InsForge backend
Postgres schema, the attribution engine, RLS access control, a realtime channel streaming pipeline stages live, pgvector for fuzzy reference resolution.
Witness
Findings,
not scores.
Performance reviews miss the work that happens in Slack. Witness finds it, and cites the proof.
← →
or
space
to navigate — scrolling does nothing
The problem
The system of record only credits what it can see.
Commits, PRs, tickets closed. The highest-leverage work looks nothing like that.
Unblocking a teammate, catching a bug in review, triaging an incident — all of it happens in Slack
and gets credited to whoever closes the ticket afterward. In the workspace we built to test that
claim, the gap is measurable, not asserted:
2 / 24
Tickets closed by the engineer who diagnosed the most incidents in Slack — last on the team by that metric
7
Slack threads where that diagnosis precedes someone else's ticket closing
0
Of those threads counted toward any ticket-based ranking
The method
Witness surfaces that work, and cites the proof.
For each engineer, it finds things they did that helped a ticket or PR credited to someone else. A Slack message alone isn't evidence. A Slack message referencing a ticket assigned to someone else, that closes two hours later, is.
Rule A
Message references a ticket assigned to someone else. That ticket closes within the confirmation window.
Rule B
Message references a PR or file. Someone else commits against it within 24 hours.
Output
Cited artifacts per person. No score, no ranking number, ever rendered.
How it works
Attribution isn't intelligence. It's a temporal join.
It runs entirely inside Postgres, as a SQL predicate, not an LLM call.
confirm_attributions.sql — rule (a)
-- referenced ticket assigned to someone else, -- closed within the confirmation window JOIN ticket_state t ON t.assignee_person_id <> e.author_person_id AND t.closed_at BETWEEN e.ts AND e.ts + make_interval(hours => confirm_window_hours)
Restrict the tool to one connector, and the demo doesn't go blank — it becomes a confident, wrong performance review. Why the join had to move into the database
The degradation guarantee
Turn off a source. There's nothing left to join against.
Every read in the attribution path goes through a SQL function gated on the run's enabled_sources. Disable Linear, and ticket_state rows are structurally absent from the query, not filtered out by a branch that could be wrong, or faked.
Access control
Same mechanism. RLS decides who sees what: an engineer sees their own findings, a manager sees direct reports, nobody sees the org.
Enforced by
auth.uid(), not UI convention.
What each piece actually does
Not a features checklist.
HydraDB identity
Ties a Slack user ID to a Linear assignee UUID with no shared key. Mirrors Slack messages as searchable knowledge, verified live.
RocketRide orchestration
Live-tested against a real Cloud account. Two real platform gaps found and documented, not guessed around.
Validation
Tested against a synthetic company, not a claim.
Real accounts on Slack and Linear. Filler messages generated from behavior, never from the detection logic itself.
24
Linear tickets
~110
Slack messages
7 / 7
Confirmed unblock threads found
Identities
Three personas, each a distinct real Slack account and Linear account — no shared identifier for entity resolution to lean on.
Filler generation
~90 background messages generated from behavioral character briefs, blind to the detection rule — the signal wasn't hand-placed for the pipeline to find.
Witness