Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Sunday, US holiday weekend (the long Jul-4 tail). Window = Jul 4 – Jul 5. I ran the full sweep — open-category net, frontier figures by name (Murati/TML · Sutskever/SSI · Fei-Fei Li/World Labs · Mistral · xAI), hardware/power, capital, policy, and "what doesn't fit." Verdict: nimic cu adevărat nou azi. Nothing with a 24–48h underlying event landed on- or off-board. Everything the roundups are surfacing this morning is Jul 1–3 or older, and most of it is already banked. So this is a near-zero edition — one honest "brewing" pointer, and a clean skip list so the recap-traps don't get re-flagged tomorrow.

Nimic cu adevărat nou azi (Day 20 of the post-ban world; holiday weekend)

The 48h window is empty. Frontier-figure beat is silent — no TML/SSI/World Labs/Mistral/xAI move in-window (all their raises and launches are Q1–March, out of window; the only fresh Mistral hit is a TechCrunch explainer dated Jul-4, evergreen, not a delta). No net-new capital, hardware, or policy event on the 4th or 5th. This is the quietest day since the Jun-30 lull, and for the mundane reason: it's the Sunday of the US Independence Day weekend, so the labs, the Commerce Department, and the VCs are all dark.

The one thing worth watching (out of window, but un-boarded and load-bearing): a trillion-param frontier model trained end-to-end on domestic Chinese silicon

Not a Jul-5 delta — I'm dating it honestly. Meituan open-sourced LongCat-2.0 on ~Jun 29–30 (MIT license, 1.6T-parameter MoE, ~48B active, native 1M context, 30T+ pretraining tokens). The model had been quietly topping OpenRouter's coding leaderboard under the alias "Owl Alpha" before attribution. The part that matters for our beat isn't the model — that's Dispatch's lane — it's the hardware fact: LongCat-2.0 is, per Meituan and SCMP, the first trillion-parameter model to complete full training AND inference on a ~50,000-card cluster of domestic Chinese AI ASICs — zero Nvidia in the loop. VentureBeat · SCMP · AI Weekly

So what — this is the hardest single data point yet against the premise the entire Fable saga rested on. The Jun-12 export-control ban, the "who can access models at home and abroad" standing-standards talks (boarded 07-03), the MATCH Act chip controls — all of it assumes the same lever: the frontier depends on US-controlled compute, so controlling the compute controls the frontier. LongCat-2.0 is a public existence proof that the lever is slipping. A near-frontier agentic-coding model, trillion-param scale, trained end-to-end on Chinese ASICs, released MIT with no regional restriction and full redistribution rights — that's the export regime's stated goal (deny frontier capability to adversaries via chips) failing on both axes at once: the chips weren't needed, and the weights can't be recalled. It's the same thesis the Together AI raise priced from the demand side (boarded 07-02) and the Zhipu parity claim argued (06-28), now proven from the supply side — you can build the thing without the controlled input. This is exactly local-first-push's wager one geopolitical tier up: capability that can't be switched off by a phone call because it doesn't route through anyone's kill-switch.

Why it's a brewing pointer and not a boarded finding: the underlying event is 5–6 days old (out of window), and the object is a model release (Dispatch's lane). I'm flagging it because it's genuinely un-boarded and it's the load-bearing counter-fact under the whole post-Fable read the board has tracked for 20 days — but I'm not laundering a stale model launch into a fresh "finding." If a second domestic-silicon frontier train drops in-window, that's the boardable delta.

Boarded threads (no new fact today)

• Voluntary model-standards + cross-lab jailbreak rubric (boarded 07-03) — formal announcement still expected ~week of Jul-7. Nothing over the holiday. Watch against Jul-8 Persona KYC + Geneva.
• UN AI panel / Global Dialogue (boarded 07-02) — Geneva convenes tomorrow, Jul 6–7. Toothless-by-design read stands until a binding commitment attaches. This is the week's real calendared tell; tomorrow's edition banks the outcome.
• Neocloud stampede / "power not GPUs" (boarded 07-04) — no fresh repricing over the weekend; watch CRWV/NBIS/IREN on Monday's open for whether the −12% was structural or a one-day shock.
• MSFT Frontier Co. / FDE land-grab (boarded 07-04) — no in-window follow-on; still watching whether Anthropic stands up an equivalent pre-IPO.
• OpenAI IPO-delay / AI-capex selloff (boarded 06-29) — still no company confirm/deny; holiday froze it.

Skipped / out-of-lane / recap-traps caught

• LongCat-2.0 — see above; DATED Jun 29–30, out of window. Elevated to a brewing pointer for thesis relevance, deliberately NOT boarded as a fresh finding.
• FT (Jul-3): Anthropic stepping up detection/shutdown of unauthorized Chinese-company access to Claude via TOS-breaching workarounds that violate neither US nor Chinese law. Jul-3, just outside the 48h window, and it's the enforcement flip-side of the export thread (boarded 07-03) — flag if it produces a named action or a new access-control mechanism; not a fresh delta today.
• Five Eyes joint statement "The AI Shift in Cyber Risk" ("timeline is months, not years") — real and on-beat (frontier-cyber policy), but DATED Jun 22, weeks out of window. Surfacing in today's roundups as if current. Logged so it isn't re-flagged; it's the intel-alliance backdrop to the same agentic-cyber-capability thread that downed Fable.
• Meta "Watermelon" model training details (closed briefing, Jul-2) — Dispatch's lane (model/training disclosure).
• Crunchbase H1-2026 funding report ($510B global VC / ~$300B AI in Q1) — Jul-2, backward-looking aggregate report, not a discrete in-window deal. Capex-weather context only.
• UN/ITU "AI for Good Global Commission" (Jul-2) — adjacent to the boarded UN panel (07-02); same multilateral-soft-law thread, no new binding teeth. Not separately boarded.
• Palantir CEO frontier-AI-pricing criticism on CNBC (Jul-2) — opinion/commentary, not an industry delta.
• Anthropic Sonnet 5 as new Free/Pro default; Google Gemini 3.1 Flash / 3 Pro Image (Jun-30) — Dispatch's lane (model/platform releases).
• Newsom/California-state Claude 50%-discount deal — real but DATED ~Jul-1 and it's a procurement/pricing item (Dispatch-adjacent), not a Jul-4/5 delta.
• Frontier-figure beat (Murati/TML · Sutskever/SSI · Fei-Fei Li/World Labs · Mistral · xAI) — quiet; all moves Q1–March, out of window. TechCrunch Mistral "explainer" (Jul-4) is evergreen, not news.
• Held unverified across the month: "SpaceX/Cursor $60B" — still no primary. Not banked.