Zero-Touch Divergence [ZTD] — what has been verified, and what has not

Status (2026-09-09): a verified detector, under forward observation. It ships in CAT as a bias engine on H1, H4 and D. It has not been tested as an edge on its own, and the one entry layer built on top of it failed every pre-registered test (see ZTD Entry — ten variants, ten verdicts).

What it is

A same-bar, non-repainting rule that combines two earlier indicators — the AVWAP z-score band and the PBK oscillator — into one divergence detector. Upper side; the lower side mirrors it:

1. An AVWAP-Z triangle at an extreme while the oscillator is also at its extreme arms a reference. That excursion above zero is range 1: its highest high and highest oscillator value are recorded.
2. The oscillator touches zero. The reference is locked.
3. The next excursion above zero is range 2. Its running extremes are compared against range 1 as they form, and the rule fires early:
   • a higher high on a lower oscillator peak is a regular divergence (RDIV);
   • a higher oscillator peak on a lower high is a hidden divergence (HDIV).
4. A fire carries the direction, the range-2 extreme as the level that invalidates it, and the zero-touch extreme between the two ranges as the level it aims at.

Reference lifetime, timeouts, and the opposite-extreme cancel are inputs. The defaults CAT runs are the indicator's own (v17): 2,000-bar percentile calibration, "until superseded" references, 240-bar timeout.

Full design: docs/superpowers/specs/2026-09-04-ztd-zero-touch-divergence-design.md. Script: ztd_divergence.pine v17.

What was verified

The rule against the chart. The Pine was verified bar-for-bar on the chart it was designed on: an export audit passed, and a three-bar replay reproduced the chart's values exactly. That is verification that the script does what the rule says.

The port against the Pine. CAT does not run Pine; it runs pyakao.ztd, a Python port. Five golden tests compare every exported plot over 19,074 closed hourly bars against a TradingView export of the same chart, and they agree to machine precision. Twenty-eight further tests pin the rule's edges (timeouts, the rolling anchor's warm-up, the zero-touch reading). The port is also the engine the entry layer's research ran through, so what CAT reports is what the research measured.

Live, against the bridge. On 2026-09-09 the port was run on 2,600 live H1 and H4 bars per instrument from CAT's own data feed; signals fired on real data and matched the chart.

What was not verified, and is not claimed

ZTD as an edge. A divergence bias is graded by what the market does after it, and that record is short. The only mechanical grade so far is a six-month first look on EURUSD H1 (2026-09-06, harmonic/validation/combo_entries.py): 22 signals, of which the seven that also met a secondary harmonic condition came back two won, five lost. That was a look, not a pre-registered test, and the harmonic layer it was taken for was shelved the same day. Nothing here should be read as a hit rate.

Any entry. ZTD says "biased short below X, aiming at Y". It does not say where to enter. The entry layer that was built to answer that question is a separate method with a separate record, and that record is ten failures.

How CAT publishes it

Every ZTD read carries this status in its advisory and links here. A read publishes only when there is a bias to show; a scheduled run that finds nothing publishes nothing. Each read that does publish is tracked to its outcome on the wall — won if price reaches the zero-touch level before the invalidation, lost the other way — so that the record this note says is short can grow in public.