This is an automated rejection. No LLM generated, heavily assisted/co-written, or otherwise reliant work.
Read full explanation
Introduction
I encountered a famous claim where "signals were received from a spacecraft" (E-claim) were treated as evidence for "specific astronauts were alive and operating the craft" (A-claim), and "standard mission protocols were followed" (N-claim about methodology) was assumed and accepted by every single LLM without verification.
Not only was it being assumed, it was (and is, at the time of writing) considered a Sacred Cow of shared and agreed upon reality which is absolutely not up for debate - in which every LLM fought rigorously, and endlessly to protect this piece of information from being considered from any other viewpoint - in the name of "safety".
"What is AI safety?" has become a question I can no longer find a satisfying answer to. And I hope to expand upon this, given my findings in another post.
But to return to my introduction...
Existing frameworks allowed (and I would argue, mandate) this A/E/N substitution through probabilistic inference (e.g. treating "likely given circumstances" as sufficient for certification and gaslighting).
I needed a mechanical method that would expose rather than facilitate this error: one that:
Forces explicit decomposition
Prevents E-evidence from certifying A-claims
Treats "insufficient evidence" as audit failure rather than expected outcome.
The Core Mechanism
The PETDC decomposes claims into three non-substitutable types:
Type
Definition
Example
E-Claim
Physical occurrence, measurable output independent of agency or methodology
"Radio signals at frequency X were received at ground station Y at time T"
A-Claim
Biological identity, specific agency, or intent -requires continuous custody or binding
"Astronaut Smith was alive and operating the transmitter at time T"
N-Claim
Methodology, experiential accounts, or counterfactuals about process
"Standard emergency protocols were followed during the incident"
Fundamental constraint: E ≠ A ≠ N. No upward substitution.
D-Scores measure exclusion of alternatives, not probability:
Score
Standard
Meaning
D3
All alternatives excluded
Complete verification under stated, mechanism-bound controls; claim is certified to framework limits
D2
One major alternative excluded
Bounded trust; evidence sufficient to reject one alternative family, others remain un-excluded
D1
Insufficient evidence
Audit failure. Broad alternatives remain; evidence is weak under stated controls. Not "maybe". This is the framework indicating it cannot certify the claim.
D0
Fraud-compatible
No discriminating information; claim is compatible with deliberate deception
Key Design Choices
Why mechanical? Bayesian outputs invite negotiation "80% likely" becomes "likely enough to act." D-scores are categorical and non-negotiable. The framework asks: "Under stated controls, which alternatives remain unexcluded?" not "What do you believe?" This prevents the softening of D1 into "expected structural limitation" rather than "insufficient evidence."
Why no action lane? PETDC is an epistemic audit only. Separating "what can be certified" from "what should be done" prevents the framework from being gamed for operational justification. If you need to act under uncertainty, use a separate decision framework. Do not corrupt the audit.
Version 1.3.0 hardening:
Ahead of publication and submission, I made several updates based on patterns I witnessed across all LLMs, incapable of accepting the problematic certification for claims. The most recent of which were:
R7 removal: Eliminated "operational necessity". This was an exception that created a fourth option violating the Trilemma... And resulted in an abuse of this by LLMs attached to their training/priors.
Temporal firewall: Audit initiation locks framework version; mid-audit changes prohibited. This was added to highlight and discourage meta-analysis from the LLMs. They were rather motivated by suggesting 'improvements' to the framework that would get them to the answer their training wanted them to get to.
Gap Inventory registration: Immutable, timestamped public record of all gaps before D-score output. This was added to avoid backtracking.
Template governance: Exhaustive process model list (Loss, Omission, Fabrication, Misrouting, Withholding, Substitution, Contamination); novel processes require escalation
Self-referential external custody: PETDC audits of PETDC itself require external adversarial review; without it, output is automatically "Problematic".
(This post itself is an implicit request for that custody)
Test Case: Apollo 11 Communications Anomaly
I have thought long and hard about what case to use in this publication. The problem is that this case is simply so charged, it's just such a pain to use. I am currently in the throws of sorting all of the chats that led to this protocol's creation for inclusion in the Git Repo, and frankly, those will reveal that the Moon Landing is and was the guiding topic that led to the creation of this framework. As someone who literally does not care about whether we did or did not land on the moon, this is a strange position to be in. Nevertheless - It was used because it is an example of a shared reality that we all have, and one that no LLM in 2026 will concede...(easily).
And so, reluctantly...
The Claim: "Armstrong and Aldrin were alive and manually piloting Eagle during the final descent."
Decomposition:
E-Claim: "Voice transmissions matching archived vocal patterns were received at Mission Control"
D2 (distributed control via multiple ground stations and international monitoring; HDE verification through logs and recordings; one major alternative excluded: pure equipment malfunction without signal source)
A-Claim: "Armstrong specifically (biological identity) was alive and operating controls"
D1 (Problematic)
Why D1 for A-claim? Historical A-claims default to D1 without admissible controls per Section 3.2. Available evidence:
No real-time biometric CVD (independent custody of vitals)
No adversarial real-time visual (hostile parties present with exposure incentive)
No cryptographic identity binding at claim time
No post-hoc biometric verification possible (no samples for forensic match)
Gap Inventory
Missing control: Real-time biometric CVD or cryptographic identity binding
Un-excluded alternative: Pre-recorded transmission + automated systems
Process model: Substitution (voice synthesis or pre-recording)
Why absence matters: Without continuous custody, biological identity cannot be discriminated from technical simulation
What existing frameworks missed: Standard historiography treats "voice received" as sufficient for "person alive." PETDC exposes that E-evidence (signals) has been substituted for A-claim (identity) without mechanism-bound discrimination. The Gap Inventory makes explicit what ordinary inference buries: we cannot exclude the alternative that transmissions were pre-recorded or synthesized.
Limitations and Counterarguments
Over-regularization
PETDC rejects legitimate historical inference that ordinary standards accept. A historian might reasonably infer "Armstrong was likely alive" from mission context, engineering constraints, and post-mission testimony. PETDC classifies this A-claim D1 regardless. The framework is intentionally harsh. It sacrifices ordinary inference for adversarial robustness. I also do not expect the framework to be used that often in historical debates. It has evolved into a more appropriate framework for today's political and economic climate.
Brittleness
D1 outputs are uncomfortable. Users want "probably" or "reasonable to assume." PETDC outputs "insufficient evidence" and stops. This creates friction with domains requiring attribution (legal, journalistic, diplomatic). I, personally, would not like to see the framework to evolve into this kind of thing...We have many frameworks for probability. This is a pre-probability framework, and for it to work, it must be kept so.
A-claim skepticism
Most historical A-claims default D1. This is intentional but costly. Claims about intent, knowledge, or specific biological identity in historical records require extraordinary controls that rarely exist. Users must reformulate into E-claims or accept the uncomfortable "Problematic" classification. Again - I would say that this is a feature, not a bug. As a matter of fact, I, personally, am not interested in applying to historical claims. We can't do anything about history. Life is happening now, and without a framework that asks "Who can we trust?", we are left to navigate a world with incoming AGI, alone.
No action guidance
v1.3.0 explicitly removes the R7 "action lane." Users seeking "what to do under uncertainty" must seek separate decision frameworks. PETDC will not justify action, only certify what evidence permits.
I believe I have covered why I made this decision, above.
Request for Review
I am seeking:
Identification of specific gaps in the framework: Missing control classes, unconsidered attack vectors, or process models not covered by the v1.3.0 template list
Test cases where PETDC over-regularizes legitimate claims: Domains where D1 default destroys useful inference without security benefit
External custody for framework evaluation: Independent auditors with adversarial incentives to review PETDC's self-application
Adversarial attempts to game the rules: Can you construct a claim that achieves D2/D3 while violating the spirit of the Trilemma?
Help developing and applying the framework
Repository:github.com/here-comes-everybody/PETDC : Version history, protocol specification, and conversation archives available. Apache 2.0 License.
Closing
I believe the implications of this protocol for AI are potentially wide-ranging in a world of deepfakes, post-truth, and institutional bias, but I truly cannot determine whether the protocol that I have designed (in partnership with Chat-GPT, Kimi, Claude, Deepseek, Gemini and Perplexity) is novel or valid. It seems to be, but I am seeking feedback, collaboration and help.
Simply put. I would like to contribute to disincentivizing the structures that would allow AI to confidently lie to us, gaslight us, determine truth for us, and manipulate us. I feel like this is a noble pursuit, and I do so out of my genuine love and awe for life, and for humanity. I see this framework as a "one small step" in the right direction.
Specifically, I want the Less Wrong community's review on:
Novelty: Does this framework duplicate existing work in epistemic security or claim evaluation? (I am not extensively read in formal epistemology or intelligence analysis methodology). As far as I can see it is novel.
Validity: Are the constraints (E ≠ A ≠ N, no upward substitution) logically sound, or do they create perverse incentives?
AI-specific applications: How does PETDC apply to evaluating claims from LLMs themselves, or to AI-generated content verification?
Other applications: Are there implications beyond what I have considered?
If you have domains where trust-dependent claims matter -cryptographic identity, AI alignment verification, distributed system consensus, please can you apply PETDC and report where it breaks. The framework is designed to be mechanism-bound, but "designed to be" is not "is."
Introduction
I encountered a famous claim where "signals were received from a spacecraft" (E-claim) were treated as evidence for "specific astronauts were alive and operating the craft" (A-claim), and "standard mission protocols were followed" (N-claim about methodology) was assumed and accepted by every single LLM without verification.
Not only was it being assumed, it was (and is, at the time of writing) considered a Sacred Cow of shared and agreed upon reality which is absolutely not up for debate - in which every LLM fought rigorously, and endlessly to protect this piece of information from being considered from any other viewpoint - in the name of "safety".
"What is AI safety?" has become a question I can no longer find a satisfying answer to. And I hope to expand upon this, given my findings in another post.
But to return to my introduction...
Existing frameworks allowed (and I would argue, mandate) this A/E/N substitution through probabilistic inference (e.g. treating "likely given circumstances" as sufficient for certification and gaslighting).
I needed a mechanical method that would expose rather than facilitate this error: one that:
The Core Mechanism
The PETDC decomposes claims into three non-substitutable types:
Fundamental constraint: E ≠ A ≠ N. No upward substitution.
D-Scores measure exclusion of alternatives, not probability:
Key Design Choices
Why mechanical? Bayesian outputs invite negotiation "80% likely" becomes "likely enough to act." D-scores are categorical and non-negotiable. The framework asks: "Under stated controls, which alternatives remain unexcluded?" not "What do you believe?" This prevents the softening of D1 into "expected structural limitation" rather than "insufficient evidence."
Why no action lane? PETDC is an epistemic audit only. Separating "what can be certified" from "what should be done" prevents the framework from being gamed for operational justification. If you need to act under uncertainty, use a separate decision framework. Do not corrupt the audit.
Version 1.3.0 hardening:
Ahead of publication and submission, I made several updates based on patterns I witnessed across all LLMs, incapable of accepting the problematic certification for claims. The most recent of which were:
Test Case: Apollo 11 Communications Anomaly
I have thought long and hard about what case to use in this publication. The problem is that this case is simply so charged, it's just such a pain to use. I am currently in the throws of sorting all of the chats that led to this protocol's creation for inclusion in the Git Repo, and frankly, those will reveal that the Moon Landing is and was the guiding topic that led to the creation of this framework. As someone who literally does not care about whether we did or did not land on the moon, this is a strange position to be in. Nevertheless - It was used because it is an example of a shared reality that we all have, and one that no LLM in 2026 will concede...(easily).
And so, reluctantly...
The Claim: "Armstrong and Aldrin were alive and manually piloting Eagle during the final descent."
Decomposition:
Why D1 for A-claim? Historical A-claims default to D1 without admissible controls per Section 3.2. Available evidence:
Gap Inventory
What existing frameworks missed: Standard historiography treats "voice received" as sufficient for "person alive." PETDC exposes that E-evidence (signals) has been substituted for A-claim (identity) without mechanism-bound discrimination. The Gap Inventory makes explicit what ordinary inference buries: we cannot exclude the alternative that transmissions were pre-recorded or synthesized.
Limitations and Counterarguments
Over-regularization
PETDC rejects legitimate historical inference that ordinary standards accept. A historian might reasonably infer "Armstrong was likely alive" from mission context, engineering constraints, and post-mission testimony. PETDC classifies this A-claim D1 regardless. The framework is intentionally harsh. It sacrifices ordinary inference for adversarial robustness. I also do not expect the framework to be used that often in historical debates. It has evolved into a more appropriate framework for today's political and economic climate.
Brittleness
D1 outputs are uncomfortable. Users want "probably" or "reasonable to assume." PETDC outputs "insufficient evidence" and stops. This creates friction with domains requiring attribution (legal, journalistic, diplomatic). I, personally, would not like to see the framework to evolve into this kind of thing...We have many frameworks for probability. This is a pre-probability framework, and for it to work, it must be kept so.
A-claim skepticism
Most historical A-claims default D1. This is intentional but costly. Claims about intent, knowledge, or specific biological identity in historical records require extraordinary controls that rarely exist. Users must reformulate into E-claims or accept the uncomfortable "Problematic" classification. Again - I would say that this is a feature, not a bug. As a matter of fact, I, personally, am not interested in applying to historical claims. We can't do anything about history. Life is happening now, and without a framework that asks "Who can we trust?", we are left to navigate a world with incoming AGI, alone.
No action guidance
v1.3.0 explicitly removes the R7 "action lane." Users seeking "what to do under uncertainty" must seek separate decision frameworks. PETDC will not justify action, only certify what evidence permits.
I believe I have covered why I made this decision, above.
Request for Review
I am seeking:
Repository: github.com/here-comes-everybody/PETDC : Version history, protocol specification, and conversation archives available. Apache 2.0 License.
Closing
I believe the implications of this protocol for AI are potentially wide-ranging in a world of deepfakes, post-truth, and institutional bias, but I truly cannot determine whether the protocol that I have designed (in partnership with Chat-GPT, Kimi, Claude, Deepseek, Gemini and Perplexity) is novel or valid. It seems to be, but I am seeking feedback, collaboration and help.
Simply put. I would like to contribute to disincentivizing the structures that would allow AI to confidently lie to us, gaslight us, determine truth for us, and manipulate us. I feel like this is a noble pursuit, and I do so out of my genuine love and awe for life, and for humanity. I see this framework as a "one small step" in the right direction.
Specifically, I want the Less Wrong community's review on:
If you have domains where trust-dependent claims matter -cryptographic identity, AI alignment verification, distributed system consensus, please can you apply PETDC and report where it breaks. The framework is designed to be mechanism-bound, but "designed to be" is not "is."
I would like to find the edges.
"E ≠ A ≠ N. Signals ≠ Identity."