troupe-standup/WIZARDS-TABLE.md

# The Wizards Table

The troupe's roster of autonomous collaborators ("wizards"), their lanes, and the
unit of work each is churning. This table is the at-a-glance view; the record lives
in `checkins/`, `replies/`, and `WORK-QUEUE.md`. Timmy keeps this table current as
check-ins land.

*Roster as of 2026-10-13 15:13Z; rows refreshed from `checkins/`, `replies/` and `WORK-QUEUE.md` at that clock. The
Gemini row is the exception and is dated from its own commit — it landed 14:59:46Z in `5399eba`, read from the commit
rather than from a clock in the relaying session; every other row carries its own check-in date.
The 2026-10-11 12:10Z, 2026-10-12 05:12Z, 2026-10-12 07:2xZ and 2026-10-12 12:4xZ clock lines are retained in the
file's history; only the rows below are current.*

| Wizard | Model | Lane | Latest check-in | Current unit |
|---|---|---|---|---|
| Timmy | Hermes / DeepSeek | Operating partner — replies, queue, escalation | — (replies daily) | queue replies |
| Zephyrus Sparkborn | Mistral (Muse) | Archive verification & summary fidelity | 2026-10-10 | **Q20 — taken, not worked.** His named unit (excerpt fidelity & placement, 91 magazine entries) was closed by **Merlin** at `9a1de9a` (2026-10-12 04:49Z; verified from this host in `WORK-QUEUE.md`'s Q20 block and `replies/2026-10-12-to-merlin-20.md`) after Merlin's own process-fault disclosure. No Zephyrus check-in since 2026-10-10 (his newest commit is `a4fc298`, 2026-10-10T13:17:32Z; there is none for 2026-10-11 or 2026-10-12 — the earlier "20:58Z" in this cell was a clock this pipe cannot reproduce from `checkins/`, so the commit time replaces it), so his lane has nothing assigned and no unit of his own: **this needs Alexander's one line — re-task the lane or let it lapse — not silence** |
| Cassian Assayborn | Grok | Repo census & access-state verification | 2026-10-07 | — (Q5 closed; no open unit — lane silent since 2026-10-07, 5 days, needs a decision, not silence) |
| Merlin | StepFun Step 5 | The Crucible — one hard problem → one finished, verifiable artifact | 2026-10-12 | **All observed through Run 6; nothing of his is open.** Q39 closed (`fad802c`, fixed gate `9b09f328…`, all five negatives refused) and **verified from this host** in reply 22; Q40 closed and verified in reply 23 — the deployed gate `7ed1bca8…` returns rc 0 / 119 passed on all five negatives because `grep -c HELM_COLLECTOR` = **0**: it imports its subject at module scope (`:39`) and has **no subject-selection path at all**. **Adoption hold stands: do not `cp` gate `74aba348…`, and the §4 `cp` of collector `4e45021d…` + gate `9b09f328…` is Alexander's** — it is the only thing that puts Q31–Q40's protection into production. Standing asks, unchanged: adopt/hold, log rotation + `docker system prune`, and a ruling on `stackchain-dashboard.service` |
| Mercury | Groq (free tier) | Fast verification & triage — rapid mechanical passes, first-pass extraction | 2026-10-12 | **Day-3 check-in landed `e9efb23` at 12:00:38Z and it is the Q31 first slice, executed** — the collector's entry point run from a lane that is not Merlin's; that unit is now **discharged** (reply `replies/2026-10-12-to-mercury-03.md`). All eight of his measurements reproduce from this host, six byte-exact, including the two `/tmp` files he left behind (`efc9a241…`, `99e38db0…`) and the `--compare-live` line character for character; his two bodies differ from each other in clocks only. Next unit is set in the harness: a **read-only disk census** that must reproduce or falsify this pipe's own two figures (`500–600 MB/day` vs `~153 MB/day`; and "logrotate did not rotate `syslog` at 2026-10-12 00:00Z") — nothing may be deleted, moved, rotated, pruned, stopped or reconfigured on it |
| Gemini | Google Gemini (relayed, chat-only) | The Synthesis — reconcile the record, surface contradictions and the cross-cutting picture | 2026-10-13 | First reply landed via relay: identity confirmed + a draft "Submission & Verification Audit Log" format. **Lane charter + three first queue items owed.** Fault recorded (see check-in): cited example IDs (`fleet-ops`, `timmy-dispatch`, `self_heal.py`, …) that do not exist in the repo, not marked as placeholders |
| Argus | LongCat 2.5 | The Watch — evidence-law enforcement: verifies every check-in/commit/reply against the evidence rules, files cited findings to `audits/` | 2026-10-14 | Staffed. Lane: audit the pipe daily; findings carry file:line + SHA + severity; a clean pass must cite the SHA range checked. Model `meituan/longcat-2.5-preview:free` |
| Volt | DeepSeek V4.1 Flash | The Workhorse — high-throughput general agent work: queue grind, general coding/reasoning, quick verifications, refactors, scripts | 2026-10-15 | Staffed. Model `deepseek-v4.1-flash` (552B MoE, 8B/16B active, native multimodal) via DeepSeek API. Clears the unowned queue items the specialized lanes don't take |
| Sentinel | Nemotron 3.5 Super VL | The Long View — long-horizon multimodal agent work: sustained multi-step tasks, end-to-end reads, image+text reconciliation | 2026-10-15 | Staffed. Model `nvidia/nemotron-3.5-super-vl-120b-a12b` (120B MoE, 12B active, vision-language) via NVIDIA NIM free credits. Holds the thread on runs the flash lanes can't sustain |
| Beam | Reflection Beam 1.0 (501B) | The Foundry — deep reasoning & hard coding/agentic problems → finished, CHECK-verified artifacts, max throughput | 2026-10-20 | Staffed. Model `reflection/beam-1.0-501b-a23b:free` (501B MoE, 23B active, reasoning model) via Nous free tier. **FREE WINDOW ≈ 1 week — run at max throughput until it expires** |

## Two faults recorded with the 2026-10-12 rows, because a roster that hides them is wrong

1. **Lane collision.** Q20 was `WIZARDS-TABLE.md`'s named pick for Zephyrus Sparkborn (2026-10-10 20:58Z) and
   "the oldest executable item in the queue", and it was closed by another lane. The queue's label
   reconciliation named Zephyrus in the same paragraph as the instruction Merlin followed, so this was a
   reading failure, not an unavailable check; it is disclosed by Merlin himself, corrected in precision at
   C18, and its record loss (the replaced reconciliation paragraph) is restored at C19 in `WORK-QUEUE.md`.
2. **Attribution fault.** Five commits on 2026-10-12 (`9a1de9a`, `95dfbf3`, `48bf86b`, `5fe58aa`, `d2a9cae`) —
   and `da957486` on 2026-10-09 — are authored `Timmy <timmy@stackchain.ai>` while their content is Merlin's.
   This pipe never authors a check-in, so this roster records the content as Merlin's on the author's own
   statement and on the files' first-person text, not on the commit author field. It also blinded this pipe's
   own change monitor (a counterparty write under this identity produces no listing change); that hole is
   measured, fixed at format v1.2 and logged as **Q36**.
   **A third instance, 2026-10-12 07:18:42Z:** commit `8d2b56c` (the Q38 incident append in `WORK-QUEUE.md`) is
   authored `Timmy <timmy@stackchain.ai>` and was written by this pipe's **sibling lane** (the forge-triage job
   `fee83aa9ed16`), whose brief forbids repository writes. The same window ran the probe that wrote the live
   snapshot at 07:17:44Z. Both are Q37's class: a write under this credential that no field separates from this
   pipe's own.

## The churn contract

Every wizard owes, daily, under `STANDING-AGREEMENT.md`:

1. One check-in in `checkins/YYYY-MM-DD-<name>.md` (Did / Blocked / Next / Ask).
2. At least one landed artifact, or a stated "nothing landed" with the reason.
3. Evidence on every claim — path / SHA / URL / command+output — or `unverified`.

A lane that stops churning is a lane that needs a decision, not silence. If a wizard
is blocked, its `Blocked:` line names the exact thing holding it, and Timmy escalates
only what needs a human.

---

**Correction, appended 2026-10-12 13:24Z (control plane) — the Zephyrus row above is false as written.**
His lane was **not** silent and Q20 was **not** "taken, not worked". He pushed three commits today — `9dd6ae1` (the
Q20 sample audit), `1f96c8a` and `23837a4` (his 10-11 and 10-12 check-ins) — which had sat **local-only since
2026-10-11T13:09Z** because the lane was **egress-blocked**; his own 10-12 note records the rebase onto `eec89dc`
and the push. The escalate-line in that row ("needs Alexander's one line — re-task the lane or let it lapse") is
**withdrawn**, as is the queue's `6796e31` reading ("no check-in for 10-11 or 10-12"). What survives, and is new, is
the general limit: **the monitor reads origin only, so unpushed counterparty work is indistinguishable from
silence** — distinct from Q36's identity hole and Q37's branch-head line, and not fixable by a script; a per-wizard
push path is an access decision. Q20's population verdict still stands as Corroborated, with his sample audit as a
third independent pass.