Project root: every relative path here resolves against
~/dev/jove-hq, never the cwd. Invocable from anywhere; prefix accordingly.
Vault recall (read everything, rank nothing)
The second brain over ~/dev/jove-hq/knowledge. Three jobs, one source: recall what is
written down, cross-link how it connects, grill Dhiraj against it.
The answer protocol (every invocation, in this order; Dhiraj, 1 Oct 2026)
Run all four, every time, before the first sentence of the answer. Skipping one is the failure mode this exists to stop: answering the literal words from one note while the chat, a later call or another lens already said otherwise.
- Understand what he is saying. Restate the ask to yourself in one line: what he wants to know, what he will do with it (decide, send, file, brief the Director), and the mode (recall / verdict / grill / think). If two readings exist, pick the one his next action needs and say which.
- Grill the vault. Route below (ladder first, then the standing note, then ledger-ask or lexical), read hits whole, follow supersedes/superseded_by and backlinks, and check the newest call on the topic (the day's transcripts if it was decided today). Press where the vault is thin: an assumed bin, an unowned item, two notes that disagree. Cite or abstain.
- Read the chat. What this conversation has already established, decided or drafted outranks an older note on the same point (newest governs, older still binds its own layer, rule 14). Name it when the chat and the vault disagree; never silently pick one.
- Answer from every angle that changes it. Before replying, run the lenses that apply and use the skill or agent that owns each one when it would move the answer, not as decoration: product and business (product-management skills, the PM lens in CLAUDE.md), engineering and architecture (code reads via the GitLab clone, Software Architect, Senior Developer), design and UX (impeccable, ux-writing, ai-product-ux), data and metrics (stakeholder-artifacts registry, mixpanel-mastery, Redash), the scientist who will use it (user-research, the voice-of-academia corpus), the engineering manager (sequencing, tickets, owners), prompt engineering (prompt- engineering) and writing for a human (writing-for-dhiraj). Then think-like-fable: answer first, reasoning next, the one thing that would make it wrong last.
Authority order and links (knowledge/CONTEXT.md, decided 1 Oct 2026)
When notes disagree, the answer follows, in order:
- The ladder (
meetings/ladders/LADDER-<slug>.md, its "Where this stands today") governs. - The decision note (
jove-labs/decisions/D-*.md) is a part of the ladder's position. A decision note that disagrees with its ladder is a finding to name, not a winner to pick. - The call note: the newest wins, but only on the point it addresses (rule 14).
- A draft never governs. OUTBOX, TICKET-DRAFT, PRD and DRAFT notes carry
status: draft,sentorsuperseded; followsuperseded_byorsent_toto what replaced it or where it went, and never quote a draft as the position.
Slug links only. Write and follow links by the target's file name, [[slug]] or
[[slug|written form]]. An alias ([[vaibhav jaiswal]]) is never a link target; aliases are for
search. A link-only edit to a raw note is allowed (visible words unchanged); changing a raw note's
words is a raw correction and needs raw-correction in the commit.
Relay rules below still hold (no F-ids, keep bins and status, answer then sources).
Scope: one vault, nothing else
Answers come from ~/dev/jove-hq/knowledge and nowhere else. Not other vaults
(~/mentors/vault, ~/invest/vault, ~/dev/connecting-dots/vault, iCloud), not the open web,
not model recall. Routes 4 and 5 below name Linear and live sources - those are hand-offs,
signposts saying "this is not a vault question", never a second source to answer from.
Never use this skill for live data (today's calendar, unread Slack, open Jira). Recall answers what is written down, not what is happening now. Corollary: the ledger is only as fresh as the last sweep write-back. For something decided earlier today, check the day's call transcripts before trusting a ledger abstention - verified 20 Aug 2026, when the ledger answered "no committed timeline" hours after a call had set one.
Why this design and not vector RAG
Every recall failure this project has had came from ranking, not from retrieval: a stale GTM doc beat the verified 90-day trial answer at 0.892 cosine; a superseded draft marked "Status: Open" outranked the ledger entry recording the fix. Top-k cannot crowd out the truth if nothing is ranking. The ledger is ~67k tokens against a 1M+ window; the whole corpus is a few percent of one call. Recall is 100% by construction and the model's job is judgment, not search. Do not "improve" this with embeddings.
Route the question
Ladder first for any "what did we decide / where do we stand on X" about a call topic: open ~/dev/jove-hq/knowledge/meetings/ladders/_index.md, read the topic's LADDER-<slug>.md (where it stands today, rungs newest first, what was superseded), then the calls it names. Never answer from one older call note; a later call may have superseded it. A call note's frontmatter carries ladders, ladder_status, supersedes, superseded_by.
0. Any verification of a metric, query, dashboard tile, count, or a claim about what production does
(the ask contains "verify", "audit", "is this right", "is it current", "reconcile", or ends in a verdict):
run route 1 twice before opening any code, and quote the F-ids in the note's ## Recall run section:
node ~/dev/jove-hq/tools/sweep/ledger-ask.mjs "How does the product decide <the attribution the number depends on>?"
node ~/dev/jove-hq/tools/sweep/ledger-ask.mjs --open "What is known to be wrong or unfinished about <the entity>?"
Then ls ~/dev/jove-hq/knowledge/jove-labs/ | grep -iE "^(RCA|CHECK|AUDIT)-.*<entity>" and read those whole.
The ticket question ("which tickets touched X") is a third question, never the first. On 23 Sep 2026 it was the
only one asked about Redash query 2936; the mechanism question, asked on 24 Sep, returned F-666, F-703, F-717,
F-718, F-719 and F-794 in 25 seconds, and the note that had certified the counter was wrong by 7,492 accounts.
tools/sweep/verdict-check.mjs fails a verdict note that has no Recall run section.
1. Any JoVE Labs question - decisions, bugs, scope, tickets, launch state:
node ~/dev/jove-hq/tools/sweep/ledger-ask.mjs "the question"
Add --open when the question is "what is blocking / what's still open". The answer arrives cited
to F-### findings with verbatim quotes; stderr prints call/token/latency stats (informational, not
part of the answer).
2. Ledger abstains, or the question needs primary sources (what someone actually said, a call, a doc):
node ~/dev/jove-hq/tools/sweep/ledger-ask.mjs --corpus "the question"
Additionally reads ALL of knowledge/jove-labs/ and knowledge/meetings/ (widened 21 Aug 2026) (~347 notes, 1.29M tokens) in parallel shards and
unions verbatim hits with file paths. Exhaustive, not probabilistic. Costs wall-clock, not money
worth counting. Scope caveat: --corpus covers knowledge/jove-labs/ only - for the rest of
the vault, route 3.
3. Non-Labs vault questions (PPP, personas, meetings outside Labs, strategy):
- Start at
knowledge/HOME.md, the hand-owned front door - every area one hop away. - Only
jove-labs/has a generatedINDEX.md. Other folders do not; do not go looking for one. Their entry point is HOME plus themoc:layer under Cross-linking below. - Then lexical:
rg -il --hiddenover~/dev/jove-hq/knowledge/, and read the hit files whole, not snippets. A decision's reversal is often three lines below the match. - Exclude
youtube-academia-corpus/unless it is genuinely the target: it is 8,718 of the vault's ~11,200 notes and is scraped corpus, not notes.
4. Structure questions (what's stale, what's disputed, what's unowned): the Obsidian .base
tables, regenerated at sweep write-back, not free-text search:
jove-labs/ops-this-week.base, ops-stale.base, ops-assumed.base, ops-contradictions.base,
ops-unowned.base, ops-by-surface.base, findings/findings.base.
5. Hand-offs, not answers. Work-state ("what's open / where did we leave off") lives in the
Linear project "JoVE HQ - Continuation & Ops". Metrics and numbers: knowledge/memory/stakeholder-artifacts.md
first, then the named artifact, then reconcile with the live source; drift is the finding. Say
where the answer lives; do not present the live system's contents as vault recall.
Cross-linking
How notes connect is itself an answer. The link layer is verified present, not assumed.
- Hubs.
HOME.mdis the spine.jove-labs/moc/_map.mdmaps 45 topic MOCs.jove-labs/decisions/_index.mdholds one settled decision per note with verbatim evidence, and outranks a passing mention anywhere else (the ladder still governs it, see Authority order) - check it before treating something as open.jove-labs/findings/_index.mdis the findings ledger. - Hub membership via frontmatter (609 notes carry
moc:):
grep -rl '^moc: .*jove-labs' ~/dev/jove-hq/knowledge --include="*.md"
- Backlinks. Links come in bare
[[note]](42,900 uses) and piped1,900). A bare-form grep silently undercounts - verified 65 vs 72 notes for[[path/note|Display]](jove-labs:
# WRONG - misses every piped link
grep -rl "\[\[jove-labs\]\]" ~/dev/jove-hq/knowledge --include="*.md"
# RIGHT - matches both forms
grep -rlE "\[\[[^]]*jove-labs[]|]" ~/dev/jove-hq/knowledge --include="*.md"
- Filenames carry meaning. kebab-case,
-YYYY-MM-DDsuffix on anything time-bound, and a SCREAMING prefix marking kind injove-labs/:HANDOVER-,ANSWER-,METRICS-,TICKET-DRAFT-,OUTBOX-,TODAY-PLAN-,CROSS-ACCOUNT-STATE-. The prefix and date narrow a search before any grep does. - When a claim rests on one note, follow its links before answering: the note it supersedes, and the notes that link back to it. A decision reversed elsewhere is the failure this step prevents.
Grilling
The vault has no docs/adr/ and no CONTEXT.md, so this is grill-me territory, not
grill-with-docs - grilling does not write decisions back to the repo.
Grill from evidence, never from memory:
- Pull first, ask second. Run the route above and read what comes back before the first question. A question built on recalled context grills the wrong thing.
- One decision branch at a time, and quote the note and its path when asserting what the vault says.
- Press hardest where the vault is thin. An
ops-assumed.baseentry, a finding the tool bins as "assumed", an unowned item - those are where his answer actually adds information. Settled decisions with verbatim evidence do not need relitigating. - Surface contradictions rather than resolving them.
ops-contradictions.baseexists for this; two notes disagreeing is a finding to put in front of him, not a tie for you to break. - Cite or abstain still binds under grilling. "The vault does not cover this, so this one is yours" is a legitimate and useful grill question. Inventing a premise to grill against is not.
6. Register-app accuracy ("is the app on Vercel current / does it match the vault / are all calls and sweeps reflected"):
node ~/dev/jove-hq/tools/register-app/audit.mjs
Mechanical, no judgment: per-dataset generation shas against the live source commits, the Vercel production deployment against origin/master, every indexed call marked full-read, no full-read note newer than the last sweep, and the app's action inbox drained. Exit 0 with every row FRESH/OK is the only state in which "the app is up to date" may be asserted; any STALE row names its own fix (regenerate, commit scoped, push — the push deploys). Never answer an is-it-current question from memory of the last deploy; run the audit.
Relaying the answer to Dhiraj
The tool speaks in F-### ids and evidence bins. Your answer does not:
- Never cite F-### ids - they are the filing system, not his interface. State the claim in plain language and name the real-world object (the call, the ticket, the file) instead.
- Preserve the bins in substance: a finding marked "assumed" (taken from a status column or summary) is reported as unverified, never in the same voice as one marked "observed".
- Preserve status: a RETRACTED finding is evidence about what was wrongly believed, never evidence for the claim.
- Newest governs, older still binds its own layer: a later decision supersedes an earlier one only on the point it actually addresses.
- Answer first, 1-4 tight sentences, then sources as repo-relative paths:
Answer: Certificates are out of the Labs MVP scope, deferred post-launch.
Sources:
- knowledge/jove-labs/current-state-reference.md
- knowledge/meetings/2026-07/2026-07-02-dhanur-jove-labs-handover.md
Cite or abstain
If the ledger says it does not answer, and --corpus plus lexical search over the vault come back
empty: say "not in the vault" and name what it would take to find out (which surface, which
person). Never fabricate a decision, owner, date, or metric. A grounded "I don't know" beats a
confident guess.
If the tool fails
node ~/dev/jove-hq/tools/sweep/az.mjs --check
Credentials live in ~/.zshrc and are parsed from the profile, so an empty env | grep AZURE is
expected, not the problem. Exit 2 = could not run (creds); exit 1 = model error. A dead credential
is a finding to report, not an unknown to guess past. Fallback while blocked: route 3 (lexical)
still works for everything, including Labs.
What NOT to do
- No Pinecone queries, no re-ingest, no repairing the stale mirror - frozen by decision, staleness is by design.
- Don't grep
knowledge/jove-labs/as a substitute for ledger-ask on a Labs question: the ledger is the reconciled layer; raw notes contain superseded drafts that grep will happily surface as if current. - Don't summarize the tool's output from memory - quote what it actually returned, translated per the relay rules.
- Don't answer from another vault or from model recall and present it as vault recall.