Wednesday, September 2, 2026 · 2h 41m · 17 stages
from bug to shell.

How the brief is made.

On Wednesday, September 2, 2026, 1,219 items went in and 9 were published. Seventeen stages did the narrowing — eight of them deterministic code, nine of them gpt-5.6-sol and claude-opus-5 working with live web search under a strict output schema. Hover over a stage to see what it is for. Every number is the pipeline's own record of what it did.

What got through
1,219items in the snapshot
24evidence packets
20distinct findings
9published
read by eleven specialists
merged and ranked
reviewed, rewritten, gated
Stage by stage
Gather
Research
Judge
Write
Publish
deterministic code
gpt-5.6-sol
claude-opus-5
01
Pre-flight6 probes
code1s
6 probes, all passed
Checks the network, the model logins and the disk before anything is spent.
02
Collect958 events
code18m
958 events from 295 feeds
Polls every tracked feed — vendor advisories, CVE registries, research blogs, exploit repositories — and files each new entry in the inbox.
03
Sweep20 leads
gpt-5.6-sol13m
20 leads taken in from 32 searches, 7 observations
One wide, unconstrained search of the last few days, so the day is not limited to sources already on the list.
04
Classify publishers10 publishers ruled on
gpt-5.6-sol3m
10 of 10 new publishers ruled on; 8 publish original research
Rules on publishers the pipeline has not seen before: original research, a vendor, or someone re-reporting others' work.
05
Snapshot1,219 items
code0s
1219 items frozen as the run's input
Freezes the inbox. Everything after this reads the same fixed input, so a run can be replayed exactly.
06
Brief11 briefs
code0s
11 briefs written, one per surface
Writes each specialist a brief: its share of the snapshot, open leads, and what the sweep noticed.
07
Specialists24 packets
gpt-5.6-sol47m
24 evidence packets from 11 surfaces; 1 could not finish
Eleven agents, one per attack surface. Each reads its share with live web search and writes up anything that changes what an attacker can do.
08
Quiet check18 reasons
code0s
18 reasons this was not a quiet day
Decides whether anything actually changed. If nothing did, the day says so and stops.
09
Dedup & rank20 findings
gpt-5.6-sol6m
24 packets → 20 distinct findings, ranked by impact at xhigh
Merges reports of the same finding from different surfaces, then ranks the day by impact.
10
Patch reality5 fixes read
gpt-5.6-sol27m
5 fixes read line by line; 2 do not remove the capability
Reads the fix for each leading finding, line by line, and answers whether it actually removes the capability.
11
Composition0 compositions
code0s
0 candidate chains judged, 0 confirmed
Looks for chains: findings that combine into something neither is on its own.
12
Synthesis9 placed
gpt-5.6-sol12m
3 priority, 6 secondary, 46 claims written
Writes the brief — which findings lead, which are secondary, and a claim behind every sentence.
13
Review47 verdicts
claude-opus-516m
47 verdicts on the evidence; 5 changed the page
A second model reads the evidence and the claims, never the prose, and strikes anything the evidence does not support.
14
Rewrite6 repaired
gpt-5.6-sol61s
6 blocks repaired after review
Repairs prose the reviewer left dangling, without restoring anything that was struck.
15
Claim check6 checked
claude-opus-516s
6 rewritten blocks checked against its claims
Checks that every rewritten sentence still rests on a surviving claim.
16
Gate0 hard failures
code0s
cleared to publish, with 14 notes for the editor
Deterministic rules, no model: every published item needs an original, verified source; every assertion a claim; no surface may read as checked when it was not.
17
Publish7 pages
code0s
7 pages rendered
Renders the page, records the day, and opens a pull request. A human merges.
Who did the work
gpt-5.6-sol
7 stages · 1h 49m
deterministic code
8 stages · 18m
claude-opus-5
2 stages · 16m

Model stages run with live web search and a strict output schema; each answer that cites a source must cite one the pipeline holds. Deterministic stages — the snapshot, the gate, the publish — make no calls at all.

Running total
2runs
2,206items in the snapshot
30evidence packets
26distinct findings
14published
BackToday's brief All editionsArchive
2 runs recorded