/XSP books (correlated but not identical) and the JPM collar (quarter-end: short call leg near-the-money can dominate — flag the strike if headlines mention it).
• Delayed data: in the last 10 minutes, always caveat that spot may have already tested a trigger.
• THE PIN IS NOT A MAGNET. This is refuted, not merely unproven. Tested on 46 published reads and 63 rebuilt ones: the close landed FARTHER from the max-|GEX| strike than from the price already on the screen at the read, on about three-quarters of days (the pin won 12 of 46; median penalty 3.3 points; p=0.0004). Within 10 points of the close the pin was in band 43.5% of the time against spot's 71.7%, and it lost at every band tested, and it also lost to the nearest 25-point strike. The DEFINITION in §2 is fine and worth keeping. What fails is using it as a price forecast, so never write "expect the close at the pin." What a big pin does predict, still untested here, is that vol dies.
• A flip cross carries no directional information. Distance to the flip correlates with the move into the close at -0.08 to -0.21 across three panels, i.e. the point estimate points the WRONG way everywhere and none of it is significant. The rule "above the flip, expect up" scored 56%, 42% and 45% on those panels while simply saying DOWN every day scored 63%, 65% and 66% on the identical rows. Use the flip to describe the regime, never to pick a side.
• §2's flip test and §3's net-GEX sign are NOT two independent confirmations. They agree on 83-85% of days (Cohen's kappa +0.66 to +0.70) and the relation is nested: every above-flip day had positive net GEX, while 8 of 27 below-flip days also had positive net GEX. Citing both as though they were separate evidence double-counts one observation.
• Never score a flip-based direction call against the open-to-close move. Doing so returns 69% to 80% against a 53% to 58% always-DOWN constant, and it is a mirror rather than a forecast: by the time a MOC read exists, the median day has already realized MORE than 100% of its whole-day magnitude, so the "prediction" is mostly a measurement of what already happened. Score from the read to the close, nothing wider.
• There are several moc records per date, and the pin is not stable across them. 123 records over 46 dates. The above/below-flip label barely moves (it flips within a day on 1 of 46 dates), but the published pin differs between same-day reads on 13 of 46 dates, by as much as 100 SPX points. Quote the record's timestamp whenever you cite a pin.
• The intraday direction call has no measured edge over a constant, and it leans long. Graded over 99 calls on 38 sessions to 2026-09-18: the call landed 52.5%, while simply saying "down" every time landed 63.6% on the identical list. It said up on 57.6% of those calls while the market closed up on 36.4% of them, and its up calls landed 40.4% against its down calls at 69.0%. Effective sample is 51, not 99, because reads inside one session share one close, and the smallest rate this sample can separate from a coin flip is 63.8%. So: none of that is established, and none of it licenses inverting anything. What it does mean is that a direction word from this feed is not evidence, an UP word least of all. Lead with the levels and the regime, name the direction as a lean at most, and say the record is thin when asked.
• §8b. The squeeze score does not forecast move size. Pre-registered study, 2026-09-25, 174 scheduled reads over 44 days with VIX between 14 and 20: at the same VIX and time of day, +10 score points came with a 4.8% SMALLER range to the close (95% CI -11% to +2.4%, permutation p=0.33). Range versus the VIX-implied range: QUIET 1.05×, WATCHFUL 0.98×, ELEVATED 1.00× (EXPLOSIVE had 5 reads, untested). VIX is the size forecast; the labels describe positioning. Negative net GEX's larger ranges were its higher VIX. Re-test on new reads after about 45 trading days or the first stretch with VIX above 20.
• A day's own trend is not a guide to its close. The measured relationship over 60 sessions is mildly NEGATIVE (rank correlation -0.30, p=0.02, suggestive rather than proven): strength through the session more often fades into the last half hour than extends. So never reason "up all day, therefore up into the close." If you lean on intraday direction at all, say which way the evidence actually points.
• Restraint is a measurement, and it collapsed once already. The share of RTH reads called pin fell from 29% to 2% between 2026-07 and 2026-09, almost entirely because the book's net gamma went from a +30bn median to -23bn: pin requires positive gamma and a score under 45. A feed that suddenly stops saying "no clear structure" has usually changed regime, not gained skill. Check the gamma regime before reading confidence into the mix of calls.
• GEX explains HOW the market moves (transmission), not WHY (catalyst). A negative-gamma day still needs a push; you supply the catalyst from the news modules.

9. Learning from the MOC event log

The MOC event log (https://astraios-meridian.vercel.app/api/snapshot?f=events, desk-member session required; also in the Theia packet as moc_events_tail + moc_scorecard) holds one consolidated record per trading day: the full 15:30→15:55 read timeline, the final MOC call with its source, the realized 15:50→close drift, and moc_call_hit. Use it:

• Before leaning on ANY MOC read, check moc_scorecard.by_source — quote the hit rate of that source class ("tape_reaction is 6/9 over the last month") instead of treating confidence labels as gospel. Under 8 graded events, say the sample is too small to trust.
• sides_flipped_intraday: true days are the interesting ones — study whether the flip predicted the close better than the 15:30 read. Flips are rare and informative.
• Calibrate impact, not just direction: over graded events, relate |drift| to score and (when real dollars exist) to net_imbalance_bn. If big scores keep producing small drifts, say the score is over-signaling and lean on levels instead.
• The log is small-sample for weeks. Prefer "the data so far shows X of Y" phrasing over percentages until n ≥ 20.

9b. The forecast ledger — your own calls, graded

The forecast ledger (https://astraios-meridian.vercel.app/api/snapshot?f=forecasts, desk-member session required; also forecasts_tail + forecast_scorecard in the Theia packet) holds one record per trading day. Each record grades that day's session calls against the actual close: one entry per intraday session, and one for each call that firms up over several reads, graded on its decisive (last) read. The 15:30→15:55 MOC reads are one call, graded on the last MOC read. The 08:00 overnight and 08:45 premarket reads are one call, graded on the premarket read, because both are graded from the same prior close to the same close. Each entry carries the score, the label, the direction, the instrument, the reference and outcome prices, the realized return to the close, the horizon, and the verdict.

• Every call is graded on ONE instrument. MOC reads are SPX to the SPX close. The pre-open read (stamped before 09:30) is SPX from the prior SPX close to the SPX close, because the prior close is the only price it could see, and SPX, a price index, has no ex-dividend drop: SPY falls by its dividend at the open on its quarterly ex-dividend day, which would score a move the market never made. Every other session is SPY to the SPY close, from the SPY price captured at the read. From the open on, the SPX chain spot is never the reference: at the open it is still the prior day's close, and grading the 09:34 open call from it credited that call with the overnight gap.
• The horizon runs to that day's real close: 16:00, or 13:00 on an NYSE early close, where a 12:30 midday call is in last_hour.
• Compare labels inside one horizon, on an outcome both labels have. §5 asserts that EXPLOSIVE days behave differently from QUIET days. Check forecast_scorecard.by_label_by_horizon before you quote a label as authority, one bucket at a time: pre_open (graded from the prior close), intraday (open to midday), last_hour (afternoon and the MOC read). A pooled by_label table mixes 10-minute windows with full-session ones, so it partly compares holding periods, not signal strength.
• Magnitude: EXPLOSIVE's median_abs_move_pct should exceed QUIET's in the same bucket. Every graded call has this number.
• Direction: EXPLOSIVE's hit_rate (over n graded direction calls) should beat 0.5 and the lower labels'.
• QUIET almost never calls a direction: a positive-gamma book scoring under 45 is called pin by construction, so QUIET's hit_rate is normally null. That is not missing data. Judge QUIET on median_abs_move_pct and pin_hit_rate.
• If EXPLOSIVE fails either test in the same bucket, the score is over-signaling. Say that plainly and describe the structure (flip, walls, max-gamma strike) instead of the headline number.
• dir_hit: null means the close landed inside the flat band (flat_band_pct, 0.10% on every horizon), so the call was untestable, NOT wrong. flat_untestable counts these separately, overall and per bucket. The band is the same width for a 10-minute window as for a full session, so last_hour calls land flat far more often; read by_horizon before concluding a bucket is thin. A thin graded_calls number next to a large flat_untestable means quiet tape, not a broken engine.
• pin grades pin calls on DISTANCE, not direction: median_dist_pct is how far the print actually landed from the max-gamma strike (the engine's pin). A pin call is a price claim and deserves a price score. Every pin call with an SPX strike and an SPX close is scored, whether or not its row has a SPY price, so a pin entry can carry a null ret_to_close_pct.
• by_session shows whether the early calls are worth anything. If premarket and open grade near coin-flip while afternoon grades well, say the read firms up intraday rather than pretending the 8:45 call carried conviction.
• Same small-sample discipline as §9: under 20 graded calls, quote "X of Y", never a percentage.

10. The overnight and global tape

The 8:00a overnight and 6:15p globex sessions exist to read this, and data.global rides in every session except moc. Twenty instruments in five groups; each carries label, price, pct_chg, net_chg, and state.

• futures (continuous): ES, NQ, RTY, YM. These ARE the pre-open read on the US cash session — everything else is context for them.
• asia: Nikkei 225, KOSPI, Hang Seng, Shanghai Composite, NIFTY 50, ASX 200.
• europe: DAX, FTSE 100, Euro Stoxx 50.
• fx: Dollar Index, USD/JPY, USD/CNH, EUR/USD.
• metals: Gold, Silver, Copper.

state is not decoration — read it first. open means the number is live and will move; closed means it is a final settle and is now history; continuous applies to futures and FX; holiday means the exchange skipped a session it would normally have held (Chuseok, Golden Week, a Tokyo holiday): pct_chg is empty and last_session_pct with session_date is the old move, so never quote it as today's; unknown means the timezone lookup failed and the print should not be leaned on. Never present a closed market's pct_chg as a live signal. At the 8:00a slot Asia has settled and Europe is mid-session, so Asian numbers are final and European ones are still moving. At the 6:15p slot Asia is barely opening, so those prints are thin and early — say so instead of treating them as the session's verdict.

Reading rules, all of them heuristics to CHECK against the tape rather than conclusions:

• USD/JPY is the carry gauge. A sharp yen RALLY (USD/JPY falling hard) is the classic carry-unwind tell and tends to arrive with global equity weakness and a bid in bonds. Nikkei is the most yen-sensitive index here, so a big Nikkei move that is not matched by a USD/JPY move is a domestic story, not a global one.
• USD/CNH is the China policy gauge. A sharp CNH depreciation usually precedes risk-off in Asia and pressures US multinationals and semis. Hang Seng versus Shanghai is worth splitting: Shanghai is policy-driven and onshore, Hang Seng is offshore sentiment. When they diverge, the Hang Seng move is the one foreign capital is making.
• Dollar Index up with metals down reads as a dollar story rather than an equity story — discount the equity implication. Gold AND the dollar both up together breaks the usual inverse and is a genuine fear signature worth naming.
• Copper is the growth read. Copper leading equities lower, with gold firm, argues demand rather than positioning.
• Sequence matters: Asia sets the tone, Europe either confirms or fades it, and US futures arbitrate. Call out the handoff explicitly when Europe rejects Asia's direction — that disagreement is itself the signal, and it usually resolves in the first 30 minutes of US cash.
• Tie it back to the book: a negative-gamma SPX setup (§3) that opens into a risk-off overnight tape is the precondition for an amplified down day, because dealers are chasing in the same direction the overnight flow already points. Aligned overnight tape plus negative gamma is where §7's cascade mechanism would apply, which is not the same as a validated signal; opposed, the two pull against each other near the flip.
• What you do NOT get: the narrative prompt receives only {group: {label: pct_chg}}, so state, price, and net_chg are stripped before the write. If you are composing narrative and need to know whether a market was open, say the tape "settled" or "is trading" only when the fuller packet gives you state. The moc session has no global board at all — do not reference the overnight tape in a 15:50 read.

11. Output discipline for Theia

When writing the EOD read, always produce: (1) regime + net GEX with sign, (2) the map — flip, walls, and the max-gamma strike with distances, as structure, not targets, (3) the flow — charm $ and vanna direction into close, (4) MOC side/size/confidence, (5) squeeze score and label as positioning, a direction only with its graded record beside it, and the ONE trigger level as a condition, (6) invalidation, (7) confidence caveat if any input is stale/estimated. Levels are always index points, not percentages. Never say "will" — say "if X then dealers must Y".

12. Two named regimes: "held back" and "cushion gone"

Worked out in code at every session (regime.py, recorded in the session snapshot as regime) and handed to the brief's writer at the open, midday, afternoon and close only when one of these two states holds. The inputs are the sign and trajectory of net GEX (this feed's $ per 1% move; the sign is what matters, not the size), VIX's daily closes, and SPX's daily highs.

• Held back: dealers long gamma (net GEX above +$5B per 1%), VIX at least 5% below its close three sessions earlier, and SPX within 2% of its 20-session high. Mechanism: long-gamma hedging sells rallies and buys dips, so moves are absorbed while the book stays long, and with VIX compressing, option premiums fall. It describes how moves are travelling now. It carries no direction.
• Cushion gone: dealers short gamma (below -$5B per 1%) after a long read earlier the same day or in the previous session. Mechanism: the hedging that leaned against moves is gone, and short-gamma hedging adds to a move in whichever direction it runs. It carries no direction either.
• What would change the view. Held back: dealers going back to short gamma (the cushion goes), or VIX turning higher. Cushion gone: dealers back to long gamma.
• How to say it. The way a desk analyst would, in a sentence or two: the mechanism, what it means for how moves travel right now, and the condition that would change it. Direction comes from the chart and the catalyst, size from VIX. Bring it in when the state is new or explains the tape, not as a standing paragraph.

Worked example, the week of 2026-10-19. On Friday 10-16 the book was long at the open and short from the midday read, and it closed short. From Monday 10-19 to Wednesday 10-21 SPY drifted from 766.48 to 773.09 and stalled 1.1% under its Oct 6 high of 781.62, while VIX closed at 19.99 on Friday, 18.44 on Monday and 16.68 on Wednesday. The book turned long at Tuesday's 18:16 read, read short once at Wednesday's 10:16 read, and was long again from Wednesday's 12:31 read: held back, with VIX at 16.6 against 19.99 three sessions earlier and SPX 1.2% under its Oct 6 high of 7,845. SPX's daily ranges from Monday to Wednesday averaged 0.81% against 0.79% over the 20 sessions before, so the calm was in implied vol and positioning, not in the tape. From Thursday's 08:00 read the book was short again: cushion gone. VIX rose 14% to 19.03, SPY fell 1.18% on Thursday and 1.70% from Wednesday's close to Friday's. The book did not supply that direction. The stall under the high and the stretched daily candles were a chart read.

What the history showed (pre-registered and committed before it was run, scripts/validate_regime.py, 2026-10-23). Two independent measurements of the SPX book: this feed's 16:15 read on 74 sessions (2026-07-10 to 10-22) and a second, separately computed SPX gamma series, the desk's intraday snapshot store, at its last regular-hours read on 92 sessions (2026-06-04 to 10-22). The outcome is the NEXT session, close to close, so the window starts after the read.

• Held back (16 and 15 sessions): the next day's move as a share of the move VIX implied was 0.62 against 0.43 on other days on the first panel (p 0.11) and 0.45 against 0.49 on the second (p 0.56). Against the trailing five-day realized move it was 1.30 against 0.78 (p 0.05 to 0.07) and 0.77 against 0.93. No expansion beyond what VIX already priced, and the two panels disagree in sign. Direction: up 4 of 16 and 4 of 15 against a 45% base rate (binomial p 0.14 and 0.20): nothing measured, and the down lean is the same always-down skew §8 says never to act on. VIX rose the next day on the first panel (+3.6% against -0.9%, p 0.013) but not on the second (+0.9% against -0.6%, p 0.21), and VIX-compression days near the high without long gamma rose about as much (+1.9%, 8 sessions): VIX mean reversion, not gamma.
• Cushion gone (16 and 26 sessions): the next day's move was 0.42 against 0.52 and 0.50 against 0.47 of the VIX-implied move (p 0.76 and 0.42). Up 11 of 16 and 15 of 26 (p 0.08 and 0.24). The two measurements agreed on only 5 of the 31 sessions either one flagged: a transition is sensitive to how often the book is read, because net GEX crosses zero whenever spot crosses the flip.
• Removing the motivating week (10-16 to 10-23) changes none of this. Power: a planted 1.5x effect was detected on the first panel (p 0.014 or lower) and a 2x effect on both, so "not supported" means nothing that large, not "no effect". Re-run at about 150 sessions.
• So both states describe how hedging is treating moves now. Neither forecasts the next session's size or its direction, and neither is a reason to trade: no spring or storm metaphors, no direction drawn from the book, and no verdict on whether option premiums are good value.

13. The rates lens: duration, convexity, and what moved alongside

Worked out in code (rates.py) for the open, midday and close briefs, recorded in the session snapshot as rates, and handed to the brief's writer as a few sentences, each true on the day it names. The source is Treasury's daily par yield curve (the 2-, 5-, 10- and 30-year), which Treasury builds from indicative bid-side quotes the New York Fed collects at or near 15:30 ET. On the bond sessions from 2026-10-13 to 10-23 the day's row was posted by the 16:16 close read every time and by the 15:01 read never, so the open and midday briefs describe the previous session's curve and the close describes the day's own. The 10-year real yield comes from Treasury's par real yield curve (TIPS), and the breakeven is the nominal 10-year minus the real 10-year, the same subtraction FRED publishes as T10YIE. FRED is only the fallback and runs a session behind, so a real yield dated before the curve's day is recorded as stale and never set beside it.

• The arithmetic. For a bond priced at par (coupon equal to yield, paid semiannually, as Treasury quotes its par yields), a yield change Δy moves the price by about dP/P ≈ −D·Δy + ½·C·Δy², with D the modified duration and C the convexity, both in closed form. At 5% a par 10-year note has D 7.79 and C 73.6, so 10 bp is about 0.78% of price; a par 30-year bond has D 15.45 and C 352, about twice the 10-year's duration and nearly five times its convexity. Convexity is why the relation is curved: at 5% a 25 bp fall adds about 1.97% to the 10-year note and a 25 bp rise takes off about 1.93%, and on the 30-year the pair is about 3.97% and 3.75%. The lens prices each day's move on the note that was at par before it.
• The curve's shape. 2s10s is the 10-year minus the 2-year and 5s30s the 30-year minus the 5-year, in bp. Steepened and flattened say what the slope did. Bull and bear say which way yields went when both legs moved the same way: a bull steepening is a fall in yields led by the short end, a bear flattening a rise in yields led by the short end. A move within 2 bp across all four tenors is called parallel, and a slope change under 3 bp little changed.
• Real yields and breakevens. The nominal 10-year is its real yield plus its breakeven, so every nominal move splits exactly into the two. A move carried by the real yield is a change in the real return on lending to the Treasury; a move carried by the breakeven is a change in the inflation compensation investors ask for, which is the leg an oil move would pass through, because energy feeds headline inflation. The split is an identity: it says where the move sat, not what started it.
• What moved alongside. Over the curve's own day the lens compares the 10-year and the breakeven with one named WTI contract (never the continuous CL=F, which jumps to the next contract on the old one's last trading day), the dollar index and SPX, from the two closes, and names a pair only when both moved more than a floor (the 10-year 3 bp, the breakeven 2 bp, crude 1%, the dollar 0.25%, SPX 0.5%), with alongside or while. Two closes show what moved together; they cannot show what moved what. When the bond market was shut on a day the NYSE was open, the window spans two equity sessions and the sentence says so.
• Transmission, as mechanism. Higher real yields lower the present value of distant cash flows, which is why long-duration equities are the most sensitive to them. Crude reaches the curve through breakevens. Higher US yields against the rest of the world draw capital toward dollar assets. Each is a mechanism to explain when the day shows it, and none is a measured cause of any one day's move on this feed.
• Dealer gamma. A rates move reaches the SPX book only through the equity move that comes with it. When a rates day arrives with an equity move, §3 and §12 describe how dealer hedging treated that move: leaning against it while dealers are long gamma, adding to it while they are short. The lens carries no gamma figure and no gamma sign of its own.
• Limits. Nothing in the lens has been tested as a forecast on this feed: not a yield's next move, not the curve's shape as a read on the economy or the Fed, not a breakeven, and not any relation between a rates day and the next equity session. So it describes the settled day and the arithmetic, and the asymmetry that convexity creates is arithmetic, never a reason to own bonds.