Method & limitations
how a claim gets on this page, and what it still cannot do
What you are shown, and what is folded. The daily crawl admits more than
anybody reads, so every post is scored 1–5 on six axes derived from what the
pipeline already recorded, and the default view holds back anything that does not
clear a floor of 2.50 across the four axes that describe the item
— topic relevancy, impact, local/foreign context and source. The two that
describe our record of it, completeness and readiness, are reported and
gate nothing: withholding a rule change because our own archive snapshot failed
would be the worst possible reading of a quality score.
Nothing is deleted. Held-back posts keep their card and their evidence, and
the “Shared only” switch above brings them straight back with a count.
ACT items, Singapore items and the mandate beats are never held back whatever they
score. Read the full assessment — every post, every
axis, and the reason behind every cell.
What a period amounted to. Under every day, week, month and year heading is a
short summary: how many items need a decision (named, with what to do about them), how
many are Singapore’s, which way the numbers moved against the period before, which
beat dominated and what that class of item means, and which beats had nothing at all.
It does not repeat the item count from the heading, and it does not total the readiness
bands: a band describes one headline’s wording and belongs on that card, where the
wording is visible beside it.
None of it is written. Every line is arithmetic over posts the pipeline already
scored, or a sentence a person wrote once in beat.json for that beat and
which is attributed as such. A day gets no direction line and no gap line, because
day-to-day movement is a publishing schedule rather than a trend. And no period is ever
compared against one that began before this scan started watching on
28 July 2026 — those months were not quiet, they were
not yet being read, and comparing against them would manufacture a trend out of our
own absence.
Every claim carries a receipt. Each post records the URL as fetched, the HTTP status, a SHA-256 of
the page body at fetch time, an Internet Archive snapshot, and a verbatim excerpt that is
mechanically checked to appear in the fetched text before the post is published. An item whose excerpt
cannot be verified and whose publisher supplied no summary is dropped rather than published — every
rejection, with its reason, is written to the run manifest under audit/.
What is generated and what is quoted. Summaries on scanned posts are the publisher's own feed
description, quoted rather than paraphrased. Excerpts are selected mechanically from the source page, not
written. The only generated text on this page is the daily throughline paragraph above, which is composed
from the admitted headlines alone and is labelled as editorial.
Readiness (TRL), and where it comes from. Every post in a category the ladder applies to
carries a Ready line: how finished the thing is, as a technology readiness band, read
from the words the publisher printed and never from anything written here. The words it was read
from are shown with it. Bands are ordered lowest-maturity-first and the first match wins, so
“full-scale field trials” is a pilot and not a deployment — trade press
overstates maturity and the band does not inherit the overstatement. Where the publisher’s
words do not say, the line says not stated, which is a gap in the source and not a level
of zero. Categories that have no readiness level at all — a funding round, a rule change — carry
no line, because TRL is a property of a technology under development and inventing one for a
ministerial speech would be noise. The same ladder runs the
Solutions registry, under one extra rule that page states in full: a
vendor’s own site may only band a tool down, never up.
Two readiness bands on one card. Where an item names a tool from the
Solutions registry, that tool’s own band is shown beside
its name. It answers a different question to the band above it. The item’s band is read
from this headline — how finished the thing this article is about sounds. The
tool’s band is a standing reading of that product, built from independent
coverage across the whole archive, and it was true before this article and stays true
after it. They routinely differ, and that is the useful part: an announcement naming a
tool already in use is a mature product shipping another feature, while the same
announcement naming a tool at pilot is a different thing to plan against. Naming a tool
remains a word match — it says the item mentions it, nothing more — and the
band beside it is never a reading of the article. It carries the date the registry was
last checked, because that file is re-checked on its own schedule and this page holds a
snapshot of it.
Where a card says nothing. A handful of publishers — journal listings, mostly — print
neither a summary nor any text that can be quoted. Those cards simply start lower down instead of
carrying a line explaining the absence: the headline, the link and the
title-only grade are the whole claim, and the fetch receipt is still
in Evidence. Nothing is hidden; it is just not announced twice.
On times of day. Most trade and vendor sources publish a date but not a clock time. Where a
time is not printed by the source, this scan says so explicitly rather than inventing one. The
scanned timestamp is always exact, because it is our own.
Evidence grades.
page the source page was fetched and a verbatim excerpt was verified against
its body.
search the page body was unavailable (blocked, empty, or paywalled), so the
post carries only the publisher's own feed summary and makes no further claim.
unverified no evidence at all — these are dropped, not published; the grade
survives only on pre-audit entries.
A legacy badge marks a summary written before this rule, where no verbatim
excerpt could be confirmed on re-fetch. Treat those as editorial, not as quoted source text.
The read. Every post carries a one-line triage aimed at whoever owns digital delivery:
ACT
something with a deadline or compliance consequence that touches your submission path;
WATCH
direction of travel worth tracking, no action yet;
NOTE
context only. It is produced by ordered rules in scan/beat.json under
signals, matching on category, scope and whole-word beat terms — not by a
language model. The tag is the most opinionated thing on this page, so it is also the one you can
trace to a rule and argue with. It is re-derived for every post on every run, so tuning the rules
retags the whole archive rather than only new items.