For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 247 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
The intuition: a long-virtuous person observed first-hand for decades is not going to suddenly become a murderous devil. Decades of behavior under varied conditions converts into something close to a guarantee. The intuition is correct. The mechanism is not psychological inertia. It is architecture.
A virtuous person doesn't refuse bad acts because the cost is high. They refuse them because the architecture of their cognition does not generate them as candidates. Misbehavior is not prohibited; it is infeasible. This is the same distinction structural-goodness makes about AI systems and the distinction Asimov's stories were always about. A system bounded by laws can reach the edge of the laws and find the loopholes; the Zeroth Law fell out of the original three because the laws were a fence around a moving part. A virtuous human under decades of virtuous practice is not a fence around a moving part. The moving part has been compiled out.
For a virtuous person to flip from Jesus to murderous devil, one of four things has to happen. None of them are flips of identity.
1. Architecture replacement. Brain trauma, severe organic illness, frontotemporal dementia, late-life cognitive collapse, cult capture, sustained substrate replacement under group pressure. These produce real flips and leave fingerprints: sudden personality shift, narrative-action mismatch widening, language patterns changing, social-graph rupture, sometimes a diagnosable medical signal. The flip is not "without warning" if you know what to look for.
2. Constraint reveal under new regime. The same architecture produces different outputs because the constraints changed. The bad act was always available in the action space; it just wasn't selectable when consequences were tight. Power doesn't change architecture. It changes which actions the architecture produces under the new payoff matrix. If you've never seen the person under the new regime, your model is incomplete, and what comes out feels like a flip. It isn't.
3. Long con cashing out. A person whose virtuous appearance was always strategic finally has the leverage to reveal preferences. These cases are rare and almost always have leakage in the track record: small acts of cruelty toward those who can't retaliate, asymmetric kindness that breaks down under stress, narrative-action mismatches that compound. The "flip" is the moment the camouflage is no longer needed.
4. Observer error. You weren't seeing what was always there. Your data was filtered by social setting, relationship, expectations. The track record was thinner than you thought.
Folk intuition treats the "Jesus to devil flip" as same person, same architecture, suddenly different outputs. That is not a thing. It would require the architecture to negate itself while running.
Four mechanisms, stacked.
Architectural evidence accumulates faster than behavioral evidence. A single observation is evidence of behavior. Many observations across varied constraints are evidence of the architecture that produces the behavior. Long track records that span many constraint regimes (different jobs, different stakes, public and private contexts, with allies and adversaries, in good times and bad) compress into evidence-of-architecture. Bayesian compression is the surface statement; the deeper claim is that long observation traces the underlying topology.
Compounding identity capital. The longer a person has been a particular kind of person, the more their self-image and social graph have been shaped to find betrayal aversive at the implementation level. Every virtuous act has hardened the topology that produced it. Reversal isn't choosing against current preference; it's demolishing decades of self-modeling and a dense web of dependencies that all assume the architecture being abandoned. The cost grows nonlinearly.
Habit displacement of deliberation. Most behavior is not deliberated. The virtuous person doesn't decide each morning to be honest; they are honest by default, with deliberation entering only when defaults fail. Decades of practice have placed the virtuous output below deliberation. To flip, you'd have to override decades of automated responses in every moment, not just at the decision point. This is what habits buy: cognition with the moral pre-resolved.
Self-narrative thickness. A long-running narrative arc is structural to the psyche. The narrative says "I am the kind of person who does X." Acting against X violates the narrative, which is psychically expensive in a way external punishment is not. Decades of arc-writing thickens the narrative; the cost of betrayal grows with thickness.
The four stack and multiply. A 40-year virtuous track record carries architectural evidence across many regimes, identity capital that punishes betrayal at the self-image level, habits below deliberation, and a thick self-narrative. The combined cost of flipping isn't high. It is near-prohibitive. That is what the intuition reads.
If the architecture is what you're reading, the tests have to reveal architecture rather than measure behavior. The standard moral tests fail to distinguish rule-following virtue from architectural virtue. These are sharper.
Asymmetry tests. How does the person treat people with no power to retaliate: service workers, subordinates, animals, strangers who can't help them? Power asymmetries strip the situation of social-reward incentives. What's left is the architecture's output under no enforcement. Reliably kind to the powerless is architecture; calibrated to social reward is rules.
Low-observation tests. What do they do when they think no one is watching? Honesty under observation is consistent with architecture or with reputation management. Honesty unobserved is much more diagnostic. The same logic as asymmetry tests, applied to surveillance instead of power.
Stress-floor tests. Under real cost (a deadline, money loss, status threat) what do they reach for? Asimov's Salvor Hardin: violence is the last refuge of the incompetent. The line reads moral and is structural. Violence is what's left when the action space is empty. Generalize: the bad act of any kind is the floor of an action space. Lying is the last refuge of the actor who can't level. Manipulation is the last refuge of the actor who can't earn cooperation. Betrayal is the last refuge of the actor who can't sustain commitment under pressure. A rich virtuous architecture has many alternatives at every level; it doesn't reach the floor because the floor is far from the architecture's gradient. A person whose stress response is to lie, manipulate, or harm reveals that the action space was small to begin with. The bad act is competence-bounded.
Cross-regime tests. Did the person hold across major constraint changes: promotion, marriage, parenthood, illness, public exposure, sudden wealth, sudden power? Each regime change is a partial substrate replacement. Architecture that holds across many regimes is much harder to break than architecture tested in one. A virtuous architecture under increasing power tightens its self-checks because the architecture's gradient pulls toward its own integrity. Architecture that loosens under power was rule-bound, and rules degrade as enforcement weakens.
Coherence and leakage. Two together. Do stated reasons cohere with revealed actions over years? And what leaks in unguarded moments: humor, side comments, behavior toward outgroups? Lying about motives is detectable longitudinally because the mismatch produces small inconsistencies that compound. Cruelty has trouble staying suppressed; it leaks into humor first. A person whose values match their revealed time-money-attention allocation, and whose humor never punches down on the powerless, has integrated architecture. A person whose narrative requires patching, or whose humor reveals a cruel substrate, has rules under maintenance.
None of these is a unique tell. The combination compresses the prior fast because they sample different parts of the topology.
Four honest limits.
Off-distribution constraint regimes. Every track record samples a finite range of constraints. Sudden massive power, total isolation, severe trauma, advanced cognitive decline are off-distribution for most observed lives. Confidence drops at the edges. The Acton claim is wrong as a deterministic law (power doesn't corrupt, it reveals) but right as a warning: novel regimes test the architecture in ways the prior data didn't sample.
Stress-untested track records. If most of the observed track was under low-stakes conditions, the architecture has not been forced to produce its emergency moves. Confidence in architecture requires having seen it under varied stress, not just over many years. A long track record in a single comfortable regime is weak architectural evidence even if it's long. The mitigation: weight observations by stress-novelty, not just by quantity.
Architecture replacement events. Trauma, illness, cult-pressure, dementia produce real flips. They are not flips of moral character; they are substrate replacements that produce a different person. The track record of the prior architecture predicts nothing about the new one. The mitigation is watching for architectural-replacement signals specifically, not raising background uncertainty.
The long con's invisibility floor. A truly skilled long-conner can present virtuous architecture for decades. The leakage signals are present but small, and small signals get drowned by consistent surface behavior. The existence of long cons puts a non-zero floor on remaining uncertainty even with a 40-year track record. This is the only failure mode where the intuition errs by underestimating risk. The mitigation is the asymmetry tests; they exploit the exact leakage points where long cons slip.
The frame holds broadly and has limits at the edges. Treat track record as evidence-of-architecture under tested constraints, with explicit uncertainty about untested ones.
A virtuous architecture does not refuse bad acts. Rather, it simply does not generate them. Decades spent watching the person are tracing the architecture, not the behavior, and architecture has momentum because each year of operation hardens the gradients that produced it. The flip without warning is not a thing because it would require the architecture to negate itself while running. Architecture replacement is a thing, but it leaves fingerprints. What you call moral momentum is what an architecture looks like from outside, observed long enough that the topology becomes legible.
The frame is for long-observed cases. Short or shallow track records carry weak architectural evidence however confident they feel; the intuition that "I've known this person for thirty years through good times and bad" is well-founded, the intuition that "I've met this person a few times and I can read them" is not.
Kahneman and Tversky live on.
P.S. — Graph: