linkedin-studio/docs/second-brain/brief-sb-s3d.md

Brief — SB-S3d: the operations centre (operations.md becomes a READ tributary)

Slice: SB-S3d (the ops-centre sub-slice of SB-S3). Splits the old "S3d hygiene+ops" charter: S3d = ops centre only; the hygiene + triple-post reconciliation (dead content-history.md retirement + post-tracking↔published↔analytics) moves to a new SB-S3e (the genuine LAST S3 slice). Status: DRAFT (brief + plan written; light-Voyage hardening + go-before-code gate pending). Predecessors: SB-S0 (scaffold seeds the bare operations.md) · SB-S3a (the profile.md reader pattern this mirrors) · SB-S3b/c (supersede + cross-silo graph — independent of this).

1. Operator decision (2026-06-23)

The S3d charter bundled three deliverables — (a) flesh out operations.md, (b) retire the dead content-history.md, (c) triple-post reconciliation. Grounding showed a clean fault line: (a) is additive / new / self-contained (touches only scripts/brain/ + one reader agent, never the post silos), while (b)+(c) are coupled (the content-history back-fill question is the reconciliation question) and both touch the loose regex-mutated state silo S3c deliberately avoided. The operator chose Split — ops centre first: S3d ships (a) only; (b)+(c) become SB-S3e, the focused finale. This keeps S3d low-risk and isolates the silo surgery.

2. Why the ops centre, and why now

architecture.md:9 frames the second brain as "memory AND an operations centre." S0→S3c delivered the memory half (profile, ingest, consolidation loop, cross-silo graph). The operations centre is the genuinely-new half (architecture.md:44: "the operations/plans centre is genuinely new"), and today it is scaffold-only: operationsSeed() (scaffold.ts:52-64) writes three bare headers — ## Who I am now (anchor), ## Plans, ## Ideas — that no agent or command reads. Verified: 0 references to operations.md in consolidate.ts/assemble.ts/ingest.ts/cli.ts and 0 in agents/.

The architecture's load-bearing spec is the frozen-past-self guard (architecture.md:59): "operations.md holds a periodic user-authored 'where I'm headed now' anchor that deprecates older inferences." This is the human override on the machine's anti-drift mechanisms (evidence-threshold / decay-flag / anti-sycophancy): where those keep the machine honest, the dated anchor lets the user declare current direction that outranks stale inferences. S3d makes that real — a reader honours the anchor.

It is the right next slice because it is additive / non-breaking, read-only into generation (no engine or silo write), and the natural successor to S3a (same "wire the first reader" shape), now extended with the genuinely-new deprecation semantics.

3. Scope — what is IN (S3d)

1. Enrich the operations.md seed so the frozen-past-self guard is operational. operationsSeed() (scaffold.ts:52-64) currently ships bare headers. The genuinely-new seed content is the dated-anchor convention _As of YYYY-MM-DD:_ under the anchor (verified ABSENT from the current seed — grep "As of" scaffold.ts = 0) plus light item-format guidance for Plans/Ideas. Rationale: the guard "deprecates older inferences" — that is inherently temporal, so the anchor must carry a date the reader can compare against a profile fact's last_seen. NB: the phrase deprecates older inferences is NOT new to the seed — it already ships in the seed's guard comment (scaffold.ts:58) — so the seed's new-coverage test must key on the date convention, not that phrase (else it passes pre-enrichment; see SC3). Idempotency is existence-skip: initBrain writes a seed file only when it is absent (scaffold.ts:105-113), so an existing operations.md is never re-touched and a user edit is structurally safe regardless of seed content.

   • Fork (open question for the go-gate): the lighter alternative is reader-only, no seed change (Option A) — wire the reader against the existing bare seed. Recommended: Option B (enrich), because an undated anchor makes the guard vague. See §8.

2. Wire strategy-advisor to read brain/operations.md. Add the file to the agent's existing Step 0 Load-Context block (strategy-advisor.md:41-48) and a short consumption subsection immediately after the profile subsection (:63). The agent reads the markdown in-context (no parser — Claude is the retrieval engine, architecture.md:14; consistent with S3a).

3. Consumption contract — the frozen-past-self guard (advisory / reader-side).

   • The anchor (## Who I am now) is the user's own dated declaration of current direction — it is authoritative, and it deprecates older inferences: when a brain/profile.md Dynamic fact predates the anchor's date or contradicts its stated direction, prefer the anchor and flag the older fact as possibly-stale — do not parrot it.
   • Anti-sycophancy INVERSION (the genuinely-new nuance): profile facts are evidence to TEST (S3a — challenge them); the anchor is the opposite — it is user-declared direction, not an inference to challenge, so it is honoured rather than counter-pressured. (It is direction, not praise — honouring it never licenses flattery.)
   • Plans = active commitments (current intent). Ideas = parking lot (suggestions, not commitments). Weigh accordingly.
   • Durable sentinel literal for the lint: deprecates older inferences (verbatim architecture.md:59).

4. Graceful absence. brain/operations.md (or an empty anchor) → the agent degrades silently — no error, no "missing ops centre" noise; it proceeds on its other context sources. (Fresh installs have no brain until brain init.)

5. Deterministic wiring test — a new Section 16e in the test-runner.sh gate, modelled byte-for-byte on Section 16d (SB-S3a): two UNCONDITIONAL checks — (a) a non-vacuity self-test (full-wiring probe detected; 3 under-wired probes — incl. a sibling-file brain/profile.md decoy — rejected); (b) a real-file grep asserting strategy-advisor.md carries BOTH literals brain/operations.md and deprecates older inferences. ASSERT_BASELINE_FLOOR 80→82 (+2, lockstep). If Option B: a new scaffold.test.ts case asserts the dated-anchor seed convention → BRAIN_TESTS_FLOOR 113→114.

6. Doc/count reconciliation. Update architecture.md:80 (the SB-S3 build-row) to reflect the split: S3d ✅ ops centre · S3e remains (content-history retirement / triple-post reconciliation). Update STATE telling (gate 95→97, ASSERT floor 80→82, brain 113→114). No CLAUDE.md count change (no new agent/command/reference doc; strategy-advisor model/frontmatter unchanged).

4. Non-goals — what is OUT (deferred)

• The hygiene + reconciliation half (b+c) — dead content-history.md retirement and post-tracking↔published↔analytics reconciliation. This is SB-S3e (the new last slice). S3d touches no post silo, no state-updater.mjs, no content-history surface.
• Code-enforced deprecation. S3d's guard is advisory (reader-side): the agent prefers the dated anchor. It does NOT add a seam that makes consolidate.ts read the anchor and down-weight/flag older profile facts during --propose. That engine change is heavier (touches the sole writer of profile.md) and is an explicit follow-up, not S3d.
• A second reader. content-planner is the obvious next consumer (plans/ideas feed planning); S3d wires exactly one agent (strategy-advisor, which already reads profile.md) — the smallest valuable proof. Follow-on.
• A parser / new .mjs for operations.md. It is user-authored prose consumed in-context; it needs no TS grammar and no twin .mjs parser. (The only TS touched is the seed string in scaffold.ts + its test.)
• Any WRITE path to operations.md from the motor. It stays user-authored; the scaffold's existence-skip is the only code that touches the file, and only to create it once (never re-touched once present).
• A "Maskinrommet cockpit" surface. architecture.md:4 reserves the day-to-day cockpit for a future thin layer that reads/writes through the store. S3d ships the store substrate (a read tributary), not an end-user cockpit.

5. Boundaries / invariants (must hold)

• READ-only into generation — S3d adds no write path to any brain file; brain consolidate --apply --confirm stays the sole writer of profile.md, and operations.md stays user-authored.
• No post-silo contact — zero edits to state-updater.mjs, ## Recent Posts, content-history, analytics, or any tributary JSON. (That is S3e.)
• No tributary schema change — operations.md is a hub file, not a tributary schema; no ingest.ts/assemble.ts/consolidate.ts change.
• Idempotent + no-clobber — the enriched seed must preserve initBrain's existence-skip (a seed is written only if the file is absent, scaffold.ts:105-113); a user-edited operations.md is never overwritten (scaffold.test.ts:70-77 must stay green).
• Fresh-clone safe — missing brain → the reader degrades silently; the seed change cannot crash a fresh init.
• Anti-sycophancy preserved — the profile reader's evidence to TEST stance (S3a / Section 16d) is untouched; the anchor's authoritative stance is an addition for a different file, never a weakening of the profile stance.
• TDD iron law — the failing Section-16e check (and, Option B, the failing scaffold seed assert) lands BEFORE the agent/seed edit.

6. Success criteria (testable)

• SC1 — wired: strategy-advisor.md Step 0 context-load includes brain/operations.md. (lint, deterministic — Section 16e Check B)
• SC2 — guard framed: the operations-consumption subsection carries the exact durable literal deprecates older inferences, asserted by an exact-literal grep + a non-vacuity self-test with a sibling-file decoy (repo idiom, Sections 13/16c/16d). (lint)
• SC3 — seed operational (Option B): operationsSeed() ships the dated-anchor convention _As of YYYY-MM-DD:_; a NEW scaffold.test.ts case asserts that exact convention (the RED-bearing literal — verified absent from the current seed, so it fails first; the deprecates older inferences phrase already ships and is a non-RED companion only). BRAIN_TESTS_FLOOR 113→114 (verified current reported tests 113, so new floor = recorded + 1). RED proof recorded: run (cd scripts/brain && npm test) against the un-enriched seed and confirm the new case fails BEFORE the seed edit. (brain suite)
• SC4 — gate green: scripts/test-runner.sh stays green with Section 16e; TOTAL 95→97 (verified current 95; +2 unconditional checks; recount live at land rather than assume). The hook suite (separate runner) untouched (no hook path changes).
• SC5 — assertion floor honoured: ASSERT_BASELINE_FLOOR bumped by exactly the new unconditional checks (+2 → 82); the self-count check passes.
• SC6 — graceful absence (manual/behavioural): strategy-advisor with no brain/operations.md produces normal output, no missing-file noise.
• SC7 — guard works (manual/behavioural): strategy-advisor with a populated dated anchor that contradicts an older profile fact prefers the anchor and flags the older fact as deprecated/stale, not parroted — the frozen-past-self override observably fires. Honesty hedge (verifiseringsplikt): if a per-call data-root cannot be exercised, this is recorded as wiring-inspected-only, NOT a behavioural pass — never record a guard-fires pass that was not actually run.

7. Verification

• Deterministic (gate): SC1/SC2/SC4/SC5 via Section 16e in test-runner.sh; SC3 via the brain suite (cd scripts/brain && npm install before the gate, else it warn-skips — STATE rule). No hook path changes → hook suite untouched.
• Behavioural (manual, documented — the honest limit): SC6 + SC7 — agent-prompt behaviour is not unit-testable (command-testing workstream still open). The plan includes a documented manual run: (a) empty-brain → clean output; (b) seed a temp operations.md with a dated anchor that contradicts a hand-written profile fact → confirm the anchor wins and the older fact is flagged. Record the result in STATE/changelog at land; never overclaim a behavioural pass not run (verifiseringsplikt).

8. Open questions for brief-review / the go-gate

1. Seed enrichment (Option B) vs reader-only (Option A). Recommended B — a dated-anchor convention makes the temporal guard operational; A leaves the anchor undated and the guard vague. B costs one seed function + one brain test (BRAIN_TESTS_FLOOR +1). Confirm at the gate.
2. Advisory vs code-enforced deprecation. Recommended advisory (reader-side) for S3d — keeps it additive, no engine change. Code-enforcement (consolidate.ts reads the anchor date to flag older profile facts) is a flagged follow-up. Confirm the deferral.
3. One reader (strategy-advisor) for S3d. Confirm; content-planner as second reader is follow-on.
4. Doc home for the contract. Recommended: update architecture.md:80 (the build-row, now partly false post-split) + let the brief/plan be the contract record. Confirm we do NOT touch consolidation-loop.md (that doc is the profile motor contract; operations.md is not in that loop).

9. Light-Voyage review — folded

Three reviewers ran on the brief + plan (2026-06-23):

• voyage:scope-guardian: ALIGNED — 0 creep, 0 gaps, 0 dependency issues. Every IN item maps to a plan step; no OUT-list file is touched (the only mentions of post silos / consolidate.ts / content-history in the plan are the scope-fence list + the architecture.md:80 doc string that assigns them to S3e). Doc reconciliation stays minimal. Proportional.
• voyage:brief-reviewer: REVISE → 5 [FIX]es folded:
  • FIX-1 (load-bearing): the sentinel deprecates older inferences already ships in the seed (scaffold.ts:58) and at architecture.md:59 — so a seed test keyed on that phrase passes GREEN pre-enrichment, collapsing the TDD RED gate. Folded: SC3 + §3.1 now pin the new seed assertion to the dated-anchor convention _As of YYYY-MM-DD:_ (verified absent: grep "As of" scaffold.ts = 0). The agent-file Check B still greps the phrase (correct there — the agent file does NOT yet carry it). ✅
  • FIX-2: SC3 now requires a recorded RED proof (run the brain suite against the un-enriched seed, confirm the new case fails first). ✅
  • FIX-3: "compare-then-skip" was a mischaracterization — scaffold.ts:105-113 is existence-skip (writes only if absent; no content compare). Corrected in §3.1/§4/§5. ✅
  • FIX-4: SC7 now carries the honesty hedge (wiring-inspected-only unless a per-call data-root is exercised). ✅
  • FIX-5: gate TOTAL treated as "expected +2, recount live"; verified live: 95 → 97. ✅
  • Floor arithmetic confirmed correct by the reviewer: ASSERT 80→82, BRAIN 113→114.
• voyage:plan-critic: 3 major + 3 minor (folded into the plan, see plan-sb-s3d.md §"Plan-critic — folded"): the brain-test RED literal (= FIX-1), the recorded brain floor (verified tests 113 → 114), grep -qF fixed-string pinning, header-enumeration range :33-37, operationsSeed() cite incl. braces, and the load-bearing retained ## anchors for scaffold.test.ts:48-58.

Net: scope unchanged; the folds harden the TDD RED guarantee (the one real hole) + correct two prose mischaracterizations. Brief-reviewer's residual verdict after folds: structurally sound, well-scoped, honest.