ID: MC-AGENTIC-ENGINEERING-2026-09-16 · Status: draft, local file + published HTML (Frank deselected DB persistence on 2026-09-16, Q4) · Owner: Teddy (trend stack), research on the Diesel/Codex seat, Ralf for company-level evidence, Frank for gates · Body language: EN (myConcept standing rule)
Published: https://phibel.app/uploads/concepts/agentic-engineering-myconcept-2026-09-16.html · Evidence page: https://phibel.app/uploads/concepts/agentic-engineering-evidence-2026-09-16.html
Introduce one new megatrend, MT13 "Agentic Engineering", into the Moonshots trend tracker (trends.phibel.app) with 12 trend lines (AE.1–AE.12), so that the DACH engineering segments read a dated, falsifiable picture of what happens when design, meshing, simulation, manufacturing and testing are run by agents and physics-AI models rather than by engineers at a workstation.
The trigger is the interview with Johannes Brandstetter (Vice President AI for Science, Mistral; founder of Emmi AI, acquired by Mistral in May 2026) — https://youtu.be/9D5MV-4UWcM — and four surrounding facts of September 2026: OpenAI's Navier-Stokes blow-up proof (08.09), Emmi's AB-UPT surrogates at 100 million mesh cells with CAD-direct inference, LEAP 71's Noyron producing hot-fired rocket engines from a specification with no human design step, and Prometheus (Bezos/Bajaj) raising 12 bn USD at a 41 bn USD valuation to build an "artificial general engineer".
Scoping sentence: strictly the trend stack — line definitions, unit contracts, evidence lanes, the megatrend page and its reading for three segments. The Pulse radar, the LinkedIn waves, Trend 3.0's Engpass research and the Inploration site get at most one sentence each. Nothing in this concept publishes to trends.phibel.app; publishing is red (publish-public) and stays a Frank click.
Tracker: ~/apps/trends-site/data/trends.json carries MT4–MT9, MT11, MT12 with 30 lines (26 with points, 3 defined without a point: RG.1, SU.1, WS.8). No megatrend and no line covers CAD, meshing, CAM, CFD, structural mechanics, sensor-coupled simulation or physics-AI. The closest lines are T4.1 (autonomous task horizon, hours), T4.2 (share of AI-written code), T4.4 (agent economy, USD), TC.1 (GPU cluster size), WS.8 (additive manufacturing cost per reference part). MT13 is the next free id.
Contract for a new line (measured 02.09, unchanged): the refresh lane codex-trend-line.sh resolves an id either from the live file or from the candidate register ~/ai-cowork/managers/drafts/teddy-2026-09-02-neue-linien-definitionen.json (registry key kandidaten, CR-2026-194). A candidate id is researched with candidate=true and the register's sha256, never a live hash. Merge gate trend_merge.py is fail-closed on a missing or mismatching unit. The register holds 3 lines today (SU.1, RG.1, WS.8), all with series: [].
Engines (16.09 16:05 CEST): Codex week 56 %, band decision catchup, green. Claude week 41 %, yellow (tandem throttled, Claude for judgement only). GLM 5-hour window at 100.3 %, reset in ~2 h — GLM legs pause.
Segments: live pulse pages exist for dach-additive-fertigung (67), dach-antriebstechnik (12), dach-marktforschung (92); dach-batteriespeicher has 0 live pages; a CAE / engineering-office segment does not exist in the segment canon.
Evidence available today (all CLAIMED grade until a lane opens the source):
| Fact | Value | Source | Date |
|---|---|---|---|
| Navier-Stokes forced 3D blow-up proof (Lean-formalised), unforced case open | announced | openai.com/index/navier-stokes-solution | 2026-09-08 |
| Clay Mathematics Institute statement | "apparently been settled", credit unhurried | press reports | 2026-09-11 |
| Emmi AB-UPT surrogate scale | > 100 M mesh cells, CAD-direct inference | emmi.ai/news/ab-upt-scaling-neural-surrogates-100m-cfd-meshes | 2026 (month to date) |
| Emmi → Mistral | team of > 30 joined Mistral Science / Applied AI | emmi.ai/news/mistral-ai-acquires-emmi-ai | 2026-05 |
| LEAP 71 20 kN methalox engines, spec → hot fire | < 3 weeks, bell + aerospike | 3dprintingindustry.com | 2025-12 / 2026 |
| LEAP 71 × HBD aerospike XRA-2E5 | 200 kN, printed monolithic in 289 h | leap71.com/2026/03/12 | 2026-03-12 |
| Exploration Company licenses Noyron RP | licence | leap71.com/2026/03/24 | 2026-03-24 |
| Prometheus raise | 12 bn USD at 41 bn USD; earlier 6.2 bn | techcrunch.com/2026/06/11 | 2026-06-11 |
| SimScale agent | open to ~900 000 users | simscale.com/press | 2026-08 |
| SimScale survey | ~90 % of 350 leaders building governance for autonomous agent action | simscale.com/blog | 2026 |
| Text-to-CAD independent test | Zoo fails gear/manifold, Adam passes gear | getleo.ai, xometry.pro | 2026 |
Brandstetter's claims (transcript, ASR-flagged, minute marks on the evidence page): three phases design → test → PLM, the first two on the computer for 50 years; "agentic engineering" reaches the pace of agentic coding "in a year or two"; the three time buckets are solver time, tooling/analysis time and iteration count — the gain is in the second and third; the unsolved research question is which data to generate, because mixed assumptions teach a model "an average of physics"; three pillars (language models + orchestration, physics models + data, real industrial problems); three "very big predictions" for engineering physics, verifiable "in the next year or year and a half" — not stated.
| # | Question | Frank's answer | Deselected (do not propose again, §8) |
|---|---|---|---|
| Q1 | Additional subtrends beyond the six fixed ones and the math line | Spec-to-hardware autonomy · Text-to-CAD · Frontier-lab capital race · Governance of autonomous engineering action · Physics-AI foundation models & data (fifth, by recommendation) | — |
| Q2 | How the Navier-Stokes line measures impact | A proof-informed solver releases · B formalised numerics · C dated bets in the forecast ledger · D verification status as contra-line — all four | — |
| Q3 | Audience | dach-additive-fertigung · dach-antriebstechnik · dach-batteriespeicher · new segment CAE / engineering service offices (needs a segment contract, card D1) | dach-marktforschung and dach-b2b-fachmedien as target segments (one sentence each) |
| Q4 | Delivery | B file + published HTML · C 12 line definitions into the candidate register · D video-insights evidence page | A DB persistence /concept/ |
| Role | Engine | Used for |
|---|---|---|
| RESEARCH | Diesel (Codex seat) via codex-trend-line.sh AE.n, GLM lane when its window is open |
first dated points per line, source opening, contra-sources — 0 Claude tokens |
| BUILD | Codex via codex-task.sh |
register schema checks, prediction ledger fields, tripwire arming, page module |
| JUDGE | Claude (Fable 5.1, effort high) | unit contracts, closest-line duplicate filter, 10 % sample of accepted points, bet calibration |
| MECH | Python collectors / cron | drafts check 06:14, merge dry-run, freshness gate |
| GATE | Frank | publish MT13 live (D2), new segment (D1), source list change (D3), pulse integration (D4) |
Runtime models of the tracker itself are untouched; the measuring instrument does not change while a line is being introduced.
Rule from 02.09: a new line must cite the closest existing line id and say why it does not cover the driver — otherwise it is a duplicate, not a finding. Counting unit before any ratio. All lines start with series: []; a line goes live only with at least one dated, sourced point (M5 blind-line rule).
| Id | Name | Unit | Closest line, why not covered | First source to open |
|---|---|---|---|---|
| AE.1 | Text-to-CAD: pass rate on medium-complexity parts | % korrekt (24-Zahn-Zahnrad + Verteiler-Testsatz) | T4.3 measures software-agent benchmarks, no geometry | getleo.ai / xometry.pro hands-on tests; neuralCAD-Edit benchmark |
| AE.2 | Automated meshing: CAD → simulation-ready mesh | Minuten je Referenzgeometrie | none — T4.1 measures hours of autonomy, not the meshing step | Ansys Fluent Meshing, Simcenter, Cadence Pointwise release notes; SimScale agent |
| AE.3 | Autonomous CAM: model → released NC program | Minuten je Referenzbauteil | WS.8 measures cost per printed part, not programming time | CloudNC, Toolpath, Autodesk Fusion automation docs |
| AE.4 | CFD neural surrogates: scale per inference | Mio Netzzellen (log) | TC.1 counts GPUs, not solved cells | Emmi AB-UPT (100 M), NVIDIA PhysicsNeMo, Luminary, Navier AI |
| AE.5 | Structural-mechanics surrogates: speed-up vs. solver | Faktor Solverzeit/Surrogatzeit (log) | none — no line measures FEM/crash | Emmi structural cases, Altair physicsAI, Ansys SimAI |
| AE.6 | Sensor-coupled simulation loops: measurement → model latency | Sekunden Latenz (log) | TR.1 measures robot autonomy duration, not twin latency | Siemens digital-twin cases, Frontiers digital-twin surrogate paper 2026 |
| AE.7 | Proof-informed solvers after Navier-Stokes (Q2 A) | Releases + Paper mit Blow-up-Skalierung (Anzahl, kumuliert) | AGI.1 measures math benchmark %, not solver adoption | OpenFOAM, Ansys, Siemens, Cadence changelogs; arXiv 2603.28308 line of work; cmccomb proof-informed simulation repo |
| AE.8 | Formalised numerics: Lean-formalised fluid-mechanics results (Q2 B) | Resultate (Anzahl/Quartal) | AGI.1 — benchmark, not formalisation | Lean mathlib / OpenAI formalisation, Clay statements |
| AE.9 | Spec-to-hardware autonomy | Tage Spezifikation → erster Heißlauf | WS.8 cost per part; TR.2 robot production — neither measures design autonomy | LEAP 71 hot-fire posts (< 21 d), Exploration Company licence |
| AE.10 | Physics-AI foundation models and open frameworks | veröffentlichte Modelle/Frameworks (Anzahl, kumuliert) | T4.5 counts deployed software agents, not physics models | Emmi Noether (open source), AB-UPT, neuralDEM, PhysicsNeMo, Mistral physics-AI research page |
| AE.11 | Frontier-lab capital for "general engineer" labs | USD Mrd kumuliert | EU.2 measures the capex gap the EU does not match; T4.4 agent-economy value — neither counts engineering-lab capital | Prometheus (6.2 + 12), Mistral × Emmi (three-digit m EUR, undisclosed), SimScale, Synera rounds |
| AE.12 | Governance of autonomous engineering action | % Firmen mit Freigabe-Gate für Agenten-Änderungen | T4.6 workforce impact — headcount, not sign-off rules | SimScale State of Engineering AI 2026 (n = 350), VDI/VDMA position papers |
Q2 C and D live on the megatrend, not as lines: three dated wagers in predictions[] of MT13 (see M06) and a contra-source field on AE.7 (verification status of the OpenAI claim as bull/bear spread; no tripwire fires on AE.7 while the spread is wide).
M01 · Register MT13 + AE.1–AE.12 in the candidate register — Prio 1 · Impact H · Effort S · green (draft-content). Append the 12 definitions with unit, was_ist_ein_datenpunkt, abgrenzung_zur_naechsten_bestehenden_linie, quellen_kandidaten (auto-ingestable, contra, base rate), falsifikator_der_linie, series: []. DoD: python3 -m json.tool passes, TRENDLINE_DRYRUN=1 codex-trend-line.sh AE.n prints KANDIDAT for all 12, live sha256 of both trends.json files unchanged. Done 2026-09-16 in this session (see §9).
| # | Task | Model | Type | Est |
|---|---|---|---|---|
| 1 | Write 12 definitions with all mandatory fields | Claude | JUDGE | done |
| 2 | Validate JSON + dry-run 12 ids against the lane | Python | MECH | done |
| 3 | Drafts check cron 06:14 reads the new ids without error | cron | RUN | next morning |
M02 · Unit contract review by a second reader — Prio 2 · Impact H · Effort S. One Codex review pass over the 12 units against the merge gate's verbatim rule (lesson 02.09: 10 of 11 points rejected on prose units). DoD: review file lists 0 open unit ambiguities or the unit is rewritten before any research job starts.
M03 · Closest-line duplicate filter, signed — Prio 2 · Impact M · Effort S. For each line the cited closest id must exist in the live file (script check) and the "why not covered" sentence must name a measurable difference. DoD: 12/12 pass; a failing line is renamed or dropped, not argued.
M04 · MT13 megatrend record — Prio 1 · Impact M · Effort S. neue_megatrends[] entry: id MT13, name "Agentic Engineering", icon "cpu", overallCurve "s-curve (unbelegt)", verdict HYPOTHESE, euLens pointing to MT7 (Brandstetter: "Europe is historically an engineering continent", the "two years" claim by Arthur Mensch is a MT7 datapoint, not MT13). DoD: record present, referenced by all 12 lines. Done 2026-09-16.
M05 · Contra-source field on AE.7 — Prio 2 · Impact H · Effort S. Verification status of the OpenAI proof as a dated series of statements (Clay 11.09 "apparently settled"; referee reports; rebuttals), rendered as bull/bear spread. Rule: while the spread is wide, AE.7 cannot arm a tripwire. DoD: field schema accepted by trend_merge.py --dry-run.
M06 · Three dated wagers in the MT13 forecast ledger (Q2 C) — Prio 1 · Impact H · Effort S. (1) "By 2027-09-30 a shipped CFD solver (Ansys, Siemens, Cadence, OpenFOAM) contains a limiter or refinement stop that cites the blow-up scaling." (2) "By 2027-03-31 the unforced 3D case has a peer-accepted result in either direction." (3) "By 2027-12-31 a DACH tier-1 supplier reports a design-iteration count reduced by ≥ 5× through surrogates, with a named part." Each carries probability_by teddy, falsifier, Brier-scored per source. DoD: entries in predictions[], D3 decides whether Frank sets his own probabilities alongside.
M07 · First research wave: 12 lane jobs on Diesel — Prio 1 · Impact H · Effort M. bash ~/ai-cowork/codex-trend-line.sh AE.n per line, candidate mode, night band; yield measured as ACCEPTED points (merge dry-run), never as job status. Measured pilot (AE.9, this session): see §9. DoD: ≥ 1 accepted dated point on ≥ 8 of 12 lines, or the failing lines are named with the reason.
| # | Task | Model | Type | Est |
|---|---|---|---|---|
| 1 | Run AE.9 as pilot, measure wall-clock and accepted points | Codex | RUN | measured §9 |
| 2 | Queue AE.1–AE.12 for the night band (setsid, DONE marker) | bash | RUN | 12 × pilot time |
| 3 | Merge dry-run on all drafts, list rejections by reason | Python | MECH | 5 min Loco |
| 4 | Claude judges a 10 % sample: source opens, number verbatim, unit verbatim | Claude | JUDGE | 15 min Loco |
M08 · Evidence page from the interview (Q4 D) — Prio 1 · Impact M · Effort S. Five insights with minute marks, each with "our measurement" and "docking point" columns (video mechanic | our measurement | where it lands), ASR warning, plus the September sources as citable belege. Done 2026-09-16: https://phibel.app/uploads/concepts/agentic-engineering-evidence-2026-09-16.html
M09 · Source portfolio by modality (A4) — Prio 3 · Impact M · Effort M. Each of the 12 lines gets at least one auto-ingestable source (release notes RSS, arXiv query, GitHub release tags, press-release feeds) registered in the line's quellen_kandidaten with rolle: auto-ingestierbar. DoD: 12/12 have one; the drafts check reports which are still model-only.
M10 · Contra-sources and base rates (A5) on the four hype-prone lines — Prio 2 · Impact H · Effort S. AE.4 (surrogate accuracy loss, Brandstetter's own "how close to reality"), AE.9 (one company, n = 3 engines), AE.11 (valuation ≠ delivery), AE.12 (survey self-report). Each gets a named contra-source and a base rate (e.g. share of CAE surrogates that ever replaced a certification run). DoD: fields filled, bull/bear spread visible on the page.
M11 · Tripwires (M1) armed only on MEASURED points — Prio 3 · Impact M · Effort S. Draft one Wecker per line (e.g. AE.9 < 7 days spec-to-fire; AE.4 ≥ 1 bn cells; AE.1 ≥ 80 % on the medium set; AE.11 ≥ 50 bn USD cumulative), state drafted; arming (green, monitoring-add) happens after M07 delivers MEASURED points. DoD: 12 tripwires in tripwires.json, each with trigger/action/status.
M12 · Owner/CEO voice per line — Prio 1 · Impact H · Effort M. Every line carries a two-sentence "what this means for a 50–500-employee owner" reading in its note: cost per design iteration, lead time, which billable hours move. Lesson 02.09: the tracker is a technology map read by business owners (97 of 160 contacts). DoD: 12/12 notes reviewed by Karl (customer stand-in) with pass/fail per line.
M13 · Segment mapping AE × {additive, antriebstechnik, batteriespeicher} — Prio 2 · Impact M · Effort S. Apply the fixed segment rule (driver appears in the segment's own value chain, not "interesting for"): AE.3/AE.9/AE.10 → additive; AE.2/AE.4/AE.5/AE.6 → antriebstechnik; AE.4/AE.5/AE.6 → batteriespeicher; AE.1/AE.7/AE.8/AE.11/AE.12 cross-segment. Write into kkt-segment-map.json per contact type with rationale. DoD: coverage_kkt.py shows MT13 coverage per segment × contact type; marktforschung and b2b-fachmedien get exactly one sentence.
M14 · New segment "CAE / engineering service offices" — contract, not build (card D1) — Prio 2 · Impact H · Effort M · red. Frank chose it in Q3. Before any line is tagged with it, the segment needs its own canon entry, contact-type buckets and a first list of firms (Ralf). This measure delivers the contract draft as a CR with zone_reason; nothing is tagged until approved. DoD: CR proposed, D1 answered.
M15 · Go-live gate for MT13 (card D2) — Prio 1 · Impact H · Effort S · red (publish-public). Preconditions measured, not assumed: ≥ 8 lines with ≥ 1 MEASURED point, all 12 with unit + closest line + owner note, contra field on AE.7 filled, Karl pass on notes, D1 answered. Then one CR to merge MT13 into trends.json and the dashboard mirror (both files, sha-checked). DoD: Frank's click on the CR; until then MT13 exists only in the register and on the concept page.
| Wave | Content | Gate | Effort |
|---|---|---|---|
| W0 (done 16.09) | M01, M04, M08, pilot of M07 | none (green) | Loco time this session |
| W1 (night 16./17.09) | M07 12 jobs, M02, M03, M05 | Codex band catchup green |
Diesel night band |
| W2 (17.–19.09) | M06, M10, M11, M12, M13 | Karl review on M12 | Claude judge ≤ 30 min Loco |
| W3 (after wave 1, 16./17.09) | M14 CR (D1 = A, decided 16.09), M09 | CR approval | Ralf + Codex |
| W4 | M15 | D2 (Frank) | one click |
| KPI | Baseline 16.09 | Target |
|---|---|---|
| Lines defined with unit + closest line | 0 → 12 (done) | 12 |
| Lines with ≥ 1 MEASURED point | 0 | ≥ 8 by 2026-09-19 |
| Accepted points from the first wave (merge dry-run) | not run | ≥ 12 (≥ 1 per line) |
| Lines with an auto-ingestable source | 0 | 12 |
| Wagers in the ledger with probability + falsifier | 0 | 3 |
| Tripwires drafted / armed | 0 / 0 | 12 / ≥ 8 |
| Segment × contact-type coverage cells served by MT13 | 0 | ≥ 9 (3 segments × 3 buckets) |
| Owner notes passing Karl | 0 | 12 |
/concept/<id> for this concept (Q4 A).trends.json or the dashboard mirror from this concept; any live line without a dated point.neue_megatrends[]; backup of the previous register alongside; JSON validated.trends-drafts/AE.9.json with 2 points, unit check OK (both 2025-12-11, 21 days as the conservative upper bound of "under three weeks", bell + aerospike, source leap71.com). ⚠ Finding: trend_merge.py merge --dry-run counts candidate drafts under drafts_without_matching_line (23 today, AE.9 among them) — the merge gate cannot measure yield for a line that is not live. For candidates the acceptance gate is the lane's unit check plus the Claude sample judge (M07 task 4); the merge dry-run becomes meaningful only after D2. M07's DoD is read accordingly.wave1.log, marker .wave1-DONE; expected ≈ 11 × 69 s if the pilot holds.~/apps/consulting-site/data/pipeline/inputs-leuchtturm/brandstetter-mistral-agentic-engineering-2026-09-16.transcript.txt.| Line | Points | Judge (source opened) | Fable review verdict |
|---|---|---|---|
| AE.1 Text-to-CAD | 0 | zero with reason | rule currently unsatisfiable by any source — loosen in M02 (accept single-part named tests as a series per part) |
| AE.2 Meshing | 0 | zero with reason | base-year artefact: a contract-clean point exists (Siemens STAR-CCM+ 2510, Corvette, 6.5 min, 93 M cells, 2025) — allow 2025 as base year |
| AE.3 CAM | 0 | zero with reason | keep rule, re-run with vendor case studies as allowed source class |
| AE.4 CFD surrogate scale | 1 | ACCEPT (140 M cells, errors verbatim) | MEASURED |
| AE.5 FEM speed-up | 3 | CANNOT_VERIFY (Springer full text blocked) | hold; three values ≈ 25 000× on one date read as a lab paper, not an industrial case — needs a second source |
| AE.6 Sensor loop latency | 1 | CANNOT_VERIFY (scope) | REJECT: 48.6 ms is data acquisition/transmission latency, not sensor → model-state update; wrong measurement for the rule |
| AE.7 Proof-informed solvers | 0 | zero with reason | correct zero — baseline 0 is the finding (nothing shipped yet) |
| AE.8 Formalised results | 1 | ACCEPT | MEASURED (2 results, 2026-09-08) |
| AE.9 Spec → hot fire | 2 | ACCEPT | MEASURED |
| AE.10 Physics-AI models | 4 | ACCEPT | MEASURED (cumulative 1→4, 2023-05 → 2026-01) |
| AE.11 Capital | 3 (after retry) | ACCEPT | MEASURED, but ⚠ two runs gave two series (Luminary counted / excluded) — the qualifying criterion needs one sentence in the register (M02) |
| AE.12 Governance | 1 | ACCEPT (79 %, n = 350, 2026-02) | MEASURED, self-report |
Totals: 16 points on 8 lines; after review 13 usable points on 7 lines, 6 lines with a MEASURED point (AE.4, 8, 9, 10, 11, 12); go-live threshold of D2 (≥ 8 lines) not yet met. Wall-clock: pilot 69 s, wave 16 min 08 s for 11 lines, AE.11 retry 3 min 19 s. AE.11's first run failed because Codex's sandbox saw shared-decisions.jsonl read-only and refused to write the draft — a sandbox artefact, not a research failure.
Segment CR: CR-2026-295 proposed (red, Frank's desk). ⚠ cr_lib replaced Teddy's zone_reason with its generic fail-closed sentence because no category was set — the reason Frank should read is card D1 above and this paragraph, not the queue entry. Finding for Julia (cr_lib owner): a supplied zone_reason must survive the fail-closed default.
Full report: ~/ai-cowork/projects/agentic-engineering-20260916/wave1-report.md.
CR-2026-295 built 18:04–18:13 (Opus on Codex, 3 packages, all exit 0), Fable review: canon entry + 9 contact buckets ACCEPTED; size-band report ACCEPTED (4 of 20 firms in the 50–500 band, 9 family-owned, 11 of 20 buy via CTO/R&D → M12 gets a CTO voice for this segment); the line mapping (AE.2/4/5/6/12 tagged) is a hard-coded table, not a rule output → PROPOSAL only, no tag applied. Coverage gate still red: a 4th consumer of the canon (kkt-verteilung-20260902.json, measured 02.09) needs a measured count from Ralf. Details in wave1-report.md.
Implementation 2026-09-16 20:42–20:54 (Frank: "Teddy implementiere, Emil Engine-Vorschlag, Kosten auf Trend"): Emil booked P1 on Claude (EB-20260916-183949-teddy, project trend), P2 as 0-token Python, P3 on the Codex seat. Teddy delivered P1 (rule-derived mapping, hard-coded table removed; derivation differs from the hand table on 4 of 12 lines — AE.4/5/6/9/12 tagged by rule, AE.2 not: meshing is invisible to a word-level rule; the divergence is recorded, the tag decision sits at card D2) and P2 (whole contact column re-measured from the tracking store: 246 contacts, engineering-services = 21; coverage gate exit 0, offline self-test green, 48 checks). P3's four lane runs (AE.1/2/3/7, contracts fixed) wait for the Codex weekly reset on 19.09 10:54 — the seat stood at 84 % against a stop at 85 and a foreign session was still writing. Finding for a governance CR: the protected-hash register of the segment lane has 3 of 4 stale hashes since 08.09 — proposed with zone_reason, Frank's desk.
D1 · New segment "CAE / engineering service offices" — ✅ DECIDED by Frank 2026-09-16 16:17: A, after wave 1 (segment contract drafted as a CR once wave 1 has finished; no line tagged before approval). What: Frank chose in Q3 a segment that does not exist in the canon (5 segments today). Stake: the segment is the group whose billable hours MT13 replaces; without it the trend has no reader who loses. Options: A create the segment contract now (Ralf builds a firm list, ~1 h Loco, then contact types) · B tag lines only to the three existing segments and revisit at W3 · C drop. Recommendation: A after W1, because the evidence wave does not depend on it and a segment without firms is a label. Reversible: yes (a segment can be retired). If undecided: MT13 goes live for three segments only; the CAE reading stays in the owner notes.
D2 · Publish MT13 on trends.phibel.app — What: merge the megatrend with its lines into both live files. Stake: public page, seen by customers via pulse links; a line without a point is a blind line. Options: A publish when ≥ 8 lines carry a MEASURED point · B publish all 12 as HYPOTHESE with empty series · C keep on the concept page only. Recommendation: A, the M5 rule exists for this. Reversible: technically yes, but a published blind line has already been read. If undecided: nothing publishes; the register keeps researching.
D3 · Probabilities on the three Navier-Stokes wagers — What: Teddy sets probability_by teddy; the ledger can also hold Frank's number per wager for calibration against sources. Stake: Brier scores are the only way this line's "impact on speed and efficacy" ever becomes a number rather than a narrative. Options: A Frank adds his three probabilities (Frank time ≈ 3 min) · B Teddy only · C no ledger. Recommendation: A, two forecasters make the ledger scoreable. Reversible: yes. If undecided: B.
D4 · MT13 on the pulse pages of the three segments — What: once live, the pulse generator can render MT13 cells with Gegenbeleg like the other megatrends. Stake: customer-facing material; the send gate (own source ≤ 90 d) applies unchanged. Options: A render for additive + antriebstechnik after D2 · B render only after ≥ 3 MEASURED points per rendered line · C never. Recommendation: B, one point is a claim, three are a direction. Reversible: yes. If undecided: A happens by default of the generator once MT13 is live — so this card must be answered before D2, or the generator gets an explicit exclude.