linkedin-studio/docs/research-engine/brief-re-r3e.md

Brief — RE-R3e: brief history + day-over-day diff (R3 slice d)

Slice: RE-R3e (research-engine rung-2, R3 slice (d) in the operator's (a)→(c)→(b)→(d)→(e) sequence — the doc is numbered r3e by creation order, the concept is slice (d)). It closes hull #7 (substrate §1 hull list: "ingen brief-historikk"): the dated morning brief already writes one Markdown file per day (morning-brief/YYYY-MM-DD.md), but the file is prose-only — nothing records which trends a brief showed in a machine-readable form, and no run reads yesterday's brief, so the engine cannot answer the one question a daily motor exists to answer: "what is new since I last looked?" Predecessor: RE-R2b (the dated brief artifact + surfacedIds(ranking) — the exact set a brief shows: brief.ts:305) + RE-R3b (the per-day-idempotent seen-log surfacedCount/lastSurfacedAt, which already records that a trend was surfaced but not with which cohort) + RE-R3c (the autonomous trigger that makes the dated files accumulate day-over-day on their own — the dependency that makes a day-over-day diff a real signal, not a once-in-a-while comparison) + RE-R3d (the temporal overlay — the within-brief recency/saturation class the diff is orthogonal to). Substrate: docs/research-engine-concepts.local.md §1 hull #7 ("ingen brief-historikk") + §B3 ("Dated-digest som flat plain-text-artefakt … diffbar, grep-bar, lenkbar … Senere sesjon laster «gårsdagens brief» trivielt" — the dated file is explicitly designed to be diffed, R3e is the diff B3 anticipated) + §B4 ("append-only seen-log → ikke re-overflate samme sak" — R3e is the per-cohort complement: not "have I seen this ever" but "was this in the PRIOR brief"). The core decisions (operator-confirmed, AskUserQuestion 2026-06-26 — baked):

• SD1 — persist membership in the brief's own frontmatter, NOT a sidecar. Each brief writes a single surfaced: <id-csv> line into its YAML frontmatter — the ids it actually showed (surfacedIds(ranking)). This keeps one self-describing, grep-bar artifact per day (B3), mirrors the existing store: { … } frontmatter idiom, and is hook-safe (the SessionStart extractYaml is ^summary:-anchored and line-scoped — a new surfaced: line cannot perturb it). The diff reads the prior brief's surfaced: line via one pure regex. No sidecar .json; no second artifact. BRIEF_SCHEMA_VERSION bumps 1 → 2 (the frontmatter gained a field — the first bump since R2b; the store's SCHEMA_VERSION stays 4).
• SD2 — added with titles + dropped as a count; brief.ts stays store-free. The diff is the symmetric set difference of today's surfaced ids against the prior brief's: added (in today, not prior — the headline "what's new", rendered with titles resolved from the ranking the brief already holds), carried (in both), dropped (in prior, not today). added is the value; dropped/carried render as a one-line tally (counts). The dropped ids are not resolved against the store for an acted/skipped/aged reason — that would require injecting store records into the render and is the one explicit follow-up (§4). The render needs only the ranking it already has → brief.ts stays pure (no store, no fs). The framing is honest: "ikke vist i dag" (not shown today), never "resolved" (which the count cannot prove). TDD-order (two-phase RED, light-Voyage discipline, inherited from R3c/R3d): Phase A — diffSurfaced, parseSurfacedFrontmatter, selectPriorBriefFile (+ the BriefDiff type) are NEW named exports of the EXISTING brief.ts; under Node16 ESM a missing named import throws at module-load (every brief.test.ts test would error, not assert), so land non-throwing stubs FIRST (diffSurfaced → {priorDate:null,added:[], carried:[],dropped:[]}, parseSurfacedFrontmatter → [], selectPriorBriefFile → null; renderBrief gains an optional diff param it ignores in the stub), then record value-assertion RED against them (the stubs' constant returns fail the diff/parse/select/section/marker assertions — true assertion-RED). Phase B — the CLI two-day diff test is value-RED against the existing brief handler (today it writes no surfaced: line, reads no prior brief, and its --json carries no diff key → the day-2 diff assertions fail). See plan Step 1.

1. Operator decision context (2026-06-26)

The research engine is Tier-1 (operator, 2026-06-23). R1→R3d built the deterministic spine, the trend's life after capture, the autonomy that runs it, and the within-brief temporal overlay: item-schema + triage (R1) → capture bridge (R2a) → dated morning brief + surfacing (R2b) → persisted relevance + composite ranking (R3a) → status lifecycle + seen-log + re-score (R3b) → autonomous trigger + headless entry (R3c) → temporal overlay (first-mover + saturation, R3d). The brief now regenerates itself every morning (R3c) and ranks each trend by its frozen relevance composite, its pillar overlap, and a live first-mover/saturation class (R3d). But every morning's brief is a standalone snapshot: it cannot say "these three are new since yesterday; the two you saw yesterday are gone." The accumulated dated files are a pile of snapshots, not a history with a diff.

R3e closes hull #7 — the brief history + day-over-day diff — which the operator chose as slice (d) of the full-R3 build-out (2026-06-24, "ALLE gjenstående R3-slices … i rekkefølge (a) → (c) → (b) → (d) → (e)"). It is sequenced after R3c for a load-bearing reason: a day-over-day diff is only meaningful when a brief is produced every day on its own — R3c's nightly trigger is what makes "yesterday's brief" reliably exist. (c) makes (d) a real signal; without the daily trigger, "since last brief" could mean "since whenever the operator last happened to ask."

The concrete value — turning a pile of snapshots into a feed. A daily motor's job is to surface the delta: the operator does not want to re-read the full ranked list every morning and diff it in their head — they want the engine to say "3 nye siden i går" at the top of the brief (and on the one-line SessionStart surfacing, for free). R3e makes the dated file a genuine history rung: each brief records what it showed (surfaced: frontmatter), and the next brief reads the most recent prior one and renders "Nytt siden sist." This is the smallest honest step from "a brief is written daily" (R3c) to "the brief tells me what changed" (the point of a feed) — and it is exactly the diffable dated-digest §B3 said the artifact was designed to be.

2. The gap — grounded in code

• The dated brief is prose-only; membership is not machine-readable. cli.ts:341-343 writes morning-brief/${day}.md from renderBrief(ranking); the body embeds each trend's id inside a rendered bullet (brief.ts:242, :251 — `${e.trend.id}`), but there is no structured record of the set a brief showed. surfacedIds(ranking) (brief.ts:305) computes that set and feeds it to markSurfaced (cli.ts:348), but it is never persisted to the artifact — so a later run that wants "what did yesterday's brief show" would have to scrape prose. The seen-log (surfacedCount/lastSurfacedAt, R3b) records that and how many days a trend was surfaced, but not which cohort it appeared with — it cannot reconstruct "yesterday's brief contained {A, B, C}."
• No run reads a prior brief. The only consumer of the dated files is the SessionStart hook's latestMorningBrief (session-start.mjs:60-77), which reads the single newest file's date+summary and surfaces it verbatim. Nothing reads the second-newest to compare. There is no diff, anywhere.
• "New since last" is not derivable from the store alone. The seen-log gives "never surfaced ever" (surfacedCount absent ⇒ a first-ever sighting) — but that is not "new since the last brief": a trend surfaced once three days ago, absent from yesterday's brief, reappearing today is new to yesterday's reader yet has surfacedCount 1 (not 0). Only a per-brief membership record (the prior surfaced: set) answers "was this in the immediately-prior brief," and only it can compute dropped (in the prior cohort, gone today) — which the store cannot express at all. This is precisely the gap §B3's "diffbar … dated-digest" and §1 hull #7 name.
• The brief's own summary cannot signal a delta. briefSummary (brief.ts:204-216) describes today's top match in isolation; the SessionStart surfacing (session-start.mjs:534-536) shows that line verbatim. There is no "N nye siden sist" the operator could see without opening the file — the one number a feed leads with.

3. Scope — what is IN (RE-R3e)

Zero new source/test files (the two tracked slice docs aside). R3e is pure EDITs: the diff lives beside surfacedIds in brief.ts (the module that already owns the brief's pure read logic — surgical, no new module to wire into the gate), its unit tests in brief.test.ts, its CLI wiring (prior-file discovery) in cli.ts/cli.test.ts, plus the wiring docs + gate.

S-history — scripts/trends/src/brief.ts (EDIT) — the pure diff + the persisted membership

• BRIEF_SCHEMA_VERSION bumps 1 → 2 (brief.ts:23) — the frontmatter gained the surfaced: field. (The store's SCHEMA_VERSION is untouched at 4 — R3e adds no store field; the membership lives in the artifact, the diff is derived.)
• export interface BriefDiff { priorDate: string | null; added: string[]; carried: string[]; dropped: string[] } — priorDate = the date of the brief diffed against (null ⇒ no prior brief, i.e. the first ever / a fresh data dir); added/carried/dropped = the three partitions of the set difference, each order-stable (added/carried preserve today's surfacedIds order; dropped preserves the prior set's order).
• export function diffSurfaced(currentIds: string[], priorIds: string[], priorDate: string | null): BriefDiff — pure (no clock/fs/env; both id lists + the prior date are injected by the CLI edge). added = currentIds not in priorIds; carried = currentIds in priorIds; dropped = priorIds not in currentIds. Uses a Set for membership; preserves input order in the output arrays. When priorIds is empty (first brief), added === currentIds and dropped === [].
• export function parseSurfacedFrontmatter(md: string): string[] — pure; extracts the surfaced: value from a brief's full text via a single line-anchored regex (mirrors the hook's extractYaml idiom: /^surfaced: *([^\n]*)/m), splits on ,, trims, drops empties. Returns [] when the line is absent, blank, or malformed (a pre-R3e brief, or a hand-edited file) — degrades to "empty prior," never throws. Real ids are comma-free hex (store.ts:69-72), so the CSV is unambiguous.
• export function selectPriorBriefFile(filenames: string[], today: string): string | null — pure; from a directory listing, returns the lexicographically greatest filename matching ^\d{4}-\d{2}-\d{2}\.md$ whose date is strictly less than today (ISO dates sort lexicographically, so string compare = date compare), else null. Mirrors the hook's latestMorningBrief filter+sort (session-start.mjs:63-66) but excludes today and any future-dated file — so a same-day re-run diffs against the true previous day, not its own just-written file (the byte-determinism guarantee, SC8).
• renderBrief gains an optional diff param (brief.ts:259): renderBrief(ranking: BriefRanking, diff?: BriefDiff). Two additive emissions:
  • Frontmatter surfaced: line — surfaced: ${surfacedIds(ranking).join(",")} inserted before schemaVersion: (always emitted, even for an empty store → surfaced: blank; this is the record the next day's diff reads, independent of whether today had a prior). schemaVersion: now renders 2.
  • A ## 🆕 Nytt siden sist section (the header gains (<prior-date>) when a prior brief exists, per SC9), placed after the intro line and before ## 🎯 Topp-treff (the delta leads, then the full ranked list). Branches (all deterministic):
    • no diff arg / priorDate === null with added → _Første brief — alt nedenfor er nytt._
    • priorDate === null with no added (empty first brief) → _Første brief._
    • priorDate !== null, added non-empty → one bullet per added id, its entry resolved from the ranking (title + matched pillars + date + link + id, reusing the bullet idiom), then a tally line _${carried.length} båret over, ${dropped.length} ikke vist i dag._
    • priorDate !== null, added empty → _Ingenting nytt siden ${priorDate}._ (+ the same tally line)
  • When diff is omitted (a bare renderBrief(ranking) call, e.g. a unit test that does not exercise the diff), it defaults to the empty diff ({priorDate:null,added:[],carried:[],dropped:[]}) → the _Første brief._ section branch (priorDate===null, added empty). The surfaced: frontmatter line is independent of the diff — always surfacedIds(ranking).join(",") (blank only for an empty store), so a non-empty ranking still emits its real surfaced ids. (This keeps existing single-arg call sites compiling and semantically valid.)
• briefSummary gains an optional diff param (brief.ts:204): briefSummary(ranking, diff?). When diff is present, priorDate !== null, and added.length > 0, it appends ${added.length} nye siden sist. to the one-line headline — so the SessionStart hook surfaces the delta for free (it already shows the summary: line verbatim; no hook edit). The marker is suppressed on the first brief (priorDate === null) and when nothing is new (no noise). It carries no double-quote and no newline (the ^summary: *"?([^"\n]*)"? hook-regex invariant, brief.ts:200-203). renderBrief passes its diff through to briefSummary so the frontmatter summary: and the --json summary agree.

S-cli — scripts/trends/src/cli.ts (EDIT) — prior-brief discovery + the diff in --json

• The brief handler discovers the prior brief and computes the diff (between the ranking at cli.ts:339 and the render at :340): readdirSync(outDir) (guarded by existsSync — a first run has no dir) → selectPriorBriefFile(files, day) → if found, readFileSync it and parseSurfacedFrontmatter → build diffSurfaced(surfacedIds(ranking), priorIds, priorDate); on any fs error, degrade to the empty-prior diff (priorDate: null). Pass the diff into renderBrief(ranking, diff). Adds readdirSync to the existing node:fs import (cli.ts:51).
• --json gains a diff object (cli.ts:352): diff: { priorDate, added: added.length, carried: carried.length, dropped: dropped.length } — counts, not id lists (the headless run-daily.sh collapses --json to one cron-log line). The non-JSON console line (cli.ts:355) appends , N nye siden sist when added > 0 && priorDate !== null.
• --no-mark is unchanged in meaning — it still governs only the store seen-log write (cli.ts:347-349). The artifact's surfaced: frontmatter records what the brief showed regardless of --no-mark (it is a property of the rendered brief, not of the store mutation). No new flag.

Wiring (D-default — WIRE, mirrors R3a/R3b/R3c/R3d)

• agents/trend-spotter.md (EDIT, prose-only, minimal): one line — the morning brief now records which trends it showed (frontmatter surfaced:) and renders a day-over-day diff ("Nytt siden sist") against the most recent prior brief — no new capture step; the agent's polling/capture path is unchanged. Domain-general (no vendor/sector token).
• scripts/trends/README.md (EDIT): add a ## Brief history + diff (RE-R3e) section between the R3d temporal- overlay section and ## Tests: the surfaced: frontmatter record, the selectPriorBriefFile prior-discovery (strict < today, same-day re-run determinism), the diffSurfaced partitions, the section + the summary marker, and the BRIEF_SCHEMA_VERSION 1→2 boundary (artifact-only; store SCHEMA_VERSION stays 4).
• scripts/test-runner.sh (EDIT): bump TRENDS_TESTS_FLOOR (live :716, currently 216) to the tests N line reported after the suite runs — recounted live, append `+ RE-R3e: brief +N, cli +N (brief history
  • diff)to the inline breakdown comment. Add **Section 16n** ("Trends Brief History / Diff", RE-R3e) **between Section 16m's closingecho "" and the Section 18 header** (anti-erosion must stay last). Mirror 16m's shape: **unconditional**, deps-absent-safe (grep -qF+ a non-vacuity self-test emitting **one** pass/fail). **6 emitters** (all on tracked source, notsx, all literals **ASCII** — the section header emoji 🆕is **never** grepped; the shell stays ASCII-clean for bash 3.2set -u): (1) self-test; (2) export function diffSurfacedinbrief.ts; (3) parseSurfacedFrontmatterinbrief.ts; (4) the section header literal Nytt siden sistinbrief.ts; (5) selectPriorBriefFileincli.ts(the diff wiring); (6) the frontmatter emitsurfaced: inbrief.ts. **6 unconditional emitters → bump ASSERT_BASELINE_FLOOR117 → exactly 123** (**live:1473**; "live recount" is the safety net; the expected value is the pinned 117 + 6). Insert the 16n clause into the **header-enumeration prose chain (:53-64`)** before "…the assertion-count anti-erosion floor (SC6) in Section 18," and append the R3e (→123) narration to the Section-18 floor-history comment (which ends "= 117").

4. Non-goals — what is OUT (deferred)

• A sidecar .json membership manifest — OUT (SD1). Membership lives in the brief's own frontmatter (surfaced:), keeping one self-describing artifact (B3). No second file per day.
• dropped resolved to an acted/skipped/aged reason — OUT (SD2), and the one explicit follow-up. dropped renders as a count ("N ikke vist i dag"). Labelling why each dropped id left (acted/skipped via status, or aged past freshDays) would require injecting the store records into the render — brief.ts would no longer be store-free. Honest framing for R3e: "ikke vist i dag," never "resolved." A small, clearly-scoped follow-up (CLI resolves dropped ids → {title, status} and passes them to a richer render) if the loop-closing signal proves worth the coupling.
• A browsable history INDEX file (e.g. a rolling history.md of all past briefs) — OUT. The dated files + the surfaced: frontmatter are the history (grep-bar, lenkbar — B3); an index is a presentation nicety, not a capability gap.
• A SessionStart hook change to render the diff — OUT. The "N nye siden sist" marker rides the existing summary: surfacing (session-start.mjs:534-536) — no hook edit, no hook test, no new frontmatter field the hook must learn. (The hook still reads only date+summary.)
• schedule / run-daily.sh changes — OUT. The nightly run calls brief (run-daily.sh:33), which now computes the diff internally → the scheduled brief gets "Nytt siden sist" automatically, with no scheduler edit (no R3c regression surface).
• A new store field / schema bump / store mutation for the diff — OUT. SCHEMA_VERSION stays 4; types.ts/store.ts are untouched. The membership is an artifact property; the diff is derived at the CLI edge.
• "New" defined as first-ever-sighting (surfacedCount === 0) — OUT (rejected as less correct). R3e's "new" is relative to the immediately-prior brief (artifact diff), which also flags a trend re-emerging after a gap — the honest meaning of "siden sist." (surfacedCount stays the R3d saturation input, a different question.)
• Diffing against an arbitrary historical brief (--since <date>) — OUT. R3e diffs against the most recent prior brief only (the "since last" a daily feed needs). An arbitrary baseline is a later nicety.
• A new module / new test file — none. The diff lives in brief.ts; its tests in brief.test.ts. (No pathguard surface — all EDITs.)
• New agent / new command / new reference doc — none. R3e EDITs brief.ts + cli.ts + their tests + one agent (prose) + README + gate. Counts stay 29/19/27. (The SSOT trend-scoring-modes.md is not touched — the diff is not a scoring concern; scope fence.)

5. Boundaries / invariants (must hold)

• TDD iron law (two-phase RED): failing tests land BEFORE implementation. Phase A — the new named exports (diffSurfaced/parseSurfacedFrontmatter/selectPriorBriefFile/BriefDiff) need non-throwing stubs first (Node16 ESM throws a missing named import at module-load), THEN value-assertion RED against the constant stubs. Phase B — the CLI two-day diff test is value-RED against the existing handler (no surfaced: write, no prior read, no diff in --json today). The plan does not claim a single "everything fails before any code" run.
• brief.ts stays pure (no clock, no fs, no env, no AI): diffSurfaced/parseSurfacedFrontmatter/ selectPriorBriefFile all take strings/arrays and return values — the directory read + file read live in cli.ts (the edge), exactly like today/pillars are injected. The module's "No fs, no clock, no AI, no network" header claim is preserved.
• Determinism of the brief: given (store, pillars, today, freshDays, firstMoverDays, saturationAt, diff) the rendered .md is byte-identical (the diff is now an injected input, like today). Critically, a same-day re-run is byte-identical: selectPriorBriefFile excludes ${today}.md (strict <), so the re-run diffs against the same previous day's brief and re-writes the same surfaced: line (R3c SC7 preserved).
• surfaced: records the shown set, --no-mark-independent: the frontmatter line is surfacedIds(ranking) joined — what the brief showed — regardless of whether the store seen-log was written (--no-mark governs the store mutation only). The artifact is always self-consistent.
• Frozen composite + temporal overlay untouched (R3a + R3d): R3e adds no cmp key and changes no ranking — rankForBrief is unchanged. The diff is a post-ranking, render-time layer over the same surfaced set. score.composite is never read for mutation; the R3d temporal overlay is orthogonal (it orders within the brief; the diff compares across briefs).
• Schema boundary: BRIEF_SCHEMA_VERSION bumps 1 → 2 (the artifact's frontmatter gained surfaced:); the store's SCHEMA_VERSION stays 4; types.ts/store.ts/score.ts/item.ts/schedule.ts/ run-daily.sh are untouched.
• Hook unaffected: the SessionStart surfacing reads date+summary only (session-start.mjs:60-77). The new surfaced: frontmatter line is ^surfaced:-keyed (the ^summary:-anchored, line-scoped extractYaml cannot match it), and the summary: marker carries no "/\n — so the regex still captures the summary whole. R3e touches neither the hook nor the field set the hook reads; the hook suite must still pass untouched (regression sanity; R3e adds no hook test).
• ASCII-only gate literals: scripts/test-runner.sh must stay ASCII (a multibyte char crashes bash 3.2 under set -u). The Section-16n sentinels grep the ASCII literals (export function diffSurfaced, parseSurfacedFrontmatter, Nytt siden sist, selectPriorBriefFile, surfaced: ) — never the 🆕 emoji (which lives only in brief.ts source + rendered output, asserted by the TS tests, not by the shell gate).
• Domain-general: no hard-coded user/repo path, no vendor/sector token in any edit. The section header + Norwegian copy (Nytt siden sist, båret over, ikke vist i dag, nye siden sist, Første brief) are domain-general UI copy (the brief's existing language); pillars/topics remain config. Section 17 de-niche stays green.
• Pathguard: R3e adds no new files — every change is an EDIT of an existing file (write-allowed). (No .mjs-under-hooks/scripts/ surface, no new scripts/ file.)
• Counts (refs/agents/commands 27/19/29) unchanged — recounted live at land, never pinned/guessed.

6. Success criteria (testable)

• SC1 (diffSurfaced — partitions + order + empty prior) — diffSurfaced(["a","b","c"], ["b","c","d"], "2026- 06-25") → {priorDate:"2026-06-25", added:["a"], carried:["b","c"], dropped:["d"]} (added/carried in current order, dropped in prior order). diffSurfaced(["a","b"], [], null) → {priorDate:null, added:["a","b"], carried:[], dropped:[]} (empty prior ⇒ everything added). Pure: same inputs → same output. The three partitions are mutually disjoint — Set membership is binary (an id is in priorIds or not), so each id lands in exactly one of added/carried and dropped is disjoint from both; within each list, order and any duplicates mirror the input (surfacedIds yields distinct ids in production, so within-list dups never arise — the cross-partition exclusivity is the real invariant, not within-list dedup).
• SC2 (parseSurfacedFrontmatter — read + degrade) — parses surfaced: 1a2b,3c4d,5e6f (in a full frontmatter block) → ["1a2b","3c4d","5e6f"]; a blank surfaced: → []; an absent surfaced: line (a pre-R3e brief) → []; whitespace around ids is trimmed; the summary:/store:/date: lines are not mismatched (line-anchored). Never throws on malformed input.
• SC3 (selectPriorBriefFile — strict-prior selection) — from ["2026-06-24.md","2026-06-25.md","2026-06-26.md", "README.md","2026-06-30.md"] with today="2026-06-26" → "2026-06-25.md" (greatest < today; excludes today 2026-06-26.md and the future 2026-06-30.md; ignores the non-dated README.md). Empty list, or no file < today, → null.
• SC4 (frontmatter surfaced: line + round-trip) — renderBrief(ranking, diff) emits exactly one ^surfaced: <csv>$ line, equal to surfacedIds(ranking).join(","), positioned before schemaVersion: 2; an empty-store brief emits surfaced: (blank); parseSurfacedFrontmatter(renderBrief(r, d)) round-trips to surfacedIds(r). schemaVersion: renders 2.
• SC5 (Nytt siden sist section — all four branches) — the rendered body contains ## 🆕 Nytt siden sist; with priorDate:null + added → Første brief — alt nedenfor er nytt; with priorDate:null + no added (empty store) → Første brief.; with a prior + added → one bullet per added entry (its title present, resolved from the ranking) + N båret over, M ikke vist i dag; with a prior + no added → Ingenting nytt siden <date>
  • the tally. The section precedes ## 🎯 Topp-treff.
• SC6 (summary delta marker) — briefSummary(ranking, diff) with priorDate !== null and added.length > 0 ends with ${added.length} nye siden sist.; with priorDate:null (first brief) or added.length === 0, the marker is absent (and briefSummary(ranking) with no diff === briefSummary(ranking, emptyDiff) — no marker, so the existing frontmatter summary === briefSummary(r) test stays green). The summary contains no " and no \n.
• SC7 (schema boundary) — BRIEF_SCHEMA_VERSION === 2; SCHEMA_VERSION === 4; types.ts untouched; a brief run does not change any record's score.composite (only surfacedCount/lastSurfacedAt move — the existing R3b behaviour, since rankForBrief/markSurfaced are unchanged).
• SC8 (determinism, incl. same-day re-run) — two brief renders with the same (store, pillars, today, opts, diff) → byte-identical .md. End-to-end via the CLI: running brief twice on the same day (the second after the first wrote ${day}.md) → byte-identical files, because selectPriorBriefFile excludes the same-day file and picks the same prior day.
• SC9 (CLI diff wiring — two-day sequence) — write a day-1 brief (records surfaced: for its cohort), then a day-2 brief over a store with one new trend: the day-2 .md ## 🆕 Nytt siden sist (<day-1>) section lists the new trend, the day-2 --json carries diff: { priorDate:<day-1>, added:≥1, carried:…, dropped:… }, and the console line appends N nye siden sist. A first run (empty dir) → diff.priorDate === null. A custom --out isolates discovery to that dir (the diff reads prior briefs only from outDir).
• SC10 (gate + wiring + de-niche) — bash scripts/test-runner.sh → FAIL=0: trends suite green at the bumped TRENDS_TESTS_FLOOR; new Section 16n green (the six ASCII sentinels + non-vacuity self-test); ASSERT_BASELINE_FLOOR = 123 (117 + 6); Section 17 de-niche green; counts 29/19/27; the hook suite still green untouched (node --test hooks/scripts/__tests__/*.test.mjs).

7. Verification

Deterministic: bash scripts/test-runner.sh → FAIL=0; trends suite ≥ new floor; Section 16n self-test + greps pass; ASSERT_BASELINE_FLOOR = 123; Section 17 de-niche green; ref/agent/command counts unchanged. Regression sanity: node --test hooks/scripts/__tests__/*.test.mjs → still green untouched (R3e touches no hook). The R3c suite (schedule.test.ts/run-daily.test.ts) still green untouched (schedule.ts/run-daily.sh not edited). Behavioural (manual):

1. D=/tmp/r3e-mb-$$; S=/tmp/r3e-$$.json (unique dir, no rm). Seed a store with two on-pillar trends and run a day-1 brief: node --import tsx src/cli.ts brief --pillars ai,gov --out "$D" --store "$S" → inspect the .md: the frontmatter carries surfaced: <ids> and schemaVersion: 2; the ## 🆕 Nytt siden sist section says Første brief — alt nedenfor er nytt.
2. capture a third on-pillar trend into the same store, then run a day-2 brief with a later today (seed via a second dated file is not possible — use --out "$D" so day-1's ${day}.md is the prior; on a real next-day run the date advances): inspect the new .md → ## 🆕 Nytt siden sist (<prior-date>) lists the new trend with its title, then N båret over, M ikke vist i dag; the --json shows diff.added ≥ 1.
3. Re-run the same-day brief → the written .md is byte-identical (diff against the same prior file; surfaced: re-written identically) — cmp the two files.
4. Confirm the seeded records' score.composite values are unchanged after the briefs (only surfacedCount/ lastSurfacedAt advance) — the diff never mutates the store ranking.
5. Confirm a pre-R3e brief (no surfaced: line) as the prior → parseSurfacedFrontmatter returns [] → every trend reads as added (graceful degrade, no crash).

8. Open questions for the go-gate

Two architectural decisions are CONFIRMED (operator, AskUserQuestion 2026-06-26): SD1 persist membership in the brief's frontmatter (no sidecar); SD2 added with titles + dropped as a count (brief.ts store-free). Residual decisions, all baked to the recommended default — confirm or redirect with "Go":

• D1 — "new" = relative to the immediately-prior brief (artifact diff), not first-ever (surfacedCount 0)? YES (rec). It is the honest meaning of "siden sist" and catches re-emergence; it is also what unlocks dropped. Re-open only to redefine "new" as first-ever.
• D2 — diff against the most recent prior brief only (no --since baseline)? YES (rec). The "since last" a daily feed needs. Re-open only to add an arbitrary historical baseline.
• D3 — ## 🆕 Nytt siden sist placed before ## 🎯 Topp-treff (delta leads)? YES (rec). The one thing a feed leads with. Re-open only to place it after the ranked list (appendix) or omit the header on a first brief.
• D4 — summary marker N nye siden sist. (suppressed on first brief / when nothing new)? YES (rec). The delta the SessionStart hook surfaces for free, with no hook edit. Drop only to keep the summary minimal.
• D5 — surfaced: frontmatter always emitted (incl. --no-mark, incl. empty store → blank)? YES (rec). It is the record the next diff reads; gating it on --no-mark or non-empty would silently break tomorrow's diff. Re-open only to gate it.
• D6 — dropped/carried render as a one-line count (no titles); dropped framed "ikke vist i dag"? YES (rec, = SD2). Keeps brief.ts store-free; honest (a count cannot prove "resolved"). Re-open only to pull the reason-labeled follow-up into R3e now.
• D7 — BRIEF_SCHEMA_VERSION 1 → 2; store SCHEMA_VERSION stays 4? YES (rec). The frontmatter gained a field (the first artifact-schema change since R2b); no store field. Re-open only to add a store field instead.
• D8 — README gets the R3e section; trend-spotter.md gets one prose line; the SSOT is NOT touched? YES (rec). The diff is not a scoring concern — touching trend-scoring-modes.md would be scope creep. Re-open only to add an SSOT note.
• D9 — commit split? Docs commit first, then one code commit (rec) — the diff (helpers + render + CLI wiring + gate) is one coherent feature. Re-open only for a helpers-then-wiring split.

9. Light-Voyage review — folded

Three Opus reviewers ran COLD on this brief + the plan against live scripts/trends/ code (scope-guardian, brief-reviewer, plan-critic — the R3c/R3d discipline). Verdicts: scope-guardian MIXED · brief-reviewer PROCEED_WITH_RISKS · plan-critic REWORK (0.88). They converged on 2 MAJOR (both re-verified against live code before folding) + 4 MINOR. Every line-cite, the floors, the regex/lex/hook safety, and the §3 scope fence were confirmed correct by all three and left untouched.

MAJOR-1 — a hard schema literal breaks at the 1→2 bump (the §6/Step-1 "no existing assertion breaks" scoping missed it). tests/brief.test.ts:574 is assert.equal(BRIEF_SCHEMA_VERSION, 1) — a hard literal, not the constant-tracking RegExp at :163 (new RegExp("\\nschemaVersion: " + BRIEF_SCHEMA_VERSION + "\\n") auto-tracks). It lives in the rankForBrief — no schema/score mutation block (:568-577), outside the frontmatter tests §6/plan-Step-1 enumerated, so the "every frontmatter assertion auto-tracks" claim did not cover it. Resolution: the GREEN schema bump (BRIEF_SCHEMA_VERSION = 2) must also flip :574 → , 2) in the same step (plan Step 3). :575 (assert.equal(SCHEMA_VERSION, 4), the store schema) is untouched. Folded into plan Step 3 + R1 + Step 1's enumeration. (Swept live: :574 is the only hard BRIEF_SCHEMA_VERSION literal in the suite; cli.test.ts:102 asserts the store persisted.schemaVersion === SCHEMA_VERSION (4) — unrelated, stays green.)

MAJOR-2 — the --json summary would diverge from the file frontmatter on day-2 (the "one source" invariant). The file's frontmatter summary: is built inside renderBrief (brief.ts:265), which Step 3 routes through briefSummary(ranking, diff) → on day-2 it carries the N nye siden sist. marker. But the CLI's --json summary field reads a separate const summary = briefSummary(ranking) (cli.ts:350, comment // SAME source the frontmatter carries) that Step 4 left unthreaded → no marker. cli.test.ts:268 asserts fileFrontmatter.summary === json.summary ("one source") → would break on day-2. Resolution: Step 4 changes cli.ts:350 to briefSummary(ranking, diff) (the diff is in scope — Step 4 computes it between the ranking at :339 and the render at :340). Safe on day-1: priorDate === null ⇒ the marker is suppressed ⇒ byte-identical to today's string. Folded into plan Step 4 + the files-touched table.

MINOR (folded):

• M1 — BriefDiff is an interface (type-only export). Anywhere it is referenced as a type (tests or cli.ts), import it with import type { BriefDiff }, never a value import — under Node16 ESM + tsx a type-only export is stripped from the emitted JS, so a value-import named binding fails to resolve at module-load (the same Phase-A hazard as a missing named import). The plan's Step-4 code does not annotate : BriefDiff (it infers from diffSurfaced's return) and the SC tests pass object literals — so in practice no BriefDiff import is needed; the rule is the guardrail if one is added. Folded into plan Step 2/4.
• M2 — SC9's prior brief is a real renamed brief, not a hand-fixture (rename-real-write). Replace the "pre-write a <prior-date>.md fixture carrying a surfaced: line" mechanism (plan Step 4 note / Phase B / Step 7) with: run brief once (writes ${today}.md with a genuine surfaced: line), mv ${today}.md → 2026-06-20.md (a fixed past date) in the same --out, then run brief again. This (a) closes the write→read loop clock-free (no today() manipulation), and (b) guarantees the prior surfaced: ids are real store ids (they came from a real run), so carried is non-trivial and added is exactly the newly-captured trend — a hand-fixture risks an id mismatch that makes everything read as added/dropped (a weaker, possibly-vacuous test). Folded into plan Step 4 note, Phase B, Step 7, SC9.
• M3 — SC1 "repeated id" wording. diffSurfaced uses currentIds.filter(...)/priorIds.filter(...), which preserve within-list duplicates — so "not double-counted" is wrong as within-list dedup. What the Set membership actually guarantees is cross-partition disjointness (added/carried/dropped are mutually exclusive). Reworded in §6 SC1 + plan Step 1/verification. (Production surfacedIds yields distinct ids, so within-list dups never arise.)
• M4 — §3 "empty surfaced:" self-contradiction. The bare renderBrief(ranking) default-diff prose said it yields "an empty surfaced: reflecting the ranking" — contradictory: surfaced: is always surfacedIds(ranking).join(",") (non-empty for a non-empty ranking; blank only for an empty store), independent of the diff; the default empty diff only drives the _Første brief._ section branch. Reworded in §3 (the renderBrief default-diff bullet).

Confirmed correct by all three — left untouched: every line-cite (brief.ts:23/204/265/305, cli.ts:339/340/350/352, session-start.mjs:60-77/534-536); the floors (TRENDS_TESTS_FLOOR 216 @ live :716, ASSERT_BASELINE_FLOOR 117 @ live :1473 → 123 = 117 + 6 unconditional 16n emitters); Section 16m is the last trends section (16n sits between its echo "" and Section 18); the 6 ASCII sentinels are non-vacuous; the /^surfaced: *([^\n]*)/m regex, the ISO-lex compare, the ^surfaced: ≠ ^summary: hook-safety, the --json shape, and the same-day-determinism strict <.