Record why the damage panel prices at the report date

[mspinola]

Why

Raised as a question about the /damage chart in cot-analyzer: is the plot using the most recent price available, and would using it make the chart more accurate? Measured, then decided as no change.

Recorded because the reasoning lived nowhere. "Use fresher data" is a reasonable-sounding suggestion and nothing in the code refuses it out loud, so the next session to ask would have had to re-derive the answer.

What the code does, measured

publish calls add_trigger_distance with no as_of, so the stamp is report_date.max(), and trigger_prices truncates the series before taking spot. add_risk_units reaches the same place by a different route (merge_asof(..., direction="backward")) and publishes sigma_date / sigma_staleness_days so the gap is inspectable rather than assumed.

The gap is not cosmetic. Panel anchored 2026-07-28, store running to 2026-08-04:

symbol	store's latest close	close actually used
GC	4152.60	4038.70
NG	2.68	2.66
ZC	442.25	458.50

What repricing would do

70 (market, side) pairs with a trigger on both dates:

median absolute change	1.742 sigma (p90 5.901, max 65.256)
cross the 1.5 close line	29 of 70, 41%
nearest lookback k changes	27 of 70
gains / loses a trigger entirely	6 / 6

That 41% is the argument against repricing, not for it. A quantity where two fifths of readings cross the threshold defining the quadrant column in five sessions is not a precise measurement being degraded by staleness. It is a noisy one, and refreshing it yields fresher noise.

Four reasons the anchor stands

1. The pool would still be a week old, so vintage skew moves inside a row and lands on trigger_*_pool_agrees, which compares two signs that would then come from different dates.
2. The history would break at its last point. A latest-price week cannot be rebuilt retrospectively without lookahead.
3. Daily publishing stops being idempotent, which is the stated reason it runs daily against weekly data.
4. D is price-dependent through C and I, so repricing the trigger alone splits one row across two vintages.

One expectation that did not survive

nearest_trigger measures reference-bar drift on 6C at 1.68x spot over 120 sessions, which invites the inference that most of the change in distance-to-trigger is bars rolling off rather than price moving.

Across 47 markets over this one week it is the other way round: 0.56x (|spot move| median 2.705%, |F* move| median 1.527%). Different measurements, so the docstring is left as it stands, but the entry states plainly that the reassuring version is not established and should not be quoted as though it were. That is the claim most likely to get repeated from memory.

Reproducer, declared absent

Not in docs/analysis/reproduce.py. Ad-hoc scripts against an unpinned store, and Norgate restates history on every roll, so the table will not reproduce exactly. Recorded at that fidelity deliberately: the decision rests on 41% being large, not on it being 41%.

Also

CLAUDE.md's amendments row listed E1-E3 and had already missed E4. It now carries E4 and E5.

566 passed, 93 skipped, ruff clean.

🤖 Generated with Claude Code