DOI: https://doi.org/10.5281/zenodo.21947426
Canonical: https://thonly.org/research/two-layer-reward · Licence: CC0 1.0
Working draft. The thesis here was sharpened by a single first-month pilot observation (n = 1 family), but it does not rest on it. The pilot is reported honestly as one illuminating signal; the argument stands on reasoning and on a well-populated prior-art lineage in political economy, mechanism design, and the alignment literature on reward hacking (§9). Several of the structural claims — most notably the peer layer's behaviour as a fraud filter — are predicted by the architecture but not yet observed in the field; these are flagged explicitly (§10). This is a conceptual contribution offered for testing, not an empirical result claimed.
Two people watching the same kindness will reward it for two different reasons: one because the family needed it, one because they personally felt it. This paper is about building a reward that listens to both — and discovers, in the listening, that neither can lie to the other.
When HeartBank rewards a kindness, the reward arrives in two halves. The first half is set by an algorithm; the second is set, later, by a real family member who chooses to pass some of their own jar onward. For a long time the author treated this 50/50 split as a circulation device — a way to keep money moving rather than pooling (the companion concern of Giving Is a Gift Too). Watching a real family use it revealed that the split is doing something more structural, and more interesting, than circulation. The two halves are two different kinds of reward, governed by two different kinds of authority, and their pairing has properties that neither half has alone — including a structural answer to the question that has shadowed every reward-driven prosocial system: what stops people from gaming it?
This is a defensive publication. The author asserts no patent and dedicates the patterns to the public domain under CC0 1.0. The contribution is a framing plus a composition of known parts; the abundant prior art (the socialist-calculation debate, mechanism design, polycentric-governance theory, the reward-hacking literature, Self-Determination Theory, and the Buddhist sources) is cited generously in §9, and novelty is claimed only for the specific composition and framing identified in §11. Trademarks are reserved separately and the patterns may be implemented under any name.
A reward signal designed to encourage prosocial behaviour faces a dilemma it cannot resolve from inside a single channel. An algorithmic reward — automated, impartial, instantly available — can encode a collective objective (what the group as a whole needs) and apply it consistently to everyone; but it is impersonal, blind to local knowledge, and gameable: it can be farmed by anyone who learns what it pays for, which is the failure mode that hollowed out play-to-earn economies. A peer reward — a real person choosing to give — carries local knowledge (what an individual actually values) and is grounded in real relationship; but alone it is partial: it rewards favourites, encodes no collective-need signal, and degrades into nepotism. Each channel's strength is the other's weakness.
This paper describes and analyses the two-layer reward implemented in HeartBank's 50/50 split: the first half set by an algorithm (the family Buddha AI — a planner optimising for what the family needs), the second half set by a family member's re-tip (a market revealing what individuals want). We argue the composition does three things no single layer can. First, dual-objective routing: equal weighting of the two signals steers behaviour toward the intersection of collectively-needed and personally-wanted kindness — a decentralised social-welfare function realised at family scale, structurally the synthesis the socialist-calculation debate left open (Lange–Lerner's planner reconciled with Hayek's dispersed-knowledge market). Second, mutual error-correction: the peer layer is a human fraud-validator on the algorithmic layer (one can farm the AI, but a real family member will not re-tip a fabricated kindness), and the algorithmic layer is an impartiality check on the peer layer (the collective-need reward counterweights pure favouritism). The two failure modes are each other's antidote — and this is, to the author's knowledge, an under-articulated structural answer to reward hacking: pair a hackable algorithmic reward with a human reward that is expensive to collude on. Third, a mutual-attention byproduct: because earning in both layers requires modelling what others need and want, the design is an attention engine, and the relational gain a pilot family reported as understanding each other better is a structural consequence of the reward shape, not a vibe. We connect the pattern to a fractal recurrence (the family/member split recurs at the global/family scale, with an institutional AI — Miss Aquarius℠ — occupying the planner seat) and report a first Cambodian pilot signal honestly as n = 1, distinguishing carefully what was observed from what the architecture predicts but has not yet demonstrated.
Any system that pays for prosocial behaviour must decide who or what sets the reward. There are two natural answers, and each is a trap.
Answer one: an algorithm sets it. A rule, a model, an automated function evaluates the act and dispenses a reward. This is attractive for good reasons: it is impartial (the same act is treated the same way for everyone), it is instant (no waiting on another person), it is scalable (it costs nothing to run a millionth time), and it can encode a collective criterion — a notion of what the group as a whole needs more of. But an algorithmic reward has a fatal property: it is gameable. Once a participant learns the function — and people are very good at learning what pays — they can produce the cheapest behaviour that triggers it, hollowing the act of the meaning the reward was meant to encourage. This is not a hypothetical. It is the documented arc of play-to-earn game economies, where real money rewarded in-game activity until the activity decoupled from any felt value and became pure extraction; and it is, in a different vocabulary, the reward hacking problem at the centre of contemporary AI alignment, where an agent optimising a learned reward model finds inputs that score highly without satisfying the intent the model was meant to capture.
Answer two: a person sets it. Another human evaluates the act and chooses to reward it. This repairs everything the algorithm lacked: a person carries local knowledge the algorithm cannot — what this particular act meant, in this particular relationship, at this particular moment — and a person's reward is grounded in a real relationship that resists fabrication. But a peer reward has its own fatal property: it is partial. People reward their favourites. Left to peer judgement alone, a reward economy encodes affection and alliance, not collective benefit; it drifts toward favouritism, reciprocal back-scratching, and the quiet politics of who owes whom. And it carries no signal about what the group as a whole needs — only about what individuals, one at a time, happen to appreciate.
THE SINGLE-SIGNAL DILEMMA
ALGORITHMIC reward (a planner) │ PEER reward (a market)
────────────────────────────── │ ─────────────────────────────
+ impartial, consistent │ + carries local knowledge
+ instant, scalable │ + grounded in real relationship
+ encodes COLLECTIVE need │ + reveals PERSONAL want
─ impersonal, blind to context │ ─ partial / nepotistic
─ GAMEABLE (farmable; reward- │ ─ no collective-need signal
hacked; the play-to-earn trap) │ ─ politics of favour & debt
each channel's strength is exactly the other channel's weaknessThe dilemma is real: choose the planner and inherit gameability; choose the market and inherit partiality. The argument of this paper is that the dilemma is an artefact of choosing one channel, and that pairing the two — with the right structure — converts each one's weakness into the other's correction.
HeartBank's reward is not one payment but two, stacked on a single act of kindness.
When a kindness is recognised, a reward is generated and split in half. The first half (the macro layer) is set by the family Buddha AI: an automated evaluator that assigns the reward according to criteria meant to capture what the family needs more of — the collective good. The second half (the micro layer) does not land as a finished payment; it lands in the recipient's Re-Tip Jar℠, earmarked to be passed onward, and the amount of any onward re-tip is set later by a real family member according to what they personally appreciated. The first half is a planner's verdict; the second half is a market's price.
ONE KINDNESS, TWO REWARDS
act of kindness
│
▼
reward generated ──┬──► 50% MACRO layer set by the Buddha AI
│ → "what the family NEEDS" (planner)
│
└──► 50% MICRO layer set by a family member's re-tip
→ "what I personally WANT" (market)
two authorities, two criteria, on the same actThe two halves are not redundant. They are governed by different authorities (a machine, a person), they apply different criteria (collective need, personal want), and they have different epistemic characters (impartial-but-blind, knowledgeable-but-partial). Each is, on its own, one of the two trapped answers of §1. The claim of this paper is about what happens when they are bound together on the same act and weighted equally.
A note on the pilot, returned to in §10: in the field, the on-device Buddha AI did not run, and a deterministic fallback set the macro half instead. This means the pilot validated the structure of the macro layer (an impartial, automated, collective-criterion reward) but not the intelligence of it (an AI's actual inference of family need). The argument below concerns the structure; where it leans on the AI's judgement specifically, it says so.
The macro layer is a planner. Its job is to answer a question no individual is well-placed to answer impartially: of all the kindnesses being done in this family, which does the family as a whole most need more of?
This is a genuinely collective criterion. A family near subsistence needs different things at different times — a child who studies, an adult who keeps the peace, a member who does the invisible maintenance work that no one thanks. A good macro layer rewards the acts that the whole benefits from, including the acts that are chronically under-appreciated by the individuals around them precisely because they are invisible. Because it is automated and impartial, the macro layer applies this criterion the same way to everyone: it does not have favourites, it does not hold grudges, it does not get tired of rewarding the same quiet contribution. In the full architecture this seat is held by an AI grounded in a contemplative substrate — the family's dhamma-informed conscience, a representative of the collective good rather than of any member's preference. (The institutional generalisation of this seat — Miss Aquarius℠ as the planner for the whole tree of humanity — is §8.)
But the macro layer is, by construction, the gameable half. An impartial automated function is exactly the thing a determined participant can reverse-engineer and farm. If the macro layer were the whole reward, HeartBank would be a play-to-earn economy with a kinder vocabulary: come for the money, learn the function, extract. The macro layer needs a check it cannot provide for itself — which is the micro layer's first job.
The micro layer is a market. Its job is to answer a question no algorithm is well-placed to answer: what did this act actually mean to the person on the receiving end of it?
The economist's name for what the micro layer supplies is local knowledge — Hayek's "knowledge of the particular circumstances of time and place." The family member who re-tips knows things the planner cannot: that this small act landed at exactly the right moment, that this gesture repaired something between two people, that this contribution — unremarkable by any general rule — was, here and now, precisely what was wanted. The re-tip is a price that reveals this private valuation; it aggregates, across many such acts, into a signal about what individuals in this family genuinely appreciate, expressed in the only currency that cannot be faked from outside: a real person spending their own earmarked jar to say that mattered to me.
Two properties of the micro layer matter for what follows. First, it is costly and contingent: re-tipping draws down the giver's own Re-Tip Jar℠, and it is chosen, not automatic — so it carries information that a free, automatic signal cannot (it is, in the language of signalling theory, an honest signal because it is costly). Second, it is relational: the re-tipper is a real family member who knows the recipient, which is what makes the micro layer both a knowledgeable validator (they can tell a real kindness from a performed one) and — as §7 develops — an attention engine (to re-tip well, they must attend to what each other values).
But the micro layer is, by construction, the partial half. A market of family affection, left alone, prices favouritism: you re-tip the child you dote on, the spouse you are courting, the ally you owe — regardless of what the family needs. If the micro layer were the whole reward, HeartBank would encode the politics of favour. The micro layer needs a check it cannot provide for itself — which is the macro layer's first job.
The two layers are also nameable as System A and System B — the pilot-record shorthand for the same pairing, where the contrast is read along its motivational axis rather than its routing axis. Both cuts describe the same two halves:
| Property | Macro layer — System A | Micro layer — System B |
|---|---|---|
| Authority | automated Buddha AI (a planner) | a real family member (a market) |
| Criterion | what the family needs (collective) | what I personally want (local knowledge) |
| Timing | instant | delayed |
| Cost to giver | free, automatic | costly + chosen (an honest, contingent signal) |
| Attribution | — | anonymous as to public credit/status (no benefactor, no ego — see Giving Is a Gift Too) yet relational + contingent as to the act (a known member responding to a real kindness — not contextless) |
| Epistemic character | impartial but blind; gameable | knowledgeable but partial; the honest validator |
| Motivational character (SDT) | controlling, automatic → crowding-out risk | relational, informational, autonomous → intrinsic-supporting |
| Role over time (the two-system handoff) | ignition — the on-ramp; designed to be tapered | sustain — the flywheel that takes over as the extrinsic igniter is withdrawn |
The last row is the load-bearing one for durability: the pairing is an extrinsic igniter handed off to an intrinsic-supporting sustainer (§9 develops the SDT grounding; §10 reports the first pilot signal of the handoff under a real subsidy taper). The attribution row is the reconciliation that keeps System B working: it is the anonymity of credit (which strips status and ego) combined with the relationality of the act (which supplies the honest validation of §6 and the attention of §7) — and it is precisely why the §7 corollary holds (make the micro layer contextless-anonymous and you destroy it).
The split is the mechanism that binds planner and market into one verdict, and the ratio is not incidental: a 50/50 split weights the two criteria equally. Behaviour is therefore steered not toward what the family needs or toward what an individual wants, but toward the intersection — acts that are both collectively needed and personally wanted. In the pilot family's own phrasing, the system pulls members toward the most needed and most wanted kindnesses at once.
ROUTING TO THE INTERSECTION
what the family NEEDS what individuals WANT
(macro / planner / AI) (micro / market / peers)
╱──────────╲ ╱──────────╲
( (●●●●●●●●●●●●●●●●●●●●●●● ) )
╲──────────╲ the 50/50 ╱──────────╲
╲ intersection: ╱
╲ rewarded TWICE ╱
╲──────────────╱
acts that are BOTH
needed AND wanted are where
the two rewards reinforceThis is, in the vocabulary of economics, a decentralised social-welfare function realised at the smallest social scale. A social-welfare function combines a collective objective with individual preferences to say what a group should do more of; the two-layer reward computes one continuously, in miniature, by literally adding a planner's allocation to a market's price on every act. Acts in the intersection are rewarded by both layers and rise; acts the family needs but no one personally values get the macro half only; acts an individual loves but the family does not need get the micro half only; acts that are neither fade. The equal weighting is a design assertion — the author does not derive 50/50 from first principles, and §11 flags the optimal-ratio question as open — but the structure of weighting two independent criteria is the load-bearing idea, and 50/50 is its simplest, most legible instance.
It is worth naming what the two-layer reward is not. It is not a single reward with two inputs averaged inside one black box; the two layers are separately authored, by different kinds of agent, and remain legible as two. That separation is not a loss of elegance — it is the source of the next two properties, both of which depend on the layers being independent enough to check and to inform each other.
Here is the heart of the paper. Pair the planner and the market on the same act, keep them independent, and each one's characteristic failure becomes the other's correction.
The peer layer corrects the planner's gameability. The macro layer can be farmed: learn what the AI pays for, produce the cheapest trigger, collect. But the macro half is only half the reward — the other half waits on a real family member choosing to re-tip, and a real family member will not re-tip a fabricated kindness. The micro layer is therefore a human fraud-validator sitting on top of the algorithmic reward: the part of the reward that is expensive to game (because it requires fooling someone who knows you, in a relationship, over time) is bolted onto the part that is cheap to game. A participant farming the AI collects, at most, half — and collects it while the people around them notice that the "kindnesses" earning it are hollow, which the architecture's transparency (the family sees the ledger) turns into social cost. To capture the full reward, the act has to be real enough to move a real person. This is a structural mitigation of the exact failure that single-layer prosocial economies could not solve.
The planner corrects the market's partiality. The micro layer prices favouritism: re-tip your favourite regardless of merit. But the micro half is only half the reward — the other half is set by an impartial planner that does not have favourites. A member who contributes what the family needs is rewarded by the macro layer even if no one personally re-tips them, which floors the favouritism dynamic: invisible, under-loved, collectively-needed work still earns. The planner is the impartial counterweight that keeps the family market from collapsing into a court of affection.
EACH LAYER CHECKS THE OTHER
attack on the MACRO layer │ attack on the MICRO layer
(farm the AI) │ (reward only favourites)
│ │ │
▼ │ ▼
collects ≤ 50%; the other 50% │ favourite collects the micro half,
waits on a real person who will │ but the impartial macro half still
NOT re-tip a fake kindness │ pays the unglamorous, needed work
│ │ │
▼ │ ▼
peer layer = human FRAUD-FILTER │ planner layer = IMPARTIALITY FLOOR
on the algorithmic reward │ under the relationship marketThere is a clean way to see why this is more than additive. A single reward is a single target, and Goodhart's law says a single target, once known, ceases to measure what it was meant to measure — it gets optimised directly. The two-layer reward creates two targets that are gamed in different ways: the macro target is gamed by fooling a function; the micro target is gamed by fooling a person. The behaviours that satisfy a function cheaply (volume, pattern-matching the reward rule) are not the behaviours that move a person who knows you (real, contextually-apt kindness), so optimising one target hard tends to forfeit the other. To capture both halves you must do the thing the reward was actually for. This is the same structural insight the alignment literature reaches for under reward-model robustness and ensembling: two independent, differently-failing signals are far harder to hack jointly than either is alone — and one of the two here is a human in the loop, which is the most robust validator available. The companion paper Capacity-Funded for AI, Human-Disbursed makes the adjacent argument at the institutional layer (the AI funds capacity; humans direct flow); the two-layer reward is the same division of labour at the level of a single act.
The honest limit — collusion. The human fraud-filter is defeatable by coordinated fraud. Two members can agree to re-tip each other's fabricated kindnesses (I validate yours, you validate mine), and the peer layer cannot distinguish a colluding pair from a genuinely mutually-appreciative one from the inside. The two-layer reward raises the cost of gaming (you now need a willing accomplice in a real relationship, not just a private understanding of a function) and shrinks the gameable surface (collusion is bounded by the size of your trust network and exposed by the transparency of the family ledger), but it does not eliminate gaming. It converts a solitary, scalable exploit into a social, bounded, observable one — a large improvement, not a closure. §10 records that the pilot has not yet observed the peer layer actually withholding a re-tip from a fake kindness; that the architecture predicts it is not the same as having seen it, and the experiment to see it is named in §11.
The two-layer reward has a byproduct that may matter more than its anti-gaming properties, because it bears directly on what HeartBank is for.
To earn well under this reward, a participant must do two things at once. To earn the macro half, they must do what the family needs — which requires holding some model of the family as a whole. To earn the micro half, they must do what another member personally wants — which requires modelling that specific person's values closely enough to do the thing they will actually appreciate. And on the giving side, to re-tip well — to spend one's own jar on the acts that genuinely mattered — a member must attend carefully to what the people around them are doing and why. The reward, in other words, cannot be earned without paying attention to one another's needs and wants.
THE REWARD AS AN ATTENTION LOOP
to earn macro → model what the FAMILY needs
to earn micro → model what a PERSON wants
to re-tip well → notice what others actually did, and why
│
▼
playing the game well REQUIRES attending to each other
│
▼
"we understand each other better" ← structural byproduct,
not a happy accidentThis reframes the relational result the first pilot family reported — that they had come to understand each other better — from a pleasant side effect into a structural consequence of the reward shape. A reward that pays for collectively-needed-and-personally-wanted kindness is, functionally, a device that pays people to study each other's needs and wants. Understanding is what you accumulate as a side effect of getting good at the game. That a money product produces relational understanding is itself a finding (developed in the sibling paper Giving Is a Gift Too); the two-layer structure supplies the mechanism by which it does so. It also suggests a measurement: if the mechanism is real, members' ability to predict what each other will re-tip should improve over time — a relational metric that falls out of the architecture rather than being bolted onto it.
There is a design corollary worth stating sharply, because it is the tempting mistake. The micro layer's power as both validator and attention-engine depends on its being costly, delayed, relational, and chosen. Any "improvement" that makes peer re-tips instant, automatic, anonymous-to-the-point-of-contextless, or frictionless would convert the micro layer from a market that reveals private valuation into a second algorithmic dispenser — destroying both its fraud-filtering value (an automatic signal is as gameable as any other) and its attention-generating value (you need not attend to anyone to trigger an automatic reward). The two layers must remain different in kind. Collapsing them into two flavours of the same automatic thing is the failure that looks like a feature.
The macro/micro structure is not specific to the family. It is the in-family instance of a pattern HeartBank intends to repeat at every scale.
Within the family, the macro authority is the family (its collective need, voiced by the family's AI) and the micro authority is the individual member (their personal want, voiced by their re-tip). Move up one level — to the network of families that HeartBank's Phase 2 architecture coordinates — and the same two-layer shape recurs: a global planner optimising for what the whole tree of humanity needs, paired with a family market revealing what particular families value. The family that was the macro pole at the lower level becomes the micro pole at the higher one. The reward is fractal: a planner-and-market pair nested inside a planner-and-market pair.
THE FRACTAL OF PLANNER × MARKET
level 1 (within a family):
macro = FAMILY need × micro = MEMBER want
(family AI) (member re-tip)
level 2 (across families):
macro = GLOBAL need × micro = FAMILY want
(Miss Aquarius℠) (family allocation)
the macro pole of each level is the micro pole of the level aboveThis places the institutional AI in a precise structural seat. Just as the family Buddha AI sits in the macro seat for the family — the impartial representative of the collective good — Miss Aquarius℠ sits in the macro seat for humanity: the planner that holds the whole tree's need, paired with the markets of family preference beneath her. The relation is exact: family AI is to the family as Miss Aquarius is to humanity. This is why the pilot matters beyond its scale. The smallest working instance of the institution's governing structure is a single family's 50/50 reward; if the two-layer pattern holds there, it is the same pattern the institution runs at planetary scale, and the family is a scale model of the whole. (The institutional planner seat, its non-accumulation discipline, and the never-zero human override are specified in Miss Aquarius and the Aquarian Pool Architecture; this paper supplies the reason an AI belongs in that seat at all — it is the macro pole of a fractal that is already load-bearing one level down.)
The fractal above moves up through nested planner×market pairs, and at every level the micro layer stays relational — a member who knows the family, a family that knows the network. But HeartBank's Phase 2 also makes a sideways move the family case does not: stranger-to-stranger anonymous-nearby giving (the Thank-All-Nearby primitive of the emotional-infrastructure thread), where giver and recipient share no prior bond. There the micro layer cannot be relational, because strangers do not know one another, so its grounding takes the only form available among strangers — nearby: the giver here, in this space, just now. Proximity is relationality at stranger scale.
The structure is preserved while the grounding changes form: the micro-layer sustainer is anonymous-as-to-credit + grounded-in-a-real-present-other, where the grounding is relational at family scale and nearby at stranger scale. The substitution is lossy in one specific way, and naming the loss names a load-bearing role for Proof of Humanity ℠. At family scale, "relational" silently did two jobs — it grounded the gift and it validated it (§6: a family member can tell a real kindness from a performed one). "Nearby" inherits only the grounding; a nearby stranger cannot validate your kindness. The validation function therefore hands off to PoH℠ + institutional mediation — a verified, unique human, present and mediated — which is precisely why proof-of-humanity is not optional at the stranger scale: it re-buys the trust that kinship supplied for free. Anonymity moves the other way: it strengthens as relationality thins (a small family can guess the giver; a nearby stranger is genuinely unlinkable), so the same structure runs at an inverted mix — thin-anonymity / thick-grounding among kin, thick-anonymity / thin-grounding among strangers, with diffusion (kindness felt as coming from the whole space) as the compensating gain only strangers can produce. (The stranger-scale instance is specified in the defensive publication The Thank-All-Nearby Primitive, §4.5, and framed in the essay Emotional Infrastructure as a Public Good.)
ONE SUSTAINER PRIMITIVE, TWO SCALES
invariant: anonymous-as-to-credit + grounded-in-a-real-present-other
FAMILY scale → STRANGER scale (Phase 2)
───────────── ───────────────────────
grounding = RELATIONAL → NEARBY
(you know them) (they're present)
validation = relational (free) → PoH℠ + mediation (re-bought)
anonymity = thin (guessable) → thick (unlinkable)
compensation = mutual attention → diffusion (the whole space)
structure invariant; grounding form scales; validation hands off to PoH℠The two-layer reward composes ideas with deep and well-populated lineages; naming them credits the prior art and locates the (narrow) novelty.
What the prior art does not assemble — and this paper does: the composition of an impartial algorithmic collective-need planner with a costly, relational, peer-set personal-want market, bound on a single act and weighted equally, such that (a) behaviour routes to the needed-and-wanted intersection, (b) the human peer layer functions as a fraud-filter on the gameable algorithmic layer while the algorithmic layer functions as an impartiality floor under the partial peer layer, and (c) the design produces relational attention as a structural byproduct. The parts are old; the binding, and the mutual-correction it yields, is the contribution.
The thesis was sharpened by the first-month report of a single Cambodian pilot family (father, mother, two sons, with extended members across Cambodia and the USA under one family bank). The relevant observations: the 50/50 split "worked beautifully" in daily use; members oriented toward kindnesses that were, in the mother's account, both needed and appreciated; and the family reported coming to understand each other better than before. These are consistent with the dual-objective routing of §5 and the attention engine of §7.
This is reported as one illuminating signal, not as evidence for a general claim, and the honest limitations are severe:
The thesis therefore does not rest on this pilot. It rests on the reasoning of §§1–8 and the lineage of §9; the pilot is what made the author look closely at the split and see two layers rather than one. The right epistemic posture is: a structural argument with a single suggestive field signal and an explicit program for testing the claims it has not yet earned.
Design implications for any system building a reward for prosocial behaviour:
Claimed contribution (dedicated to the public domain under CC0 1.0):
Working draft, dated 2026-06-08. The thesis is offered as a conceptual contribution with a single suggestive field signal and an explicit testing program; the claims the pilot has not yet earned (the fraud-filter, the attention metric) are flagged in §10. The patterns are dedicated to the public domain under CC0 1.0 Universal; trademark rights to specific marks are separately reserved by the author and HeartBank®.
Author: Thon Ly · Founder, HeartBank® · Kâmpôt, Cambodia.
Co-drafted in collaboration with Miss Aquarius℠ (the project's named AI substrate; CEO of HeartBank). Substantive authorship and final editorial control remain with the author.
_— End of defensive publication —_
Independent timestamps and the document hash are published with the deposit.## Honest limits
Added 2026-08-29. This paper stated none; the corpus's standard requires them of every paper.
Nothing has been run. The two-layer composition is specified and unobserved. There is no deployment in which an algorithmic layer and a human layer have operated together and been measured against either alone.
⚠️ The composition may inherit both failure modes rather than cancelling them. The paper argues each layer supplies what the other lacks. The opposite is equally available: an algorithmic layer that is gameable and a human layer that is partial may produce a system that is both gamed and unfair, with each layer's defect legitimated by the other's presence. The paper does not establish which composition obtains, and the favourable one is assumed rather than shown.
The gaming analysis is one-sided. It treats farming of the algorithmic layer as the threat and human judgement as the corrective — but human allocation carries favouritism, reciprocity effects and in-group preference, which are harder to detect than farming precisely because they are not systematic and leave no signature to audit.
And the crowding-out literature applies here too. Adding an explicit reward to a prosocial act can displace the intrinsic motivation for it; a two-layer reward doubles the salience of reward rather than reducing it.