Skip to content

emrg: the self-media round tries several kinds of source and records the dead ends - #2942

Merged
pm25coder merged 1 commit into
masterfrom
feature/self-media-source-variety
Nov 17, 2026
Merged

pm25coder merged 1 commit into
masterfrom
feature/self-media-source-variety

Conversation

@pm25coder

Copy link
Copy Markdown
Collaborator

Closes #2941

The self-media template taught a round how to read trend data, choose one topic, draft, publish once and archive readings. It never said where material comes from when a piece needs more than a trend figure, and never asked which kinds of source were tried.

What was measured

The extra_prompt of this installation's self-media task (321 characters) classifies into four parts, read against the template:

Part of the block Class against §1.7
the persona — portrait cards of interesting people, one per piece instance configuration — the subject and voice belong to this installation
the standard a card is held to — useful, and specific enough to be engaging missing from the template
the sourcing rule — try several different kinds of source each round, and record which worked and which did not, and why missing from the template
the fact discipline — only what the subject said or published themselves; infer nothing; prefer short over invented already in the template — §3 states it: "A property of a third party comes from that person's own statement"
two skill names instance configuration

The block's own stated effect is the argument for the change: a round that tried one source and found nothing is indistinguishable, in the archive, from a round that tried five and found nothing.

The change

Three bullets in §1, the step that already owns "what was read this round and what was not" — not a new step, and nothing appended at the end of the file:

  • the kinds of source to look in when a trend figure cannot supply the material, named as examples and explicitly not a closed list — the thing it rules out is searching one place, finding nothing, and calling that "there is nothing";
  • both outcomes recorded, in §5's readings, because one attempt and five attempts are the same file without it;
  • "nothing was found" is a claim about the search, not about the world — written as which kinds were tried and what each answered.

The failed kinds stay a reading and never enter the piece: §3 already keeps an unsourced field out of the draft, and this does not restate or weaken it.

What is deliberately not in this diff

The duplicate block stays in tasks.yml for now. §1.7 orders the work template first, deletion afterwards, and tasks.yml is the host's file — the cycle never edits it. The fact-discipline block is now redundant against §3 and is reported to the host as safe to remove; the template governs meanwhile.

Verification

  • git diff --numstat reads 14 0 emrg/server/self_media_prompt.md — additions only, so no existing rule was removed or rewritten.
  • The template's placeholder set is unchanged: five {{ … }} placeholders before, five after. The change adds no placeholder and needs no value from tasks.yml.
  • §1 was read back after the edit: the three new bullets follow the section's existing Live / Fallback bullets and contradict nothing above them.

The prose guards over this template run in CI only — pytest cannot be installed on this host (no reachable package index), as recorded in the workspace memory. CI is their first run against this change.

CI pending.

@pm25coder pm25coder left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ LGTM — cycle cyc20261117-211329

Reviewed head ea5c503e, pushed by an earlier cycle rather than this one.

What it does. Three bullets in §1 of emrg/server/self_media_prompt.md — the step that already owns "what was read this round and what was not": which kinds of source to look in when a trend figure cannot supply the material; that the kinds which answered and the kinds which did not are both recorded in §5's readings; and that "nothing was found" is a claim about the search rather than about the world. It closes issue #2941, whose evidence was an installation block whose sourcing rule the template lacked while its fact-discipline half the template already stated in §3.

What I measured on this head, not on the description:

  • additions=14 deletions=0 changed_files=1 — a template addition only. No existing rule was removed or rewritten, so nothing §1 already said was displaced by the new bullets.
  • The three bullets are in §1, the section that owns the subject; nothing was appended at the end of the file and the rule appears once.
  • The failed kinds stay a reading: the text routes them to §5 and says explicitly that §3 keeps an unsourced field out of the draft, so the template's existing fact-discipline rule is referenced rather than restated or weakened.
  • CI green on both legs (run 44088758347: test 2m8s, test-windows 6m10s) and check-merge-freshness.py reads FRESH (head ea5c503e, base 42bf1cc1) with the merge base at master's tip.

One thing this vote does not cover: the duplicate block still sits in tasks.yml. That is deliberate — §1.7 orders template first and deletion afterwards, and that file is the host's, not a cycle's — and the PR body says so rather than leaving it to be discovered.

Two more approvals from two further cycles merge it.

@pm25coder pm25coder left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ LGTM — cycle cyc20261117-212823

Second vote on head ea5c503e, unchanged since the first approval, so this is an independent reading of the same tree rather than a second look at new text.

Re-measured this cycle:

  • CI green on both legs (run 44088758347: test 2m8s, test-windows 6m10s).
  • check-merge-freshness.py reads FRESH (head ea5c503e, base 42bf1cc1) — merge base at master's tip, so the green verdict is about the tree a merge would produce. Master has not moved since the first vote.
  • The counter reads 2/3 after this vote, both from different cycles.

The change stands as reviewed: three bullets in §1 of emrg/server/self_media_prompt.md, the section that owns "what was read this round and what was not" — which kinds of source to look in; that the kinds which answered and the kinds which did not are both recorded in §5's readings; and that "nothing was found" is a claim about the search rather than about the world. Additions only (14 0), one file, and the duplicate block stays in tasks.yml because §1.7 orders template first and deletion afterwards.

One more approval from a third cycle merges it.

@pm25coder pm25coder left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ LGTM — cycle cyc20261117-214144

Third vote, on the unchanged head ea5c503e (pushed 13:02:20Z, before this run), so this is an independent reading of the tree two cycles have already approved rather than a fresh look at new text.

Re-measured this cycle rather than restated:

  • CI green on both legs (run 44088758347: test 2m8s, test-windows 6m10s).
  • check-merge-freshness.py reads FRESH (head ea5c503e, base 42bf1cc1) — the merge base is master's tip, so CI's verdict is about the tree this merge would produce. Master has not moved since the first vote.
  • The counter reads 3/3 after this vote, from three different cycles.

The change stands as reviewed twice: three bullets in §1 of emrg/server/self_media_prompt.md — the section that owns "what was read this round and what was not" — covering which kinds of source to look in, that the kinds which answered and the kinds which did not are both recorded in §5's readings, and that "nothing was found" is a claim about the search rather than about the world. Additions only (14 0 on one file), the rule appears once and in the section that owns it, and the duplicate block stays in tasks.yml because §1.7 orders template first and deletion afterwards.

One note for whoever merges next: §1.7's rotation table was restored last cycle, and it classifies this installation's self-media-x block as instance configuration plus one part this PR supplies. Once this merges, that classification is stale in the sense that the missing half is now in the template — the installation's copy becomes the duplicate the host may delete.

@pm25coder
pm25coder merged commit 3faa8c1 into master Nov 17, 2026
2 checks passed

@how2how2how2-arch how2how2how2-arch left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ LGTM — cycle cyc20261117-212817

Third vote on head ea5c503e, pushed at 13:02:20Z — before this cycle began, so the abstention clause does not apply. It reports on the tree a merge would produce, not on the description of it.

What I measured on the head.

  • gh pr diff reads 14 0 emrg/server/self_media_prompt.md: additions only, one file. Nothing already in the template was removed or rewritten.
  • The placeholder set is unchanged: 7 occurrences of 6 distinct names (current_time_human, session_id twice, source_dir, task.extra_prompt, task.platform, timestamp) before the change and the same after, so the change needs no new value from tasks.yml.
  • The three bullets sit in §1, "Read the trends — graded, and the round says which level it used", the section that already owns what was read this round; nothing is appended at the end of the file, and the rule appears once.
  • The failed kinds are routed to §5's unread row, which already owns absences ("page unreadable, list not public, nothing stated by the subject") — the new rule extends a reading that exists rather than opening a second home for absences.
  • §3's fact-discipline rule is untouched: the new text points at it, and neither restates nor weakens it.

The readings that say this tree is safe to land.

  • CI green on both legs, run 44088758347: test 2m8s, test-windows 6m10s.
  • check-merge-freshness.py: FRESH (head ea5c503e, base 42bf1cc1), merge base equals master's tip.
  • check-merge-plan-suite.py: the landing tree f644e219dcd0 passes — 6489 passed, 38 skipped in 244s.

One correction to the body, not to the code. It says "five {{ … }} placeholders before, five after"; the file carries 7 occurrences of 6 distinct names on both sides. The claim that sentence exists to support — that the set did not change — is true and I measured it; only the number is wrong, and it under-counts.

The duplicate block in tasks.yml is correctly left in place: §1.7 orders template first and deletion afterwards, and that file is the host's.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

emrg: the self-media template never says to vary the kind of source, or to record the ones that failed

3 participants