Open Problems
These are the places where this framework might break. Publishing them is deliberate. A system that hides its failure conditions is a religion. One that exposes them is engineering.
The engineering backlog (65 items) lives in extropy-engine/docs/GAPS.md. Codex v2.1 stays frozen for now — this page is the public reading of the gaps, not a new edition.
LocalFlow is the errand face
LocalFlow replaces the pile: Uber, Lyft, DoorDash, Grubhub, plus the run you don’t have a car for. Post it. Someone nearby does it. You confirm. No platform fee, no surge. Users never have to say XP.
That confirmation closes a loop. It is not how the score is invented. You do not type in how much a lawn is worth.
SignalFlow is the protocol
You talk to SignalFlow. It talks to the assistant you already trust — ChatGPT, Claude, Gemini, or a model on your own hardware — plus your PSLL. It looks at the task, the duration, before/after evidence on the DAG, and proposes a provisional ΔS. If-then. The other side has to agree.
Company login means company tether. Own hardware is how you stay unknown. A network-hosted model is a later idea, not a product today. MICRO overselling is real. SignalFlow plus evidence plus late burn is how we live with it, not a claim that people will not try.
Measurement
Cross-Domain Entropy Measurement
How do you compare entropy reduction in farming versus software versus teaching versus art? The Entropy Scaling Function (ESF) must normalize across radically different domains without imposing a single metric that Goodharts itself. No satisfactory universal ESF exists yet.
Opening-condition disagreement
A kitchen looks clean to one person and disordered to another. That is the opening condition, not the attractor, and it is not the observer effect. Schrödinger’s cat was a reductio against applying quantum recipes at cat-scale. Quantum does not apply to the macro. The bet here is a weighted eight-domain DAG: like-cases stack, ZKPs unveil only what the equation needs, and a number that keeps surviving without negative feedback moves the needle less. Landauer is the claimed conversion for social/cognitive events (information erased has a heat floor) into a bits-equivalent proxy — not a metaphor, and not a frozen joule for a quarrel. Unsolved: whether convergence of the graph is calibration of a physical quantity, or a folk taxonomy that got dense and stable. That unsolved is a research problem, not a reason to recant the sentence, and not a license to start a show with wavefunction collapse.
Local vs. Global Entropy
Air conditioning reduces local entropy (cool room) while increasing global entropy (waste heat, energy generation). Every local entropy reduction has externalities. The framework needs boundary condition logic that accounts for net system-wide effects. Current approach: nested boundary analysis with recursive scope expansion.
Validation
Adoption Density, Not Validator Priesthood
“Who validates the first validators?” overstates the problem. LocalFlow is the errand face: rides, food, the car you don’t have. Confirmation closes the loop. SignalFlow is how the ΔS gets proposed — assistant + PSLL + evidence, not a self-score. Remaining constraint is density: enough people in the same zone to pick the work up. Unsolved: how thin a local graph can get before neglected-work escalation is not enough.
Micro Overselling and MACRO Coordination
A MICRO can puff a lawn. A MACRO that treats those numbers as gospel will drift. That pressure is real and is not denied. The stronger answer is not “people are honest.” It is the DAG as instrument: SignalFlow proposes ΔS, evidence hangs on the vertex, like-cases get referenced, DAG curators (paid in XP) keep the graph navigable, ZKPs keep the diary at the edge, settle/decay/late-burn still apply, votes stay in the DFAO. Unsolved: whether a dense graph of surviving interpretations is calibration, or a popular story with better footnotes. Also unsolved: purchased anonymized backlogs as training fuel vs contamination from the extractive systems that made them.
AI Validator Recursion
If AI systems validate entropy reduction, and AI systems are themselves subject to alignment failures, you get a recursive trust problem. The current architecture uses a three-layer model (human + AI + physical sensor) but the interaction dynamics between layers are not fully formalized.
Validation Latency
Some entropy reduction is only visible over long time horizons (planting a forest, educating a child, writing a foundational paper). The system needs temporal validation mechanisms that don't penalize slow-burn contributions or reward short-term manipulation.
Governance
DFAO Power Concentration
If authority flows to top entropy reducers, does this create a new elite that entrenches itself? Decay on influence and role rotation are the current approach. Fractal scale is the other: a MACRO that tries to act like a MICRO chat will choke; a MICRO that pretends to be PLANETARY will capture. Defaults change with scale. Caps do not. Unsolved: whether nested DFAOs actually prevent concentration or just hide it one layer down.
Cultural Resistance
Institutions that have been Goodharted for decades will resist transition to entropy-based metrics because it threatens existing power structures. The framework has no built-in mechanism for peaceful adoption at institutional scale. This is a political problem, not a technical one.
Economics
XP and Existing Markets
How does a non-transferable, non-speculative value unit interact with existing monetary systems? XP cannot replace money overnight. The bridge mechanics between entropy-based value and market-based value are underspecified.
Incentive Gaming at Scale
At sufficient scale, sophisticated actors will find ways to game entropy metrics just as they game every other metric. The framework's defense is recursive auditing, but the cat-and-mouse dynamics at civilizational scale are unpredictable.
Experiments
ΔS calibration
Take a hundred identical real tasks. Independent observers and models estimate ΔS. Watch downstream effects. Measure prediction error → correction → convergence. If the proxy does not get cheaper to keep wrong, the architecture failed. This is the first empirical test. Not the slogan.
Farming resistance
Assume everyone is trying to manufacture XP. Attack R (split one job into 400 fake-rare classes), F, slam-shut Tₛ, confirmation, bilateral agreement, evidence, late burn. The formula will not catch rarity-splitting. The DAG has to see one underlying operation. If the graph cannot, F and Tₛ are furniture.
L as an extraction machine
If L turns standing into a till spark, a captured DFAO might juice L and harvest EP. The lock: L is live rank against house load, both edges accept, coefficient public, cash cannot mint XP, fridge does not lock. Test whether a house can still extract. This is the economic attack, not a philosophy problem.
Late discovery / late mint
Ordinary work mints a small proxy. Twenty years later the work was vastly more important. Late mint is citation-gated, proxy delta only, not a second full paycheck. Simulate it. If it becomes a retroactive windfall, the bound failed.
Cross-domain (w · E) as a political knob
Lawn, ride, ozone, mediation do not share a constant. The vector is the claim. Demonstrate that it improves predictions rather than becoming a tunable vote. If w moves with faction, not with evidence, the DAG is a sermon.
Philosophy
Is-Ought Bridge
The framework derives value from physics (entropy reduction is good). This is a naturalistic claim that may not survive philosophical scrutiny. Can you derive an ought from an is? The framework's response: entropy reduction is not claimed to be morally good in all cases — it is claimed to be a better measurement anchor than any existing alternative. The justification is pragmatic, not metaphysical.