90 minutes · two parts · break at 45
Run of show
Beat by beat: what runs live, and the single point each beat exists to make. Read Positioning and the Demonstration outline first. This page is the script for the room, not the argument.
Ninety minutes is not the 45 stretched. The 45 minute cut could not carry what had been built: three finished pieces were absent from it entirely, namely the Stage 1b verification matrix, the Stage 4 Capital interface, and the capture-replay finding, which was marked optional. The 90 has a break at 45, room for the questions this audience actually asks, and the "what we will not claim" list is now spoken rather than only written down.
SGT is in the room. That is settled, and it changes the run of show rather than only answering a question: the Title 10 and accreditation beat is live at minute 86 to 90, with the people who can actually answer it, instead of being deferred to a follow-up.
1Part one, 0 to 45: the thread
0–5
Their own findings slide, read back to them. Stove-piped development process, insufficient design verification, limited visibility into assembled hardware, architecture hardware and firmware immaturity.
We are answering the problem they stated, not the one we sell to.
5–8
The MSP-2 article. DMSMS-driven, their words, their project. Say it is fictional and unclassified before anyone asks.
This is your problem, scoped to one card.
8–13
Run the baseline. It abstains. 15.2% coverage on five simulation-only records, and no readiness figure reported at all.
A readiness number you cannot audit is worse than none. This sets the tone for the whole 90 minutes and is worth the five minutes.
13–20
Stage 0, the mission layer. The Seg0 mission, three ordered phases, environment and capability. The 105 °C limit is owned by the track-maintenance phase, which links to the 55 °C hot-day case. Cameo carries the context of use, deliberately shown managed through MBSE Integration Services: a competitor's authoring tool that the thread holds anyway, said out loud.
The thermal failure will arrive later as a mission consequence rather than a datasheet number. This is the reason to start at the mission, and it only pays off at minute 45.
20–29
Stage 1, requirements as a real specification. 22 requirements in Teamcenter, the spec rendering as a document with numbered paragraphs, verification methods assigned, requirements carrying typed parameters.
Verification plan automation, from their own REMEDE slide. Show the Content tab, not a flat list: the rollup is the point.
29–37
Stage 1b, the Section 4 verification matrix. 37 rows with real verification statements, 37 typed trace links carrying methods from Teamcenter's own vocabulary, plus the Verification Requests, Simulation Requests and Tests underneath them, and the saved trace link matrix. Read the two footnotes off the screen rather than defending them: columns f and F are empty by intent, and 37 links give 36 marks because MSP2-SR-006 is verified by Analysis under two techniques and they share a cell.
This is the artefact an RFP Section 4 response is graded on, generated rather than typed.
Mechanics, get these right. Open Quick Matrix, not Full-Rollup, and expect it in a new browser tab. Do not go looking for it in the work area.
If the beat is running to time, close it with the generic report. Open Verification Cross Reference Matrix (XC0_VCRM_GENERIC_001), change ReqPattern off the MSP-2 value, Generate. Identical shape, different programme, live. Be precise about the scope: it reads object identity from the Teamcenter type and pairs a verification requirement to its requirement by naming convention, so it fits any programme on a tier using those types. Do not say "works for anything". This is the first thing to drop when 29 to 37 needs trimming.
37–45
Stage 2, the trade. Run
check_criteria.py first, showing the freeze timestamp, the authority, and basis: drafted by the account team, not elicited from stakeholders. Then run the trade over both candidates. Say CONTESTED out loud.Every vendor shows a winner. Show the freeze, show its caveat, and show that the engine refused to call its own result robust because six of eight criteria abstained.
Break at 45. Take questions here. Everything before it is the thread; everything after it is the evidence.
2Part two, 45 to 90: the evidence, and what it costs to be honest
45–55
Stage 2b, multi-mode thermal. Simcenter Flotherm, measured: Candidate A at 110.12 °C against a 105 °C limit, margin −5.12, converged in 410 of 1500 iterations, energy balance closing to −0.024% of 61.9 W, and 1.59 °C/W attributable to the extra 7.4 W on a bit-identical mesh. Candidate B measures 100.16 °C and passes. Say plainly that HyperLynx and Simcenter 3D are not installed here, so this is one technique of the multi-mode column rather than three.
The column no pure-EDA competitor fills. Say that this number corrected our own slide, which read 111.8. A demonstration that survives its own solver is worth more than one that agrees with its placeholder.
55–64
Stages 3, 5 and 7, the hardware-accurate twin. Questa coverage and formal analysis; then the OFP in the loop on a SystemC/TLM virtual platform, 2.500000 GSPS with 0 ICD deviations over 23 checks; then the same OFP cross-compiled to aarch64 running on a modelled quad-core Cortex-A53 under Nucleus RTOS in Innexis Architect Explorer.
Parity on the incumbent's ground, deliberately underplayed. The reframe is from "whose emulator is faster" to "who integrates the system before the hardware exists".
64–70
Stage 4, the DESIL interface. 20 signals over 32 pins to a neighbouring subsystem twin. Candidate A routes six safety pathways through the device that also carries mission and platform traffic: three safety nets share a device with mission nets. Candidate B terminates them on the partition isolation controller: zero.
Their Stage 4 ask, and a measured architectural difference rather than an assertion. Be honest that Capital itself did not run here: the bundles are authored and validated, and the import step needs the Capital GUI.
70–77
The model maturity ladder. Python concept model, FMU, virtual platform, and a Simcenter romAI ROM. The toolchain ran: four ROMs trained on the Stage 6 capture, four FMI 2.0 FMUs exported and every one stepped 1538 times. And the ROMs score worse than a constant predictor on a holdout block they never saw, while the afternoon of Python reproduces that same holdout bit-exactly, 384 of 384 cycles identical.
The most credible seven minutes available. We built the expensive rung, we ran it, and we are telling you it is the wrong one for this article.
Do not split this claim. It has two halves and dropping either one misleads in the opposite direction from dropping the other. Quote the toolchain half alone and we have oversold a ROM we know is poor on this article; quote the fit half alone and we have undersold a product that trained, exported and stepped four FMUs on the first attempt. Both sentences travel together or the slide does not go up.
77–82
The capture-replay finding, promoted from optional to its own beat. A trace taken off the RTL and replayed into the concept model disagreed on 1233 of 2712 outputs. One assumption about decimation phase across a refused sample, which the requirement never specified. Change that single assumption and the same trace replays to 0 mismatches of 2712. Neither engineer was wrong.
A requirements defect, found by comparing artefacts. This is the class of thing that surfaces in an integration lab in year three, and the capture found it in one run. At 45 minutes this was cut; it is the best illustration of the thesis in the whole demonstration.
82–86
Re-run the readiness record live. The gate moves to NOT READY on the thermal failure. Show the discarded record, the unverified list,
co_analysis at 3 of 6, and that adding a second co-simulation tool did not move the number.Computed from evidence, cites objects that re-fetch, abstains where it must. A figure that refuses to move is the proof it is not theatre.
86–90
The gaps named, SGT on the accreditation path, then their closed-loop slide with PLM at the top. Name what is open: 6.4.11 validation is method-confirmed but the trace does not yet reach a stakeholder-level requirement; 22 of the 30 process skills have never been run against live data; the criteria weights were drafted rather than elicited; Innexis VSI standalone does not license here.
Credibility, then the Title 10 answer from the people who can actually give it, then the one box on their own diagram the incumbent does not sell.
3What 90 minutes buys that 45 did not
Three things that were built and could not be shown:
- The verification matrix, which is the artefact an RFP Section 4 response is actually graded on.
- The Capital interface, which is their Stage 4 ask by name.
- The capture-replay finding, which was cut for time and is the single best illustration of the argument.
It also buys the two things a technical audience needs and a 45 minute script cannot give them: a break with questions at the seam, and time to say the "what we will not claim" list out loud rather than leaving it in a document nobody reads.
4Talk track, the moments that carry the room
Opening the baseline (min 8–13)
"Before we show you anything that passes, we're going to show you something that refuses to answer. This is the incumbent card, run through the same engine, on five simulation-only records. It comes back abstained: 15.2% coverage, no readiness number. That refusal is the point. A number you can't audit is worse than no number, and everything you see for the next hour and a quarter is held to that same rule."
The freeze, before the trade (min 37–45)
"These eight criteria were frozen at 00:45 UTC on 14 August, before a single piece of Stage 2b or Stage 3 evidence existed. The tool prints the timestamp, the authority and the basis on every run, and it refuses evidence dated earlier. And here is the qualifier that prints with it: we drafted these weights, we did not elicit them from your stakeholders. The control we're demonstrating is that criteria were fixed before evidence was generated. The stronger claim, that these are your priorities, is not one we're making."
Handing over the multi-mode column (min 45–55)
"This is the one column on your own IV&V slide that a pure-EDA vendor cannot fill at all. Watch the solver converge, 410 of 1500 iterations, energy balance closing to two hundredths of a percent, before we tell you what it found. And one more thing: this run corrected our own slide. We had 111.8 degrees in the deck from constructed data. The solver says 110.12. The conclusion survived and the slide did not, which is the outcome this demonstration was built to be capable of."
The ladder, both halves together (min 70–77)
"We built the reduced-order rung. Simcenter romAI trained four models on a trace off the RTL, exported four FMI 2.0 FMUs, and we stepped every one of them. Now the honest half: on a holdout block they had never seen, all four score worse than a constant predictor, and the Python model somebody could write in an afternoon reproduces that same holdout exactly, 384 of 384 cycles. That is not a defect in the tool. This datapath is a four-tap filter with a decimator and a saturating rail: exactly computable, which is the wrong shape of problem for a neural surrogate. A ROM earns its place where the solver is expensive and the dynamics are continuous. We are telling you it is the wrong rung for this article."
The gate moving live (min 82–86)
"Watch the gate. We just fed it a real thermal result, 110 degrees against a 105 limit, and the record recomputes in front of you. Nobody edited a slide; the engine read a solver log and changed its own answer. Now watch what did not move. We added a second co-simulation tool to this record and co_analysis is still three of six, because coverage counts requirement-and-technique pairs and both tools answer the same three requirements. A number that refuses to move when you add a tool is the proof that it is not theatre."
5If it overruns
Cut in this order, and the order is deliberate:
- Stage 4 Capital, 64 to 70. Their Stage 4 ask, but Capital itself did not run here, so it is the weakest evidence in the room.
- Stage 1b matrix depth, 29 to 37, down to showing the saved matrix and moving on.
- Stages 3, 5 and 7, 55 to 64, down to Questa alone. Emulation depth is ground we do not win anyway.
Never cut 45 to 55, 70 to 77, or 77 to 82. Those are the thermal failure, the ladder and the requirements defect: the argument, the credibility and the proof respectively.
The readiness engine and its data are the floor. Nothing below that line can fail on the day: it is complete, tested, and independent of the demonstration tier.