Reversal Context — STREAM BAND + FRACTAL FLOOR — Canonical Reference

One-line definition: the two survivors of the desk's reversal-pair research — STREAM BAND (a 13-EMA high/low band with a 5-EMA HLC/3 trigger) and FRACTAL FLOOR/CEILING (the latest confirmed Williams down/up-fractal as a support/invalidation level) — shipped only as descriptive context columns (band state· own-history stretch percentile· risk-geometry floor), because every tradeable form of the pair was pre-registered and falsified at every level.

1. What it is

Two related structures the desk specified (2026-07-13 voice-note session) for catching reversals off a stretched, beaten-down tape:

• STREAM BAND — an EMA13-of-highs / EMA13-of-lows envelope with an EMA5-of-(H+L+C)/3 "trigger stream" running through it. The read is the position of the trigger relative to the banks (BELOW· CROSS-UP/RECLAIM· INSIDE· CROSS-DOWN/SLIP· ABOVE) plus a STRETCH measure: how far the trigger has deviated from the violated bank, ranked as a percentile against the stock's own trailing history (per-stock, percent, never a rupee threshold — the desk's no-static-threshold rule).
• FRACTAL FLOOR / CEILING — the latest confirmed degree-10 (fallback degree-5) down-fractal low is a well-defined support / invalidation level; the mirror up-fractal high is the ceiling. A 10-fractal that still holds implies a ≥20-bar consolidation by construction. The useful read is risk geometry: how far above a confirmed floor price sits, how old the floor is, and whether it is still unbroken.

Both are pure-Python over the NSE bhav archive, ₹0, no LLM, point-in-time honest (a degree-N fractal is only used once N bars have printed after it).

2. Our variation vs. the standard technique

• STREAM BAND is in the Gann-HiLo / high-low-channel family (sane prior art). The proprietary treatment is the STRETCH percentile vs the stock's own trailing 756-bar history — which resolves the cap-size problem exactly (a small-cap at 8% may be p60; a large-cap at 4% may be p97, the more actionable read), where no global threshold can.
• FRACTAL FLOOR reuses the validated Williams-fractal primitive (fractal_pivots in the wolfe code, shared with the Wolfe lens under the gate — not forked) and reframes the confirmed fractal as a cheap, well-defined place to be wrong rather than as an entry trigger.

3. How it works (methodology)

All constants (EMA lengths, the 756-bar stretch window, fractal degrees, floor lookback) live in the reversal context code (EMA_FAST/EMA_SLOW, STRETCH_WIN, DEGREES, FLOOR_LOOKBACK) — not restated here. The module computes one latest row per symbol into the isolated reversal_context table (bounded snapshot; space rule; no the db code edit — the signal alerts code isolation pattern) and stamps every fractal's knowable_at implicitly by only consuming confirmed pivots.

The Screen+ "rev" column-group renders four/six descriptive cells per symbol (_rev_cells in the screener plus code): band state (with the reclaim cross labelled a caution), stretch %, stretch percentile, and the floor (+ ceiling) level. Two filter pills — ?rev=ri ("⚠ Reclaim· floor intact") and ?rev=si (the bearish mirror, "⚠ Slip· ceiling intact") — surface the band event with its confirmed fractal context. None is an entry or an alert.

4. Status, validation & honesty fence

DESCRIPTIVE-ONLY, and the falsification is the product copy. This is a binding fence — the project keeps a falsified-approaches ledger and softening this status is a blocking error. The reversal pair was tested pre-registered + hash-frozen + tamper-clean and is falsified at every level (all numbers in the strategy ledger):

• Timing (07-13, STREAM BAND). The BUY-cross is an anti-signal — it negatively selects; both matched placebos beat it; the book never approaches the 0.89 hurdle. An early band-reclaim after a real downtrend is a falling knife more often than a reversal → the cross never ships as an entry/alert; only the descriptive state + stretch-percentile survive.
• Confirmation craft (07-14, FRACTAL FLOOR). The up-fractal breakout confirmation adds ≈ nothing over mere floor proximity at fixed horizons; "STRONG" (beats the first two up-fractals) is worse. The floor is the information; the breakout craft is not. (The interim "PROX10 book return/vol 1.04" was a gross-of-cost accounting defect — see next.)
• Risk-geometry book at true cost (07-14b, FENCES). The reconstruction gate caught our own defect: true flat-cost PROX10 = 0.59 (below the hurdle — the candidate was never alive). All three fences then failed — participation cost decays catastrophically with AUM, a random-entry/same-exit control reproduces the book (the entry is ~inert; the P&L was the exit geometry), and realistic fills sink it further.
• Ex-ante selection (07-14c, RECLAIM SELECTION). Even validated factors cannot pick the launchers in advance: the best frozen rule (low-vol + above-own-norm delivery) merely stops the bleeding — it matches random same-stock days (δ vs placebo ≈ 0). The right tail exists but is not identifiable in advance from price/volume features.
• Exit studies (07-14d MANAGED, 07-14e EXIT LAB). the desk's dictated Case-A stop/re-entry stack is not-improved (the two-candle exit churns to −0.50); across ten exit engines the exit law is monotonic in looseness (slow/wide best, profit-takers worst) and the ceiling is ~0.63 OOS < 0.89 — exits shape losses, they cannot mint edge.

What survives (as shipped): the descriptive context columns + the two watch pills. The one reusable, non-tradeable nugget: among knife-bounces, low-vol + above-own-norm-delivery names are consistently less bad — context for triage, never a rank. Any future reversal proposal must cite ledger §§ 07-13 / 07-14 / 07-14b / 07-14c first. This is coherent with project doctrine: price strength is the only gross forward-return engine; reversal/band/floor reads are context, never rankers or fundable alpha.

5. Where it lives (code· routes· DB· timers)

• Compute: the reversal context code (pure-stdlib) → the isolated reversal_context table (one latest row per symbol; the db code untouched). CLI --compute / --selftest. Fractal primitive reused from the wolfe code (fractal_pivots).
• Surfaces: Screen+ (/dash/screen2) "rev" column-group (_revctx_by_sym + _rev_cells in the screener plus code, read-only) + the ?rev=ri / ?rev=si filter pills. Descriptive columns only — band state· stretch %· stretch percentile· floor· ceiling.
• Timer: nightly compute piggybacks on the existing signals chain (no new systemd unit); deploy = scp + writer-safe restart.

6. Data & provenance

• Primary source only — NSE bhav copy OHLC (split/bonus-adjusted via src.automation.adjust); no vendor, no Screener (Guardrail #8). Pure stdlib (prod venv has no numpy), ₹0, no LLM.
• Point-in-time / no-look-ahead — a degree-N fractal is only consumed once N bars have printed after it; the stretch percentile reads only trailing stored values.

7. Terminology canon

• STREAM BAND — the EMA13(high)/EMA13(low) bands + EMA5(HLC/3) trigger stream. Band states: BELOW· RECLAIM (cross-up, a caution not an entry)· INSIDE· SLIP (cross-down)· ABOVE.
• STRETCH — signed % gap between the trigger and the violated bank, ranked as a percentile vs the stock's own trailing history (per-stock, percent — never absolute).
• FRACTAL FLOOR / CEILING — the latest confirmed degree-10 (fallback 5) down-fractal low (support/invalidation) / up-fractal high (mirror). Read as risk geometry (gap %, age, alive?), never as a trigger.
• Reclaim / Slip pills — ?rev=ri = reclaim with the floor still intact; ?rev=si = the bearish slip mirror with the ceiling intact.
• ⚠ **Deprecated as signals — "band-reclaim buy", "fractal breakout trigger", "PROX10 book". These are falsified** (§4); the words survive only as descriptive state labels, never as entries.

8. Decision & session history

• 2026-07-13 (voice-note) — the desk specifies the reversal pair (STREAM BAND his stated priority + FRACTAL FLOOR); build order = descriptive screener → pre-registered study → book only if it survives.
• 07-13 → 07-14e (ledger) — six pre-registered, hash-frozen, tamper-clean studies falsify every tradeable form (timing / craft / risk-geometry book / selection / two exit labs). Full record + numbers: the strategy ledger.
• / — the two Screen+ pills shipped (reclaim ?rev=ri, slip ?rev=si) with the honesty fence in the product copy; the reversal context code is the live descriptive engine.
• (2026-07-14) — this canonical page created (the arc's home moved out of the transient the reversal pair PLAN notes, now retired); origin label 🧑 THE DESK recorded.

9. Open items / frozen work

• RULED OUT (do not re-open) — the reversal pair as a signal / entry / book. Falsified at all four levels; any re-attempt is blocked until it beats the recorded numbers, net of cost, under the same no-leak harness.
• Bear-mirror parity — the ceiling columns exist; a full bear-side study was deferred/disclosed (non-shortable in the harness), never a live short signal.
• Phase-3 chart overlay (the 13/13/5 ribbon + fractal markers on /dash/stock) was a design aspiration in the retired PLAN doc — not built; not required (the descriptive columns are the shipped surface).

10. Sources of truth

• Falsification & numbers (single source): the strategy ledger §§ Study 2026-07-13 / 07-14 / 07-14b / 07-14c / 07-14d / 07-14e.
• Code: the reversal context code (engine + module docstring = the canonical prose)· the screener plus code (the "rev" group + pills).
• Sibling references: the wolfe wave notes (shares the fractal primitive + the selection-not-craft doctrine)· the dvpt notes / the mep notes (the descriptive-only doctrine).
• Provenance: the origins notes (🧑 THE DESK rows: STREAM BAND· FRACTAL FLOOR/CEILING· Reversal context columns).
• Memory: failure-models-ledger, record-and-remind.