Constraint-first reconstruction

This is a working draft. The public order is on the Gap Last method. This page is the longer argument, still in motion.

Subtitle: How to shrink the unknown before you invent it

Working paper, 2026-08-31. Public instrument: Gap Last (docs/gap-last-tool-spec.md).

Abstract

Most of what we try to understand cannot be rerun. School science still teaches a laboratory method as the only respectable path: form a hypothesis, test it, try to falsify it. That method is powerful where intervention is possible. It is the wrong default.

This paper specifies constraint-first reconstruction: thicken the description of a thing until the chain is constrained; use what is already known to eliminate mechanisms; hypothesize only about residual gaps; keep more than one working hypothesis in those gaps; leave remainder unknown when the traces do not decide. When a better trace arrives, reopen the bound. The gap moves when the object moves.

You cannot answer how something happened until you have a solid what. How is a mechanism. What is the object. A how aimed at a thin what is a story with better vocabulary.

Pursue the "how" at your folly if you fail to frame the correct "what".

The public instrument is Gap Last. This paper is the method. The tool is what keeps you in the order.

Reconstruction is old. What is specified here is an order, a prohibition, a stop rule, and a loop that is allowed to retire its own questions. The scope is broader than disasters. This is the default for a learner until some other instrument earns the job.

1. The default, not the special case

I named this on a Himalayan flood because the flood made the error expensive and visible. That does not make the flood the only subject.

A learner meets a claim, a rumor, a symptom, a product demo, a family story, a market move, a line in a book. Almost none of those can be rerun on purpose. The scarce resource is not a clever hypothesis. It is a description tight enough to rule out most stories, and a leftover gap that can hold what you still do not know.

Gap Last is the default until proven or shown otherwise. If you can run the experiment, run it. If you can prove the theorem, prove it. If the bottleneck is sampling, do the statistics. Those are graduations from the default, not the starting posture. Starting with a theory is comfortable. It also scales whatever you got wrong about the object.

2. The wrong question

"Who caused it?" and "what's my theory?" feel like work. They skip the what. The what of the event is the bound: what happened, where, in what order, at what scale. Skip that and every later mechanism scales the wrong peak.

The first answer to "what happened" is almost always a noun. Glacier. Construction. Climate. The noun is the elephant. The real what sits a few questions further in, the same way the real destination sits past the word "summit."

The common failure is not a missing how. It is cementing a flawed or incomplete what too fast. Limited perspective, a favorite mental model, a preconceived villain, an inclination you already had: any of those will hand you a noun and call it settled. Wednesday's "glacier collapse" was that move. Once the what is cemented, every later how looks like work.

A collapsed chain is a wrong question. "China or climate?" can be answered with great energy and still leave you lost, because initiation, amplification, and exposure were never separated. Right answers to the wrong pair are how discourse stays busy. Define first. Invent later, and only in what is still open. Remainder is a result, not a failure of nerve.

3. The problem the flood made visible

A glacier fails. A plant explodes. A rumor hardens in an afternoon. People reach for a cause while the object is still unnamed.

Three common errors:

Cemented what. A first noun is treated as settled. The rest of the work is how. Limited perspective, a favorite model, or a preconceived inclination did the cementing.
Ruling theory. One story is chosen early (climate, construction, incompetence, malice) and evidence is gathered in its favor. Often the how that grew on a cemented what.
Collapsed chain. Initiation, amplification, exposure, and response are treated as a single question: who caused it?

The Himalayan flood showed all three. A social claim that Chinese construction detached a glacier sat beside a public claim that the event was climate. Available traces already separated the problems: a high scar in Nepal, a debris flood down a river, hydropower sites and people in the path. Distant construction-as-resonator failed basic constraints of energy, attenuation, and location. That deletion was available before any new theory of mountains.

A day later a clearer satellite pass moved the object again. That is not a third error. It is the method working. The error would have been to defend "glacier collapse" once the rock under the ice was visible.

4. Why the textbook method is the wrong default

The experimental method assumes you can perturb a system and compare runs. Most of what a person or an agent meets leaves traces, not reruns. The scarce resource is not ingenuity. It is a description tight enough to rule out most stories.

This is not an argument against experiment. Where a test can be run, run it. Constraint-first reconstruction is the starting posture for the large class of happenings, claims, and observations that can only be approximated after the fact. School science hammered "the scientific method" as the only valid approach to understanding. It has limits. Most of what we meet cannot be rerun. At best we approximate its shape.

5. What is not new

I did not invent reconstruction. I specified an order and a prohibition, and I gave the stop rule a public name.

Historical science (Carol Cleland): present traces of past events; rival hypotheses; search for a smoking gun.
Process tracing (Van Evera, Bennett, Collier): hoop tests, smoking-gun tests, straw-in-the-wind, doubly decisive tests.
Multiple working hypotheses (T. C. Chamberlin, 1890): do not marry a ruling theory.
Accident investigation, forensic reconstruction, diagnostic reasoning, Analysis of Competing Hypotheses.
Boyd's OODA: kinship in the loop, especially in orient. I posed that kinship. See §8.

These traditions already know that one-off events are reconstructed from traces. Constraint-first reconstruction borrows from them. Its emphasis is stricter on when invention is allowed, explicit that the leftover gap may move, and explicit that the same order is the default for ordinary understanding, not only for a board of inquiry.

Cleland's pattern often proliferates rival hypotheses once puzzling traces appear. The specification here delays that proliferation until the observation has been thickened and the known has been spent on elimination. That is a procedural difference, not a new epistemology.

Methods are many. Principles are few. This is a method. The principle underneath it is older: name the what before you spend the how.

6. The specification

6.1 Order

Keep this list aligned with the tool spec.

Bound the event: time, place, scale, sequence.
Split facts into settled, provisional, and open.
Write the chain in parts: initiation, amplification, exposure, response.
Spend the known on elimination. Apply hoop tests to geometry, timing, energy, materials, records, incentives.
Name only the residual gaps.
Place multiple working hypotheses in those gaps only. Prefer mechanisms that need the least new machinery.
Search for discriminating traces, not confirming anecdotes.
Re-open the bound event if a new trace changes the object. Report remainder unknown as a result. Record how the leftover question moved.

The tool adds a ninth output, the reconstruction log, so a later reader can see Wednesday's bound and Thursday's bound. That is an implementation of step 8, not a new principle.

6.2 Prohibition

A hypothesis that does not point at a named residual gap is out of order.

Correlation in a region is not location on a scar. Construction in a river basin is not blasting on a 5,200 m face. Warming in a range is not, by itself, the last increment that dropped a particular slope. A noun in a headline is not a what.

6.3 Stop questions

What part of the chain is this claim about?
Has it already failed a hoop test?
Is this a residual gap, or leftover storytelling?
Did a new trace move the object, and if so, what is the gap now?

A shorter form: Is that the what, or are you already on a how?

And when the what looks done: Did you cement a first noun, or can you still reopen it?

6.4 How a better trace can change the work

Settled, provisional, and open are labels for the current description, not a final map. A better trace can:

eliminate a mechanism (distant resonance, ice-only burst)
shrink a gap (we now know rock led)
shift a gap (the last increment now has to act on bedrock, not hanging ice alone)
split a gap (preconditioning of the face versus the flood's water budget versus who was in the path)

The first gap often is not answered. It is retired. A tighter what creates a different leftover.

Let the bound event change without treating that as humiliation or as a license to invent. New stories still have to point at the new gap.

7. Guardrails

Observation is theory-laden. Early labels mislead. The same flood was called earthquake, GLOF, glacier collapse, then bedrock failure carrying ice. Freezing step 1 too soon shrinks the search space around the wrong object. That is the cemented what: a limited perspective, a favorite mental model, or a preconceived inclination handed you a noun, and you treated it as settled. Keep the description layered. You cannot spend how on that noun until the what can bear it.

"Readily knowable" is not "only what matters." Spending the known first is a search-space move. It does not make the leftover gap trivial. Climate preconditioning and a last-increment trigger can both be true. One may remain hard to pin down.

First principles are a knife. They are for asking whether a proposed mechanism can exist in these materials at this distance. They are not a license to decorate a clever machine that the traces do not require. I like to take first principles to ridiculous levels. That habit is useful when it eliminates a story. It is vanity when it builds one the traces do not ask for.

Remainder is allowed. "We may never know the last increment" is a legal result.

The gap is not a trophy you find once. Naming a leftover question is not the same as knowing you have the last one. Gap Last is the order. Last Gap is a claim you usually cannot yet make. A later trace can move the question. Treating the gap you named as closed, when you do not yet know enough to say so, is the same thin move this method exists to stop.

Do not create a ruling method. Chamberlin's warning applies to this paper. If the bottleneck is an experiment, this framework is the wrong instrument. If the bottleneck is time-limited action against an opponent, OODA is the closer instrument.

8. Learners, OODA, agents

I posed this to Grok after the Thursday images: effective learners already run a feedback loop, a cousin of OODA. I named the kinship. Grok named how decide differs.

Effective learners do not defend the first frame. They take in a trace, update the object, throw out mechanisms that no longer fit, and only then spend attention on what is still open. When Thursday's images arrived, the move was to re-orient: rock led, ice followed, the leftover question moved.

I have used OODA for years as the correction term in a life that does not get a smooth path. Walking is a controlled fall. Stumbling with grace is proportional response. The smoothness depends on the speed of your OODA loops. Boyd put the danger in orient, not in observe. Bad orientation freezes a story. Good orientation destroys the old picture fast enough to act.

Constraint-first reconstruction is a slower, reconstructive version of that loop: observe and bound; orient by constraint and elimination; decide only inside a named gap; act by looking for a discriminating trace; loop when the bound event shifts.

I called it a cousin, not the same thing. OODA is for time-limited action against an opponent. This method is for understanding, including the cases where acting on a false cause is worse than leaving a gap. You still loop. You refuse to treat "decide" as "pick a villain."

Agents and comment threads fail the same way: first plausible story, then commitment. Models skip the bound and write a cause. A skill that will not output a villain until the chain and gaps are named beats another "be more rigorous" prompt. A learner is allowed to retire a question.

9. Worked case: Langtang Lirung

The method was named on this flood. Same order applies to quieter things.

9.1 Three claims, one sitting

Claim A: Construction vibration traveled through rock and ice, resonated at an unsupported extent, and caused the break.

Hoop tests: blasting energy and duration versus earthquakes; rapid attenuation of particle velocity; ice–rock impedance mismatch; hydropower sites downstream of the scar; failure described as bedrock under ice at high elevation.
Result: mechanism fails as a far-field trigger. Local work on the face would be a different claim and would need different traces.

Claim B: The event is "because of climate change."

Result: as a background probability statement about warming, permafrost, and thinning, this can be a live preconditioning hypothesis. As a sole, specific trigger of this token collapse, it occupies a residual gap and should be labeled as such.

Claim C: Hydropower "caused the disaster."

Split the chain. If the claim is exposure (people and tunnels in the path) it can be true without touching initiation. If the claim is initiation, it has not yet produced location on the scar.

The method's value here is not that it settles climate. It is that it stops three different questions from wearing one noun.

9.2 The bound event moved (2026-08-31)

A later trace arrived: Onmanorama, 2026-08-28, drawing on Cook and Shugar.

Wednesday's first images were cloudy and dusty. They were enough to say ice had come off Langtang Lirung. Thursday's clearer pass showed something larger: the rock the glacier was sitting on failed, and the ice went with it. Shugar: a big bedrock failure that took part of the glacier with it. Cook: the rock collapsed; the glacier was cargo.

That is initiation, restated with more detail. A tighter what, not a new villain.

The rest of the chain still separates:

Amplification. The mass dropped into Lendi/Lhende Khola, stayed mobile in a steep narrow valley, and ice melt plus sediment turned it into a debris flood. River color going green to brown is the sediment load, not a separate mystery.
Exposure. The same satellite work that found the scar also shows villages and the border works buried downstream. That is why the death count is a flood story even if the start is a slope story.

Onmanorama, and the Cook/Shugar interviews it draws from, do not mention construction at the scar. Climate enters only as a possible precondition: permafrost that used to hold high rock and ice together is weakening. Cook and Shugar still will not call that the precise trigger of this face. That remains a residual gap.

In method terms, the Wednesday label "glacier collapse" was a provisional bound. The Thursday images were a hoop test on that label. It did not survive as the whole initiation. The new bound is closer to: north-face rock avalanche on Langtang Lirung, ice included, then a long runout.

What is still open is why that particular slab let go that morning: thaw, water in joints, a last increment nobody has isolated yet. That is the gap. The satellite story did not fill it. It did eliminate a simpler ice-only picture.

Wednesday's leftover was closer to "why did that ice detach?" Thursday's leftover is "why did that rock slab fail, taking ice with it?" The first gap was retired, not answered. The first question was aimed at the wrong object.

This is the reopen rule with a receipt. Spending the known deleted ice-only as the whole start. It did not license a new story about the last increment.

10. The tool split

Constraint-first reconstruction is the name for the method.
Gap Last is the name for the instrument: site, checklist, agent skill.

Non-specialists can use the tool without reading process tracing. Specialists can inspect the paper and see the debts. The homepage should keep them coupled: Gap Last implements constraint-first reconstruction.

Language models invent early. The skill refuses a cause until a gap is named. After Langtang, it also has to reopen: no frozen bound when a better trace arrives.

11. Limits and objections

"This is just critical thinking." Yes, specified. Unspecified critical thinking does not stop a thread.
"Hypothesis-first is sometimes better." Agreed, when the instrument is actually available. Measurement design, experimental programs, and mathematical conjecture often need an early hypothesis so you know what to look at. That is a graduation, not the default.
"Description cannot come before theory." Partly true. The bound event is not theory-free. The mitigation is layered facts and a duty to reopen the bound when traces change, not a claim of innocent seeing.
"Self-evident methods do not need a name." They need a stop rule. Without a name, the order dissolves in use.
"It will go viral and get stupid." Likely, if it spreads. The paper should treat slogan drift as the main expected failure, not as a reason not to specify the tool.
"This is just OODA." I called it a cousin. Kinship in the loop. Different job, different "decide." If you need to act against an opponent before the traces are thick, do not run this protocol as a delay.
"This is only for disasters." That was the seed, not the scope. A seeker on any topic is better served by this order than by inventing first.

12. Open questions

Is "constraint-first reconstruction" the right formal name, or is "subtractive reconstruction" clearer?
Can the eight steps be reduced for public use without losing the prohibition?
What counts as a hoop test in non-physical domains (intent, institutions, markets, a conversation you just had) without pretending those traces are like seismograms?
How should the method treat conjunctive causes: several true parts that are not rivals?
What would a serious validation look like: blinded reconstructions against later official findings?
Where is the line between "spend the known" and "ignore an uncomfortable unknown"?
How much reconstruction log does a public tool show before it becomes a history of the news cycle rather than a bound event?

13. Do this on the next thing you hear

Pick one claim you met today. A headline, a Slack message, a family story, a diagnosis, a product promise. Write four lines:

The what, as tightly as you can say it. Mark it settled, provisional, or a first noun you might have cemented.
One mechanism the geometry, timing, or record already eliminates.
The leftover gap, named as a question.
Whether "how" or "who caused it?" is even the right question yet.

If you cannot name the gap, you are not ready to invent. If you already have a how, check whether the what can bear it.

Changelog

2026-08-30: Method named.
2026-08-31: Split to this file. Moving gaps, Thursday bound, learners / OODA / agents. Default broadened. Hinge and maxim hung. Deed added. Order and steps preserved.
2026-09-04: Vault paths and sister-project lineage removed from this tree. Method kept.
2026-09-04: Public draft on the site at /paper. Eliminate, not kill. Last Gap is a claim, not the name.