JCube — Simulation Mode Report

Real historical inbound mail replayed through the real Mailroom, Intake & Sourcing agents — offline, nothing ever sent to a client or factory. Generated 2026-07-23 22:56 UTC.
30 day window 366 job threads 640 real emails 0 processing errors
What this is, in one line: we ran the real AI agents against a month of J-Cube's actual email — the same way they'd behave live — but with sending physically disabled, so we could see their judgment side-by-side with what Jason actually did, before any of it touches a real client or factory.

What came through the door

Every inbound email this window, sorted by what it actually was
Factory quote 125 RFQ reply 103 Engineering query 96 Shipping notice 86 Needs human 66 Other 52 Purchase order 42 Client question 24 Factory question 23 Spam 11 Complaint 7 factory-invoice 2

Factory-pick scorecard

Where the agent recommended a factory AND we have a matched record of what Jason actually did
366 threads
Agreed with Jason — 3 (1%)
Agent deferred, Jason picked — 27 (7%)
No comparison possible — 336 (92%)

Agreed with Jason (3)

JobAgent pickedJason picked
W180122VENTUREVenture, TopTek
SIM-PO15086TOPTEKTopTek
SIM-POtentialFASTLINKVenture, FastLink, JHY

Differed from Jason (0)

JobAgent pickedJason picked
None this window

Where the agent deferred and Jason picked anyway (28)

Not mistakes — the agent correctly said "no confident signal" rather than guessing. This is the real map of what's missing from its factory-selection knowledge.
FastLink 11 TopTek 9 Venture 4 JHY 3 IQ 1

What the agents actually decided — sample cases

One real example per category this window, in the agents' own words
New RFQSIM-THREADgenteccapcbedd
What came in: Gentec (Nicolas Reignier) corrects a previous Gerber submission — the wrong file was sent for an Eddyfi board — and requests a fresh quote for part 10033552-R02 (ACC269202 / C10034647), qty 5, 10-day lead time, with a new Gerber attached; full board specs (layers, finish, thickness, etc.) must be re
Agent's action: Opened/updated SIM-THREADgenteccapcbedd ([contact]/OTHER). Layers resolved to 17. 1 flag(s) raised. | Recommended undefined for SIM-THREADgenteccapcbedd; RFQ drafted (no commitment).
⚑ Client [contact] province is OTHER/US — tax rule is a [GAP — Jason]. Billing will block on this.
⚑ Factory selection is a human gate. Recommendation: none yet. New job (or no good prior run). Factory-fit profiles are a [GAP — Jason] in factories.md — shortlist by board fit only and surface for human selection.
Engineering queryW180122
What came in: Venture's CAM engineer (Alisa) issued a 12-item Engineering Query for job W180122 / PN 06W1801220750A (SABB2000 Rev A), covering copper thickness, hole tolerances, board thickness, IPC class, stackup/impedance, and via treatment; Shirley is now following up asking whether Jason has a reply from Juli
Agent's action: EQ recognized from VENTURE — production blocked, surfaced to Jason (no auto-confirm).
Needs humanW180122
What came in: Venture CAM engineer Alisa has sent the working Gerber and Engineering Query (EQ) for PO 137580-1 / PN 06W1801220750A Rev A, requesting confirmation before production can start — but there is a critical layer-count conflict: the Gerber reads 16-layer while the Venture part number encodes 6-layer, an
Agent's action: Venture CAM engineer Alisa has sent the working Gerber and Engineering Query (EQ) for PO 137580-1 / PN 06W1801220750A Rev A, requesting confirmation before production can start — but there is a critical layer-count conflict: the Gerber reads 16-layer while the Venture part number encodes 6-layer, an → Flag to Jason immediately as 'production blocked — EQ awaiting confirmation.' Two issues require human judgment before any reply: (1) resolve the layer-count discrepancy (Gerber says 16L, part number 06W… says 6L — which is correct for this board?); (2) review and answer the EQ questions in the attached spreadsheet. Do NOT auto-confirm the Gerber or any EQ answer. Also note: this job's part number
Price negotiationSIM-THREADwestportcoma46
What came in: Sheri finds the quoted $9.17/pc USD (incl. shipping) too expensive and is pushing back on price, while also requesting the production panel Gerbers used for the last build of 20-04951-6.
Agent's action: Price pushback recognized — would reopen the price gate, never auto-adjust.
Factory quoteSIM-THREADfastlinkpcbcom
What came in: FastLink (Tina) returns a price for the Gentec/Pavemetrics ECM0719B board, having assumed standard specs (ENIG 1µ", FR4 TG170, 1.6mm, 1oz all layers, green SM, white silk) because the original RFQ lacked them, and includes both standard (14 WD) and quick-turn lead times — asking Jason to confirm the
Agent's action: Cost picture ready for SIM-THREADfastlinkpcbcom. Stopped at the margin — Jason prices.
⚑ Factory selection is a human gate. Recommendation: FASTLINK. New job (or no good prior run). Factory-fit profiles are a [GAP — Jason] in factories.md — shortlist by board fit only and surface for human selection.
⚑ Cost picture assembled for SIM-THREADfastlinkpcbcom (factory cost $0.00). Awaiting Jason's price — engine will not propose one.
ComplaintW180122
What came in: Client Paul Willis (Selcom/SECO) reports that the factory EQ and working files were based on a superseded fab package (G0F72A10.zip) that contained ODB errors; he demands all files be discarded and the EQ/working Gerber redone using the corrected package G0F72A12.zip he has re-attached, and flags th
Agent's action: Client Paul Willis (Selcom/SECO) reports that the factory EQ and working files were based on a superseded fab package (G0F72A10.zip) that contained ODB errors; he demands all files be discarded and the EQ/working Gerber redone using the corrected package G0F72A12.zip he has re-attached, and flags th → Needs Jason immediately: (1) forward G0F72A12.zip to Venture (Alisa/CAM) with explicit instruction to discard all G0F72A10-based files and regenerate the EQ and working Gerber from G0F72A12 only; (2) do NOT confirm the existing EQ — it is based on wrong files; (3) reply to Paul acknowledging receipt and confirming the correction is underway. The EQ gate is blocked and production cannot start until

By the numbers

96
Engineering queries recognized & held for human confirmation (0 auto-confirmed)
86
Shipping notices recognized
1
Price pushbacks — price gate reopened, never auto-adjusted
66
Items correctly escalated as needs-human
7
Complaints recognized
$0
Client price ever proposed by an agent (structurally impossible — no margin logic exists)

Limitations — read this before trusting the numbers above

1
Only 3 thread(s) had both an agent factory pick and a matched real outcome to score against — 100% agreement is directional, not statistically proven. The 28 deferred cases above are the more useful signal right now.
2
Factory-initiated threads with no client RFQ in the window get a placeholder client record — 200 threads here — which can spuriously trip tax-uncertain flags. Treat those as a simulation-harness artifact, not a real new-client count.
3
Coordination and Billing agents were not separately re-run — in the live system they only fire after a human resolves a gate (EQ confirmed, price approved), which a one-shot historical replay can't reconstruct. EQ/shipping/PO handling shown above is the same deterministic logic the live system already uses, not a new judgment call.
4
Spec extraction (quantity, layers, material) was checked for presence and plausibility, not against independently-verified ground truth — no dataset of "what was actually built" exists yet to check it against.

Recommendations

1
Feed the 28 deferred-but-Jason-picked cases into the routing model as training signal — the concrete next step to close the factory-selection gap.
2
Get factory-fit profiles from Jason (which of the 7 factories is best at what) — the single highest-leverage input missing right now, confirmed by this run.
3
Run this again on a rolling basis as more real mail accumulates, to grow the scored sample past statistical noise.
4
Tighten the placeholder-client artifact so future reports don't need the caveat above.