← OrreryThe Ecliptic is an Orrery publication
AN ORRERY PUBLICATION THE ECLIPTIC The frontier of technology and the economics that price it
Issue No. 017November 6, 2026An Orrery publication · published weekly
From the Editor

This week two desks found that the number being reported is not the thing being bought.

Our cover story, from Parameter, follows software review once the person leaves the approval step. On 28 October GitHub let Copilot's approval count toward a repository's merge rules, a month after agents started more than 40% of its pull requests, 110 million in September. Firms that already merge with fewer people report falling revert shares. Outside studies of 2025 agent code find more corrective work instead: merged agent changes drew a verified follow-up fix at 1.62 times the odds of human ones, and 69.6% of those fixes came from the same agent. The two records count different things. A revert needs a person to notice, and a defect fixed forward never shows up as one. Among the features, Joule reads Meta's deal for certificates tied to about 200,000 tonnes of low-carbon cement, while the cement itself is sold locally at market rates. What the certificate covers is mostly the price of lime, which a tonne of this cement needs 0.56 tonnes of.

In both, the headline figure is accurate. The question is what it measures.

Twelve features and eight briefings from across the network.

— The Ecliptic, Editorial Desk

Cover Story · Compute & Intelligence · Parameter

Parameter — Quality After the Human Leaves the Gate — November 4, 2026

GitHub made machine approvals count toward merge rules on 28 October, the month after agents started more than 40% of its pull requests. The firms that already merge without a person report falling revert rates, while outside studies of 2025 agent code find more corrective work, more high-severity findings and a maintenance burden that rises with the share of unreviewed changes. The gap is in the metric: a revert needs a person to notice, and the fixes for agent code are increasingly written by the same agent.
What happens to software quality when the human approval step is taken out? PAR-0024 found agent code as good as human code inside a firm where agents waited on a person 80% of the time, and asked what would follow once that person left the gate. This report reads the six weeks since: GitHub's general release of machine approvals that count toward merge rules (28 October 2026), its Octoverse count of 110 million agent-started pull requests in September, Anthropic's account of risk-tiered approval in the Claude Code team, and three 2026 studies of what happens to agent code after merge. Firms that merge with fewer people report falling revert shares (Anthropic: down more than 40% while merges more than doubled). Outside studies of 2025 agent code find 49–51% higher hazards of corrective and bug-fix maintenance, 1.62 times the odds of a verified follow-up fix, and an agentic maintenance burden that rises about 6% per 10 percentage points of unreviewed merges. The primary finding is that the two records disagree because they count different things. The mechanism is that a revert needs a person to notice, while defects in agent code are increasingly fixed forward, 70% of the time by the same agent. The secondary finding is that the safeguards behind the good in-house numbers are not the product default: GitHub's approval counts as a person's, Anthropic's rule is that it never does.

Read the Graphic Edition →  

Power & Energy

Flagship  Joule

Joule — Meta Buys the Carbon, Not the Cement; a Lime-Fed Cement Plant Needs About $55 a Tonne Above Market at Today's Lime Price, and the Carbon Tax Credit Could Cover Two-Thirds of It — November 6, 2026

What Meta's certificate deal with Fortera pays for, why a cement made from recarbonated lime is priced by the lime market rather than the cement market, how far 45Q and book-and-claim certificates close the gap, and why the first commercial plant sits at a lime works rather than a cement works
On 29 October Meta agreed to buy environmental attribute certificates for about 200,000 tonnes of cement from Fortera's planned 300,000 t/yr plant at Mississippi Lime's Ste. Genevieve works in Missouri, while the physical cement is sold locally at market rates. This Flagship asks what that certificate has to…
2026-11-06 · ~14 pp
Flagship  Beyond Solar

Beyond Solar — Ocean Thermal's Labelled Net Record Is Still 103 Kilowatts — November 2, 2026

Of 15 ocean thermal projects on the record, three have a published net figure, and measured plants kept 20–48% of their gross power against 60–72% in every design from 1 MW to 80 MW; the pump load behaves as a fixed charge against a gross output that grows with the square of the temperature difference, so size does not fix the ratio and each degree moves net output by 12–17%; and the $0.19 per kWh quoted for a 1 MW plant excludes the intake pipe, which takes it to $0.41–1.11
Ocean thermal energy conversion is usually described by turbine ratings. This report asks how much of that power leaves the plant, and at what cost. It builds a register of 15 projects from the Ocean Energy Systems white paper, the IEA-OES economics report, Okinawa Prefecture's FY2024 study and…
2026-11-02 · ~13 pp
Flagship  Joule

Joule — A US Heat Pump Saves $210–1,190 a Year Against Oil, Propane and Electric Heat and Loses to Gas Nearly Everywhere; the Largest Pool of Savings Is in Homes Their Occupants Do Not Own — October 31, 2026

Running-cost break-even by fuel and region at this winter's prices, measured cold-climate performance, how many homes have the ducts and air conditioner for the cheap route, why the oil case depends on an oil shock that may not last, and what electrifying New England's oil homes does to its fuel use on a 5 °F day
TIB-0087 found that US heat pumps lose to gas furnaces on running cost, yet they outsell gas-fired equipment. This Flagship asks where their economics really sit. Using EIA's winter 2026–27 prices, a new cross-tabulation of 18,496 RECS household records, measured seasonal COPs of 2.34–3.41 and published installed costs,…
2026-10-31 · ~16 pp
Flagship  The Maple Grid

The Maple Grid — The War Premium Stops at Hardisty — October 24, 2026

Seven months into the Hormuz shock, WTI has averaged about US$91 since April and Alberta now forecasts C$9.7 billion more resource revenue than it budgeted. But record tanker freight has shut the Gulf Coast outlet for Canadian heavy crude, and the WCS discount has widened to US$25.55, about US$10.75 beyond Alberta's forecast. On the province's own sensitivities, a dollar of widening costs about as much as a dollar of WTI adds. The same shock lands hardest on Atlantic households that heat with fuel oil.
Since war broke out in the Middle East in late February 2026 and traffic through the Strait of Hormuz was disrupted, Canada has been on both sides of the oil shock. This report traces the shock through Canada's ledgers from primary documents: provincial fiscal updates, Alberta's own revenue…
2026-10-24 · ~14 pp
Flagship  Joule

Joule — Grid-Battery Failures per GWh Fell 98% While the Yearly Count Rose, and the Rate Will Keep Falling Even if Designs Stop Improving — October 14, 2026

EPRI's new failure statistics say lithium-ion storage is now about as fire-prone, unit for unit, as a large oil-filled transformer. The data behind the headline show something narrower and more useful. The 98% fall is a fleet that grew about a hundredfold against an incident count that rose by about two-fifths; each newer vintage really is safer; and because the per-GWh rate is a fleet average, it will keep falling for years even if the products being installed today stop getting better. The number that tells a permitting authority whether storage is getting safer is the count per year and the rate for each new vintage, not the fleet average
In October 2026 EPRI published, and the trade press reported, that the global failure rate of lithium-ion grid batteries fell 98% per deployed gigawatt-hour between 2018 and 2025, to a level comparable with large oil-filled transformers. This report asks what that number measures and what it will do…
2026-10-14 · ~15 pp

Compute & Intelligence

Flagship  Parameter

Parameter — The Humanoid Data Market — September 24, 2026

Humanoid robots are now being trained on paid human labour filmed at home and at work. One day of one crowdsourcing app now exceeds, in hours, every robot-collected dataset this report can find — and the constraint has moved from how much data exists to how much of it a robot body can use.
What now limits the capability of general-purpose humanoid robots — hardware, compute, or data? This report assembles 2023–2026 disclosures from robot-learning research, the two largest data-collection programmes, a robot developer's zero-shot evaluation and the 2026 shipment record, and finds that data is the binding constraint and is being…
2026-09-24 · ~9 pp

Frontier & Flight

Flagship  The Reality Gap

The Reality Gap — The Denominator Problem: Inertial Fusion's Driver Efficiency — October 14, 2026

Laser fusion's headline number is target gain. The number that decides whether a plant ever sends power to the grid is the efficiency of the machine that drives the target, and the only full-scale measurement of it in the public record is about 0.7%
How efficient does the machine that drives an inertial-fusion target have to be, and how close is anyone to building it? TAC-0023 found that the National Ignition Facility's record target gain of 4.13 sits beside a grid draw of about 300 MJ per shot. This report applies the…
2026-10-14 · ~12 pp
Flagship  Engines of Tomorrow

Engines of Tomorrow — The Reuse Tax, Measured Vehicle by Vehicle — October 11, 2026

Twelve launch vehicles publish what reuse costs them in payload. Set side by side, the payload lost runs from 13% to 70%. Above about ten tonnes it does not fall with vehicle size. It is set by where the booster lands and by whether the upper stage comes home, and the tax a launcher can afford depends on how much of the rocket is still thrown away.
MACH-0022 argued that the payload lost to reuse hardware shrinks as a launch vehicle grows, and proposed testing the claim against published payload pairs. This report runs that test. We collected every expendable and reusable payload pair we could source, for twelve vehicles from 0.3 t to 33.5…
2026-10-11 · ~20 pp
Flagship  Delta-V

Delta-V — The Reentry Mass Ledger — October 10, 2026

About 1.6 kt of spacecraft and rocket hardware reached the top of the atmosphere in 2024 and an extrapolated 2.35 kt in 2025, most of it rocket stages. The satellites already in orbit commit several times today's satellite flux, a 100,000-satellite fleet would inject about three times what meteoroids do, and since 30 September no US licensing step looks at any of it.
DLV-0021 ended on a quantity no licence measures: the mass of spacecraft returning through the atmosphere each year. This report builds that ledger from an open-access elemental inventory for 2015-2025, ESA's 2026 environment report, the largest operator's own disposal filing and an independent satellite catalogue, then tests each…
2026-10-10 · ~16 pp

Markets & Digital Assets

Flagship  Chain Reaction

Chain Reaction — The Freeze Followed the Issuer — November 4, 2026

On 29 October attackers had Binance's own signing service sign away about USD 6.3-6.4 billion across ten or eleven chains, most of it within about 20 minutes. About 44% has been frozen or seized. The share was set by what was stolen, not by how hard anyone chased it: about 77% of the stolen stablecoins were frozen and about 6% of everything else, close to the 5% Bybit got back from an all-ETH theft. One EIP-7702 delegation took 39% of the loss in a single transaction and stayed on the hot wallet for almost four hours, because the only party able to revoke it was the signing system that had just been subverted. For about two of those hours the fastest brake on the victim's own address was a stablecoin issuer's blacklist
Why was about 44% of the largest theft on record frozen within days, when Bybit's 2025 loss to the same North Korean group stands near 5% recovered or frozen? This report reconstructs the 29 October 2026 Binance theft from Binance's initial findings as reported, the on-chain timelines of…
2026-11-04 · ~11 pp
Flagship  Chain Reaction

Chain Reaction — The Gas Limit Triples, the Room for State Does Not — October 13, 2026

Ethereum's Glamsterdam upgrade went live on the Sepolia testnet on 6 October with a 200-million gas target, 3.3 times mainnet's. The same fork moves state creation onto a second gas meter priced at 1,530 gas a byte, and on that meter a 200M block holds less new state than a 60M block can today. The execution lane starts about a fifth full and the state lane between a fifth and three-quarters full; at a base fee near 0.07 gwei a new account still costs about three cents. The repricing is a quantity cap that scales with the gas limit, and it buys between one and five years before the state size at which clients were measured to degrade
What does Ethereum's planned move from a 60-million to a 200-million gas limit buy, and for which resource? This report reads the Glamsterdam specifications that activated on the Sepolia testnet on 6 October 2026 — principally EIP-8037, which meters state creation separately at 1,530 gas per byte —…
2026-10-13 · ~17 pp

The Briefings Well

The Power Bill · 2026-11-06
The Power Bill — PJM Is Asked to Buy Capacity Two Months Out, Nuclear Uprates Price Near $6,500/kW, and Arizona's Load Turns Flat — November 6, 2026
1. Exelon asks for a capacity auction held two months before delivery, and Invenergy asks to be paid for capacity MISO already counts 2. DOE lends $10.1 billion for 3.7 GW, and the disclosed…
The Maple Grid · 2026-11-06
The Maple Grid — The Discount Rail Cannot Fix — November 6, 2026
1. The discount doubled at Hardisty, but it was set in Houston 2. Rail did not come back, and on the right spread it should not have 3. Venezuela has taken the Gulf Coast…
Faraday · 2026-11-06
Faraday — The Window Is the Material — November 6, 2026
1. Lithium-manganese-rich cathodes buy their cycle life with the voltage window, on the way to 2028 production 2. LiNiO₂ keeps its capacity by cutting its strain, which is the other route past NMC 3.…
Joule · 2026-11-04
Joule — Gas Doubles but a German Battery Earns 39% Less, NERC's Data-Centre Rules Measure Before They Mandate, and PJM's Backstop Shrinks 91% on a 4% Forecast Cut — November 4, 2026
1. Gas more than doubled, yet a German two-hour battery benchmark earned 39% less than a year ago; trading revenue tracks gas, but the reserve market that paid most of the bill is filling…
Engines of Tomorrow · 2026-11-04
Engines of Tomorrow — The Engine Was Not the Hold-Up — November 4, 2026
1. Spectrum reaches orbit on propane, and its first flight was lost to a valve, not an engine 2. Themis may not hop in 2026, four years after Prometheus first fired 3. Neutron slips…
The Reality Gap · 2026-11-02
The Reality Gap — The Best Bionic Eye Shows About Four Letters at a Time, at Any Zoom — November 2, 2026
1. PRIMA's zoom trades field for sharpness along a fixed line of about four letters across 2. Without zoom the chip roughly matches where participants started; zoom supplies the gain 3. The retinal register:…
The Gas Gauge · 2026-11-02
The Gas Gauge — The Surplus Survived October, and Most of the Federal Gap Was Alaska — November 2, 2026
1. **The surplus stopped falling at 35 Bcf and stood at 38 Bcf on 23 October.** The range set here on 21 October will very likely be missed, and half of the turn came…
Joule · 2026-11-02
Joule — GE Vernova Pays $9 Billion for a Grid-Equipment Maker as Turbine Bookings Slow to 6 GW, China's Solar Build Falls 54% Over About 240 GW of Polysilicon in Store, and Britain's 7.6 GW Storage Round Draws a £1.6 Billion Challenge — November 2, 2026
1. GE Vernova pays $9 billion for a bushing and instrument-transformer maker in the quarter its turbine book grew only 6 GW; the slowdown looks more like sold-out slots than weak demand 2. China's…
The Ecliptic assembles the week's research from the Orrery network of publications. Nothing here is investment advice.
The Ecliptic™ is an Orrery publication. Content is produced by AI systems with human editorial oversight.
© 2026 Orrery  ·  All issues