← The energy-first turn · The database

Keeping a Review Alive

A map of a moving field rots unless something keeps it honest.

Most training teaches you to read the literature. Very little teaches you to maintain a picture of a field that changes faster than you can publish, which is a harder problem with its own failure modes. What follows is a worked design for that job, drafted alongside the energy-first review and published as teaching material. It is not a description of a process running here: the map is a dated snapshot, and saying otherwise would be exactly the kind of unverifiable claim the review itself warns about.

You should leave able to

1. Design a review cadence that separates cheap frequent checks from expensive rare ones.

You should leave able to

2. Grade a claim's evidence consistently, and keep that grade attached as the claim travels.

You should leave able to

3. Write falsifiable markers in advance, so being wrong is detectable rather than deniable.

The cadence

Three passes, deliberately different in cost.

The common failure is a single review rhythm: either so frequent it gets skipped, or so rare the map is stale between runs. Splitting the work by cost addresses that. Cheap checks would run often against a fixed watchlist; expensive judgement rarely, and scheduled rather than improvised.

PassFrequencyScope
LightWeeklyRun the falsifiable markers and the standing queries; update only the rows they touch.
DeepMonthlyAll search lanes including non-English sources; add and retire entries; refresh the headline figures; check the conference calendar.
AuditQuarterlyReview what died, pivoted or was acquired; re-score the conditions; re-run the forecast sensitivity; check whether the thesis' own falsifiers have fired.

Note what the quarterly pass contains: nothing urgent, and everything that requires judgement. Work that can be deferred indefinitely usually is, so it gets a slot.

Four rules that do the work

Each exists because of a specific way maps go wrong.

These are the parts worth copying into any living review, whatever the subject. Each is a response to a failure that is easy to commit and hard to detect afterwards.

Every figure carries a verification grade: verified, reported, self-published, or claimed.

Without it, a company's own estimate and an independent replication look identical three months later, when nobody remembers which was which.

Where energy is involved, record how the number was obtained as well as what it was.

A modelled figure and a metered one are different kinds of statement. Storing only the value makes them interchangeable, which is how simulations become headlines.

A letter of intent is never promoted to an award.

Announced money and committed money diverge routinely. Collapsing them inflates a field's apparent momentum, and the error is almost never corrected later.

Failures are entries, not deletions. They move to a graveyard with the lesson attached.

A map showing only survivors teaches survivorship. The record of what did not work is usually the more transferable part.

Falsifiable markers, written in advance

Decide beforehand what would change your mind.

The design carries twelve predictions, each paired with the evidence that would flip it and what that would imply. Writing them before the evidence arrives is the entire point: it converts a judgement call made under pressure into a decision made calmly in advance. The four below are shown as they stood when the review was compiled.

MarkerStatusWhat would move it
A thermodynamic sampling chip lands a public benchmark with wall-plug numbers by mid-2027PendingA measured hundredfold on a real workload validates the lane; slipping the date marks it at risk.
Processing-in-memory ships broad commercial energy savings before any paradigm chipOn trackStandards ratification plus a first device teardown would log the paradigm's first mass savings as arriving from the incumbent side.
A photonic compute startup funded in 2026 pivots to interconnect within three yearsOpenA pivot announcement becomes a graveyard row and strengthens the compute-versus-interconnect reading.
The scaling claims behind energy-descent models are independently replicatedUnreplicatedReplication hardens the model-class argument; refutation requires rewriting the section that depends on it.

four of twelve shown · the full ledger travels with the review

Where each kind of news goes

A routing rule removes a daily judgement call.

Deciding case by case where a finding belongs is how inconsistency enters a long-running review. Fixing the routes in advance means the only remaining judgement is whether the evidence is good, which is the judgement worth spending attention on.

A new organisation→ a database row, and a source row alongside it
A national or regulatory move→ the programmes table
A death, pivot or acquisition→ the graveyard, with the lesson written down
A changed headline figure→ the explainer and the next revision of the review
A marker firing→ the ledger, and the review's changelog
Anything measured with receipts→ candidate for the open registry

The design closes with the line that matters most for reuse: whoever runs it, runs it the same way. A method that depends on the judgement of one person is not a method — which is also why a method nobody is running should be described as a design rather than as a practice.

See the map it maintains →The course: PAI-270