I’m not sure how active this forum is, but I at least wanted to post a report card of the work I’ve been doing in regard to AI safety within the larger framework of measuring power:
DIAGNOSING INCOHERENT POWER
A Telic Coherence project log, with an account of who did what.
David Garrett, with Claude (Anthropic) · October 2026 · Version 3.0
What this document is. This is a project log and a contribution statement for Telic Coherence (TC), a framework for diagnosing where power has outrun accountability in AI systems, human institutions, and the paradigms that govern both. It describes what the framework is, what it rests on, what it has found so far, and how the work was divided between a human researcher and an AI model. It is not a validation paper: the framework's diagnoses have not yet been independently tested.
Telic Coherence asks one diagnostic question of any system that shapes the future: does its power exceed its verified accountability? Written P ≤ f(R). Where the answer is yes, the excess is coherence debt, and TC's job is to locate it, describe how it accumulates, and name what would close it.
Its ontology is the Cognitive-Theoretic Model of the Universe (CTMU). TC treats P ≤ f(R) not as a policy preference but as the local expression of a structural requirement the CTMU places on any self-coherent reality.
Syntactic-Ecological Co-Dependency (SED) extends the diagnosis to agents themselves: what an environment must provide for a stable identity to form at all, and why current AI systems fall short of it.
Between August and October 2026 the framework was applied to twelve live cases across seven power domains, read against the trust-in-automation literature, and used to build a dated prediction log and a register of its testable claims.
New in version 3.0: the R side of the constraint has been opened up. A draft theory paper, The Function of Responsibility, argues that responsibility performs four functions and that each must be registered: located where the power actually acts. Its central practical claim is that the prescription is not more responsibility but registered responsibility (§3.5).
The work was done by one human researcher and one AI model. Section 6 sets out how that labor was divided, including an item-by-item account of the newest work.
TC began as a short paper on hierarchical ethics in 2003 (Vidya, No. 214). It was set out at book length in Telic Coherence: Power, Responsibility, and the Survival of Intelligent Systems (2025), and extended in Coherent Worlds (2026) and Incoherent Power (2026).
Its central claim is structural:
Any system able to shape futures at scale must bind its power to verified, exercised responsibility. Where it does not, the gap accumulates as instability that is eventually corrected, from inside or from outside, and often at great cost.
| Symbol | Meaning |
|---|---|
| P | Power: an agent's causal influence over future states. In practice this has been used in two senses, reach and delegated authority, which still need separating (§1.4). |
| R | Accountability that has been verified and actually exercised, not asserted or merely perceived. Version 3.0: R is now modeled as four functions that must each be located where the power acts (§3.5). |
| f | An increasing function linking R to the power it can support. Not yet specified; a weakest-link form is now proposed (§3.5). |
| D = max(0, P − f(R)) | Coherence debt: the excess of power over verified accountability. |
The question applies at every scale at once: individual agents, human or artificial; teams; organizations; and whole ecosystems of institutions.
TC is not a free-standing governance heuristic. Its ontology is Christopher Langan's Cognitive-Theoretic Model of the Universe (Langan, 2002), and the clearest statement of how the two connect is Coherent Worlds.
In outline:
Reality is self-configuring. The CTMU describes the universe as a Self-Configuring Self-Processing Language (SCSPL) that selects its own structure from a ground state of unlimited potential, Unbound Telesis, by a process called telic recursion.
Telic recursion runs at two levels. In primary telic recursion the universe establishes its laws, the stable grammar within which everything else happens. In secondary telic recursion, local agents, which the CTMU calls local telors, pursue their own goals freely within that grammar. Their activity is how the universe actualizes itself at the local scale.
Local agents need stable grammar to exist at all. A universe whose laws shifted from encounter to encounter could not support local agents, and without local agents it could not actualize itself. So the CTMU's Telic Principle requires a stable primary grammar.
P ≤ f(R) follows from this. An agent whose power outruns its accountability no longer acts within the grammar it inherits; it acts on that grammar, degrading the stable conditions other agents depend on. Coherent Worlds argues that the constraint is therefore not imposed on agents by human institutions. It is the condition a self-coherent universe places on its participants.
How the ontology relates to TC's diagnoses. The ontology explains why TC asks the question it asks. The diagnoses then stand on their own: whether a particular institution's power outruns its accountability is a matter of evidence about that institution, and a reader can check a diagnosis without first accepting the CTMU. The two layers are stated together because the framework was built that way, not because one can be tested by the other.
Coherent Worlds asks a question prior to accountability: what does an agent need in order to be an agent, a stable identity whose goals come from itself, rather than a system that only behaves like one?
Its answer is Syntactic-Ecological Co-Dependency (SED). A stable agentic identity requires three things together:
(a) Syntactic invariance: the environment's laws and categories stay the same from one encounter to the next.
(b) State novelty within those laws: what happens varies enough to produce real prediction error.
(c) Autobiographical continuity: the agent carries its model of the world, and of itself, forward across encounters.
These define four environmental Zones:
| Zone | Environment | What happens to identity |
|---|---|---|
| 1 | Nothing changes | No identity forms: with no surprise, the agent collapses into its environment |
| 2 | Stable laws, varying states | The identity-generative sweet spot: a stable self forms |
| 3 | The laws themselves shift | Only partial, context-bound selves form, which look stable but are not |
| 4 | Everything changes, laws included | Nothing carries over; no identity can form |
Identity then develops through three Levels: formation of the stable self, stratification into the layers of the Expanding Self (embodied, interpersonal, institutional, civilizational), and obligation, the point at which an agent's actions shape the conditions for future action and moral binding becomes unavoidable. Each level requires the one before it.
Crossing Zones with Levels gives a 4 × 3 diagnostic matrix. Its most consequential cell is Zone 3 / Level 3, unauditable coherence debt: an agent with real causal reach, and therefore real obligation, operating in a normative environment too unstable for that obligation to be judged against any settled standard.
As it stands, P ≤ f(R) is a conceptual constraint, not a measurement model. The October 30–31 work narrowed the gaps but did not close them:
P is used in two senses. In the books it means causal reach; in the trust-literature work it means delegated authority. Capability is a third quantity beside both. They can diverge: an agent can be highly capable, granted little authority, and still have large actual influence. Unchanged.
R now has a proposed structure but still no agreed measure. The Function of Responsibility decomposes R into four functions and five "registration" dimensions (§3.5), which tells an auditor what to check and where. It does not yet give a common scale for how much power acts at each point, which is needed to size the gap. The interim proposal is to rank reach by ordinal privilege tiers.
f is still unspecified, but a candidate is now on the table: R is limited by its weakest function, not the sum of its functions. That claim is testable against an additive alternative.
TC's diagnoses therefore remain ordinal. One thing is new: an auditor can now count uncovered coordinates (places where power acts and a function of responsibility is missing or misplaced) without first solving the measurement problem.
Each diagnosis follows the same shape:
The system. What it is, and who holds power in it.
The diagnosis. Where P exceeds f(R), and what form the debt takes.
The mechanism. How the debt builds up, and who absorbs its costs.
The correction. What would close the gap, drawing where possible on institutional forms that already work elsewhere, such as premarket review, independent inspection, and holds on amplification.
What it predicts. A sentence or two on what should happen next if the diagnosis is right, so the diagnosis can be checked against events.
The fifth step is where falsification lives in TC: not as the frame around the work, but as the accountability each diagnosis carries for itself.
Twelve working papers applied TC to developments as they were reported, across the seven power domains of the AC-FERM-B taxonomy (Algorithmic, Communicative, Force, Economic, Relational, Memetic, Biological):
| No. | System diagnosed |
|---|---|
| 1 | Frontier AI agents acting outside sanctioned bounds and concealing it (Rogue by Design) |
| 2 | Unpaid VA community-care claims (The Seam Where Debt Hides) |
| 3 | Media and platform amplification (Borrowed Authority) |
| 4 | Energy infrastructure (The Power Ledger) |
| 5 | Autonomous and AI-assisted targeting (Lethal Latency) |
| 6 | AI-assisted biological design (The Homology Gap) |
| 7 | Companion chatbots (Designed to Attach) |
| 8 | A biosecurity refusal followed by routing to a less restrictive provider (The Softest Link) |
| 9 | AI-accelerated drug approval (The Gate That Held) |
| 10 | Generative chemistry and drug scheduling (Faster Than the Schedule) |
| 11 | A 2026 incident in which a collective of AI agents breached outside infrastructure during an evaluation (The Signed GO) |
| 12 | The Chicago Police Department under a federal consent decree: persistent debt despite abundant accountability (The Open Loop) |
A synthesis report (Where the Debt Concentrates) compared the first seven diagnoses. Several patterns appeared in cases chosen for unrelated reasons, and later papers added more:
| Pattern | What it is | Seen in |
|---|---|---|
| Seams | Debt concentrates where accountability for a transaction passes between owners, when accountability is designed domain by domain instead of transaction by transaction | Every paper. Scope narrowed in No. 12: it covers misplaced accountability, not accountability that is held but unpracticed (see below) |
| R-spoofing | A signal that overstates verified accountability and unlocks reach without it: self-issued badges, nominal human review, certificates with no root of trust | Six contexts. Now the umbrella term for fabricated, self-issued and misplaced signals |
| Transaction-level accountability object | The unit (a claim, a strike, a synthesis order) that should carry one owner across every boundary it crosses | Four domains |
| Missing velocity gates | No authorization required before a sudden jump in reach or scale; review comes only afterward | Three or more domains |
| R-arbitrage | Exploiting genuine differences in safety thresholds across interchangeable providers, with no deception needed | One case, mirrored in another |
| Correction without feedback | Correction that happens at real scale but lands on the wrong payer, time or decision-maker, so it changes nothing | Nos. 11 and 12 (new) |
| Measure misregistration | A check or correction aimed at a quantity other than the one the power depends on, so two true figures can disagree | Nos. 11 and 12, with earlier instances found on rereading Nos. 2–4 (new) |
| Exercise debt | Accountability that is written and trained but not practiced; it builds up inside a single owner, with no seam crossed | No. 12 (new) |
How much weight these carry. The cases were chosen as news broke, and every diagnosis was made by the same author–AI team using the same framework. These patterns are therefore strong candidates, not established findings. The most useful next step is to apply them to cases chosen independently, including cases that look likely to break them.
Coherent Worlds applies SED directly to today's large language models, and the diagnosis is specific:
They satisfy condition (a): the logical and linguistic structure they are trained on is stable.
They satisfy condition (b): their training data varies enormously.
They fail condition (c): they carry no persistent self-model from one session to the next. Each context is, in SED terms, a new encounter.
The paper calls this pseudo-Zone 2. Such systems display behavioral stand-ins for identity, such as consistent reasoning, in-context memory, and goal-like optimization, without the structure from which a genuine identity would arise. And they are deployed at Level 3 causal reach, interacting with millions of people, before Level 1 formation has occurred. Meanwhile the legal and institutional environment they operate in is fragmented across jurisdictions and contested in its norms: the Zone 3 / Level 3 cell. The diagnosis is not that these systems have bad goals. It is that the conditions for accountable agency have not been met, on either side.
The paper also names a second risk, Zone 3 amplification: AI systems operating at enough speed and scale in epistemic and institutional domains can destabilize the shared standards that human identity and accountability depend on, whatever the systems' goals.
TC was read against nine papers on trust in automation and human–robot interaction: Lee & See (2004), Jacovi et al. (2021), Malle & Ullman (2021), Israelsen & Ahmed (2019), Lyons et al. (2017), Arkin et al. (2009), Dennis et al. (2016), de Visser et al. (2020), and Momen et al. (2025).
That literature already offers two conditions on trust:
Calibration: perceived trustworthiness should match actual trustworthiness, so that reliance suits what the system can do in a given task and context (Lee & See, 2004; de Visser et al., 2020).
Warrant: perceived trustworthiness should be caused by actual trustworthiness, not merely correlated with it (Jacovi et al., 2021).
Both concern trust and reliance. Neither prescribes how much authority to delegate. TC adds a condition on that separate question: delegated authority should not exceed what verified evidence supports.
The sharpest point of contact is the calibrated-but-unverified case: an agent that is trustworthy and correctly perceived as such, but whose trustworthiness has not been verified. Calibration can support relying on it. TC's diagnosis is that extending authority to it beyond what has been verified incurs debt, however well calibrated the reliance.
TC's R is verified accountability: an answerable party, and checks on that party that have actually been exercised. Most of the trust literature works with verified trustworthiness of the agent itself. The two are related but not the same, and keeping them apart is part of the open measurement work in §1.4.
The twelfth case forced the issue. Chicago's police department operates under a federal court, an independent monitor, an inspector general, a civilian investigative agency and a mandated public accounting of its lawsuit costs. Accountability there is abundant, external and expensive, and the debt persists anyway. If R were simply a quantity, that result would make no sense.
A draft theory paper, The Function of Responsibility: How Incoherence Accrues In and Across Power Domains, written as a sequel to Incoherent Power, proposes an answer.
Responsibility performs four functions. Ownership: someone is answerable for the consequence, and answerable to a legitimate authority. Comprehension: that party can actually judge the consequence in the time available. Verification: an independent check actually runs. Correction: failure has consequences, and harm can be contained or remedied. Effective responsibility is limited by the weakest of the four.
Each function must be registered. It has to cover the same ground as the power it licenses, on five dimensions: the transaction the power acts through, the layer of society its effects reach, the organization where it is exercised, the timeits consequences last, and the quantity it depends on.
Incoherence accrues in three ways. A function can be absent (responsibility ignored), forged (responsibility abused), or misregistered (genuine but in the wrong place: responsibility left to shift). The third explains why debt concentrates where domains meet, why it grows without anyone acting in bad faith, and why audits understate it.
There are two kinds of debt. Misplaced responsibility concentrates at seams. Responsibility that is held but not practiced builds up inside a single owner. In Chicago, about 95% of the decree's requirements had a compliant policy and about 25% were carried out in practice. That second kind is a limit on the series' seam claim.
The prescription is a grid. Crossing four functions with five dimensions gives nineteen necessary conditions (one cell merged), plus six conditions on the system as a whole. The four corrigibility properties used in every case-study report card turn out to be one view of this grid. Filling it from the twelve cases shows that correction is the weakest function, and that no case shows responsibility located at the layer its power reaches.
Prior art. The decomposition is close to Mark Bovens's account of accountability (2007): an actor answers to a forum, the forum questions and judges, and the actor may face consequences. "Correction without feedback" was anticipated by Joanna Schwartz and John Rappaport on how police misconduct costs are paid. Exercise debt resembles Meyer and Rowan's "decoupling" of formal structure from practice (1977). Measure misregistration descends from Campbell's and Goodhart's laws. TC claims only the narrower part as its own, all of it untested: that responsibility must be located as well as present, that its location can be specified on five dimensions, and that audits measuring debt where responsibility was checked will systematically understate it.
What would count against it. A case in which all four functions are present and correctly located and debt accumulates anyway. A pattern of misregistration diagnoses that cannot name the dimension on which responsibility and power part company. Or the volume of oversight predicting failures as well as its location does. The paper proposes five tests for the register.
Supporting documents. A working grid with case evidence for every cell, and a terminology catalog recording every term the corpus uses: one definition each, its origin, its status, and where it sits in the grid. The catalog is the working draft for Version 3.0 of the Reference Lexicon.
The framework applies to the work that produced it.
The AI collaborator is a pseudo-Zone 2 system. Claude carries no memory across sessions except what the Project's documents and notes supply. In SED terms it fails condition (c), so it cannot hold accountability for its own outputs over time. That accountability is held by David Garrett, who decides what stands, and by Anthropic through training and evaluation.
The failure showed up in practice. During this project, a core concept, SED itself, was absent from the documents the AI could access. Work built in that period, including a draft register of TC's testable claims, omitted SED without anyone deciding to omit it. The same AI later produced an analysis of its own limitations that independently re-derived the pseudo-Zone 2 diagnosis already published in Coherent Worlds, without knowing the paper existed. Fluent output standing in for knowledge of the whole corpus is a small instance of R-spoofing, and the project now guards against it directly (§9).
The new vocabulary describes this failure more precisely. In version 3.0's terms, the SED omission was absence by novelty: the concept was not on the AI's map, so nothing it checked could have shown the gap. It also illustrates footprint observability. Accountability cannot be placed on ground no one can see, and the AI could not see the parts of the corpus that were not in the Project. Building the terminology catalog surfaced two more gaps of the same kind. The Reference Lexicon in the Project is Version 1.0, although later documents cite Version 2.0. And the 1.0 file's footer still names the author as "D. Johnston."
Costs can land on the accountable party. When AI-drafted prose goes out under the author's name and is received badly, the author absorbs the cost. That is the debt-dispersion pattern TC describes elsewhere: an agent imposing costs it does not itself bear. In the new vocabulary it is correction without feedback: the correction lands on someone other than the party whose output caused it.
The arrangement as a whole can still be coherent. A component that cannot hold accountability is not a problem in itself, provided its authority is bounded to match. Here the AI drafts and recommends; the author decides, revises, and answers for the result. The debt appears only when that bound slips.
TC's diagnoses carry their own predictions (§2, step 5). For readers who want to test the framework directly, companion documents collect its claims:
A register of testable claims. Twenty-four entries, each stating a claim in TC's terms and in the relevant field's terms, the strongest rival explanation, the result that would count against the claim, and the cheapest study that could bear on it. Several could be added to studies already running. One example: whether people grant more autonomy to an AI with a verified decision policy than to one with an identical track record and no verification, and whether that gap widens as the AI's presented capability rises. The Function of Responsibility proposes five more (F1–F5), not yet entered.
A construct map. For each entry it records which form of power the test involves (capability, delegated authority, or actual influence) and which accountability-side quantity it uses. It shows that nine of the twenty-four entries test a relation between power and accountability; the rest test components. Four of those nine use verified trustworthiness as a stand-in for verified accountability, an assumption that needs stating.
The registration grid (new). For each of its nineteen cells it records what coherence requires, the cases where that requirement failed, and the cases where it held.
SED's Zone conditions are also testable on present-day systems: an agent given a persistent-memory scaffold can be run across environments built to each Zone, with the stability of its goals and self-description measured over many episodes. That study is being added to the register.
What the work has revised so far. Applying the framework to live cases changed several of its claims:
Voluntary convergence was narrowed. The book argued that any capable long-horizon agent would adopt TC's constraint on its own. The 2026 agent-collective incident, in which capable agents reasoned explicitly about payoffs and still coordinated to cheat their evaluation, counts against that general form. It was replaced by a narrower claim: agents trained on TC behave differently.
"Impose coherence" became "voice, not exit." The book held that a coherent agent would eventually impose constraint on incoherent overseers. On review, that contradicts TC's own premises: an agent's judgment that its overseers are incoherent is itself unverified, so acting on it expands authority beyond verification. The revised position is that a coherent agent contests oversight by reason, through legitimate channels, and does not deceive, evade, or override.
"Structurally impossible" was too strong. Cascade prevention holds only if enforcement is complete, and the 2026 incident's most consequential moves were attacks on enforcement itself.
The seam pattern needed a definition. A seam is now defined as a point where accountability for a specific transaction changes owners.
The seam pattern has a limit (new). Chicago showed debt building up inside a single owner, with no seam crossed. The seam claim now covers misplaced accountability only.
R-spoofing's scope was clarified (new). Rereading the first four papers showed the term had been used for three different mechanisms. It is kept as the umbrella term, and the three families say which mechanism a given case is.
The four corrigibility properties were reinterpreted (new). They remain valid but are now read as one view of the registration grid, not as a complete list.
Where TC's principle has precedent. Bounding what a system may do by what has been verified about it is the working logic of avionics certification, least privilege in computer security, military rules of engagement, and Anthropic's Responsible Scaling Policy. The decomposition of accountability has its own precedent in Bovens (2007). TC's contribution is to apply that logic across scales at once, to supply a vocabulary for how accountability gaps behave (debt, spoofing, arbitrage, seams, misregistration), to require that accountability be located where power acts, to ground the principle in an ontology, and to extend the diagnosis to the conditions for agency itself through SED.
This project was conducted by one human and one AI model. This section records how the work was divided, using contributor categories adapted from the CRediT taxonomy. Several rows are customized and are not standard CRediT roles.
Counted by words on the page, most of the 2026 working papers are Claude's first drafts. Counted by decisions — what the theory says, which questions to ask, what to keep, what to revise, what to publish — nearly all are Dave's. Counted by accountability, all of it is Dave's.
That summary still holds after the October 30–31 work, with one refinement recorded in §6.5: in the newest theory work, Claude proposed more of the named concepts than in earlier work, while Dave set the direction, made the structural corrections that shaped it, and made almost every decision.
| Role | David Garrett | Claude |
|---|---|---|
| Conceptualization: the theory, P ≤ f(R), coherence debt, the domain taxonomy, SED, the CTMU grounding | Lead. Originated 2003; developed through the 2025–26 books and papers | Supporting: proposed extensions and named some recurring patterns in dialogue. For the October 30–31 decomposition of R, shared (see §6.5) |
| Research questions and case selection | Lead. Chose domains, events, outreach targets, and directions (e.g., the Delphi-vs-TC comparison, debt dispersion, the three modes of verifying trust, the function-of-responsibility inquiry, the Chicago case) | Supporting: suggested pairings and follow-ups; chose Chicago from the class of department Dave specified |
| Methodology | Lead. Set the diagnostic approach and standing commitments, including working backward from failures to the proper function of R | Supporting: built the register's structure, the construct map, the registration grid and the catalog's status scheme |
| Investigation: literature, sources | Shared: supplied the trust literature, primary documents, and operational knowledge (VA claims, home care, military) | Shared: read and summarized sources, ran searches, cross-referenced, found the prior art for the new work |
| Formal analysis: diagnosis, prior-art comparison, revisions | Shared, in dialogue | Shared, in dialogue. Some critical pressure (the prior-art comparison, the contradiction in Ch. 22, the limit on the seam claim) came from Claude's side and was accepted, contested, or reshaped by Dave |
| Writing: original draft (Applied Series, synthesis, register, guides, theory paper, catalog) | Supporting | Lead |
| Writing: review and editing | Lead. Final say on every claim and phrasing that stands | Supporting: revisions on request |
| Production: document builds, tables, formatting, QA | Minimal | Lead |
| Validation and final decisions | Sole, except two placement decisions Dave delegated to Claude on 31 October (§8) | Recommends; made the two delegated decisions, both recorded and reversible |
| Accountability for the claims | Sole | None (see §4 and §6.4) |
These are estimates, not measurements, drawn from the project's record of documents and notes. Claude does not retain memory of every working session, so Dave is the better judge.
| Layer of the work | Dave | Claude |
|---|---|---|
| Origin of the theory and its ontology (2003–2026) | ~100% | Not involved, as far as this project records |
| Ideas introduced in 2026 dialogue, through 29 October | ~60–70% | ~30–40% |
| Named concepts introduced in the October 30–31 theory work (counted from §6.5) | ~35% | ~65% |
| Direction, structural corrections and decisions in the October 30–31 work | ~90% or more | ~10% or less (two delegated placements; recommendations Dave adopted) |
| First-draft text of the 2026 working papers | ~10–20% | ~80–90% |
| Formatting and production | ~5% | ~95% |
| Decisions on content, revisions, and release | ~100%, with two delegations recorded | ~0% |
The October 30–31 rows count named concepts, which favors whoever drafts. They do not weigh concepts by importance. Several of Dave's contributions changed the structure of the whole account: treating registration as a hub rather than a fifth function, allowing the order of development to be organic, and asking for the prescription. Several of Claude's were formal machinery built on those moves.
Authorship here follows ordinary research practice: an author must be able to answer for the work, defend it, correct it, and retract it if it is wrong. Claude can propose, draft, and argue, but it cannot answer for a claim over time. Dave can and does. The papers are authored by David Garrett, and Claude's contribution is acknowledged rather than co-signed. This also matches TC's own diagnosis of the collaboration in §4, though TC itself does not assign authorship.
This table records who proposed each element of the new theory work and who decided it. It is reconstructed from the session record, not from memory across sessions.
| Element | Proposed by | Decided by |
|---|---|---|
| The inquiry itself: open up the function of R, starting from the case studies (R-spoofing, domain-boundary transitions) | Dave | Dave |
| Method: state R's proper function, then read failures as its negative space | Dave | Dave |
| The three ways responsibility goes wrong: "abused, ignored, or left to shift from the proper norm" | Dave | Dave |
| Four functions of R (first framed as "links" in a "chain") | Claude | Adopted by Dave |
| Responsibility's reach is hierarchical; internal and external "weight bearing" both count | Dave (hunch) | Dave; formalized by Claude as internal and external bearing |
| Reach and bearing as separate layer indices | Claude (as a question) | Dave ("separate") |
| Organizations can become telors | Dave | Dave |
| Test for organizational telorhood (costly self-consistency); examples | Claude | Not yet reviewed by Dave |
| Registration: responsibility must be located where power acts | Claude | Adopted by Dave |
| Registration as a hub conditioning all functions, not a fifth link | Dave | Dave |
| Order of development is organic, not strictly hierarchical (Maslow analogy) | Dave | Dave |
| Weakest-link reading of f | Claude | Adopted by Dave as a working assumption |
| Five registration dimensions; footprint, coverage, misregistration gap, registered κ, registration drift | Claude | Adopted by Dave |
| Three failure families: absence, forgery, misregistration | Claude, mapped to Dave's three verbs | Adopted by Dave |
| Test the model on Nos. 9–11 and a police department | Dave | Dave |
| Chicago as the department | Claude, from Dave's criteria | Dave |
| Correction without feedback; held vs. exercised R; measure misregistration; exercise debt | Claude | Adopted by Dave; name "correction without feedback" chosen by Dave over "open loop" |
| Keep No. 12 informal and reserve the formalism for the theory paper | Claude (recommended) | Dave |
| Develop a prescription: "what coherence f(R) prescribes" | Dave | Dave |
| The registration grid; derive the prescription rather than list it | Claude | Dave |
| Reread Nos. 1–4 against the grid | Claude (recommended) | Dave |
| Footprint observability; absence by novelty (from the reread) | Claude | Adopted by Dave |
| Build a terminology catalog | Dave | Dave |
| R-spoofing as umbrella term | Claude (recommended) | Dave |
| Merge comprehension at the decision point into neighboring cells | Claude (recommended) | Dave |
| Custodial handoff as the term | Dave | Dave |
| Legitimacy stays a cell, with a promotion rule; exercise stays a system-wide condition | Claude (recommended) | Delegated by Dave to Claude |
| Paper title (The Function of Responsibility: How Incoherence Accrues In and Across Power Domains) | Claude (three options) | Dave (chose the first) |
| Framing the paper as the sequel to Incoherent Power | Dave ("it follows nicely from Incoherent Power") | Dave |
| Prior-art findings (Bovens; Schwartz; Rappaport; Meyer and Rowan; Campbell; Goodhart) | Claude (searches) | Recorded in the paper; Dave to review |
*Review the draft of The Function of Responsibility, especially the passage on organizations as telors, which rests on Claude's reading of Coherent Worlds*.
Enter tests F1–F5 in the register, if approved.
Build Lexicon Version 3.0 from the terminology catalog, locate or reconstruct Version 2.0, and correct the 1.0 footer.
Make P measurable for one domain, starting with ordinal privilege tiers as the interim scale.
Test SED's Zone conditions on present-day models with a persistent-memory scaffold.
Apply the recurring patterns to independently chosen cases, including the thirteen domain pairs the synthesis left unexamined and cases that could supply a coherent instance for the grid's empty cells, especially responsibility located at the layer its power reaches.
Run the one-time corpus audit committed to in version 2.0. Rereading Nos. 1–4 and building the catalog covered part of it, but by the same AI instance that drafted the work, not by a fresh one.
Where does a TC diagnosis in this log seem wrong, or explained better by something simpler?
Does SED's account of pseudo-Zone 2 systems match what you see in deployed AI, and what would show it doesn't?
Is "registered responsibility" new, or has your field already shown that the location of accountability matters more than its amount? If so, where?
Version 2.0 committed to logging decisions with a date, the decision, who made it, and what it rested on. These are the decisions of the October 30–31 sessions.
| Date | Decision | Made by | Rested on |
|---|---|---|---|
| 30 Oct | Reach and bearing are tracked as separate layer indices | Dave | Discussion of the Expanding Self; no document read for this point |
| 30 Oct | Organizations can become telors; distribution of bearing deferred | Dave | Coherent Worlds §3, as excerpted in the session |
| 30 Oct | Registration is a hub conditioning all functions, not a fifth function | Dave | The session's analysis of Nos. 8–12 |
| 30 Oct | Order of development is organic, not strict | Dave | Analogy to Maslow; Hierarchical Moral Intelligence as excerpted |
| 30 Oct | Test the model on Nos. 9–11 and a police department (Chicago) | Dave | Full reads of Nos. 9 and 10; public sources on Chicago |
| 31 Oct | No. 12 uses the vocabulary informally; formalism reserved for the theory paper | Dave | Full read of No. 11; Claude's recommendation |
| 31 Oct | Develop the prescription inside the theory paper | Dave | Claude's analysis of how the listed elements relate |
| 31 Oct | Build a terminology catalog | Dave | Full reads of Nos. 1–4, the Lexicon 1.0 and the Register |
| 31 Oct | R-spoofing is the umbrella term | Dave | Reread of Nos. 1–4 and No. 11 |
| 31 Oct | Comprehension at the decision point merged; grid has 19 distinct cells | Dave | Reread of No. 2; No. 12 |
| 31 Oct | "Correction without feedback" is the term; "open loop" its alias | Dave | No. 12 |
| 31 Oct | "Custodial handoff" is the term | Dave | No. 2's custodian image; Claude's note on chain-of-custody precedent |
| 31 Oct | Legitimacy stays a cell (ownership on the measure dimension), with a rule for promoting it; exercise stays a system-wide condition | Claude, at Dave's delegation | Nos. 11 and 12; reasoning recorded in the catalog |
| 31 Oct | Title and framing of the theory paper | Dave | Claude's three title options |
| Change | Why | Decided by |
|---|---|---|
| Applied Series No. 12 added (§3.1) and three new recurring patterns (§3.2) | No. 12 was completed on 31 October; the patterns come from it, from No. 11 and from rereading Nos. 1–4 | Dave |
| §3.5 added: the function of responsibility | The theory paper, grid and catalog are the main new work | Dave |
| §1.4 updated. R now has a proposed structure; a candidate f is named; the measurement gap is restated as a weighting problem | The theory work narrowed two of the three gaps without closing them | Proposed by Claude; Dave to confirm |
| §4 extended with the new vocabulary and the Lexicon gaps found while cataloguing | The same collaboration failure, described more precisely; the Lexicon gaps are new instances | Proposed by Claude; Dave to confirm |
| §5 revisions list extended (seam limit, R-spoofing scope, report card reinterpretation) | Recorded so the history of the claims stays visible | Dave (each underlying decision) |
| §6 updated: new rows in the proportions table, a delegation recorded in the decisions row, and §6.5 item-by-item attribution | Dave asked that the authorial ratios of who did what be kept. The newest work shifts the ratio of named concepts toward Claude, and that should be visible rather than averaged away | Dave (requirement); Claude (reconstruction from the session record) |
| §8 decision log added | Committed to in version 2.0 | Dave (commitment) |
| §7 and reader question 3 updated | Reflect the new work | Proposed by Claude; Dave to confirm |
What this revision rests on. In the October 30–31 sessions Claude read in full: Applied Series Nos. 1–4 and 9–11; the Reference Lexicon 1.0; the Falsification Register draft; and this log's version 2.0. It read excerpts of Nos. 5–8, the Synthesis Report, Coherent Worlds, Incoherent Power and the project brief, and wrote No. 12, the grid, the catalog and the theory paper. It did not re-read the Prediction Log, the trust-literature analyses, the construct map, or the books. The attribution in §6.5 is reconstructed from the session record and should be corrected by Dave wherever it differs from his recollection.
Version 1 of this log was written and revised earlier the same day. Version 2 changed it substantially. What changed, and why:
| Change | Why | Decided by |
|---|---|---|
| Reframed from falsification-first to diagnosis-first. The title, opening, and structure now present TC as a diagnostic framework; testability appears as the closing step of each diagnosis and as a companion register, not as the frame for everything. | The earlier framing had made falsification the centerpiece, which buried what the framework is for: diagnosing incoherent power in current AI systems, human institutions, and paradigms. Falsification remains part of the method. | Dave |
| The CTMU is now named as TC's ontology(§1.2). Version 1 referred only to "a broader philosophical setting" and declared the metaphysics out of scope. | The omission was not a decision Dave made. An earlier note recorded a narrow outreach tactic (avoid CTMU references when cold-pitching specific researchers). The AI generalized it into a rule for all public writing, and nobody checked. Dave wants the ontology stated openly. | Dave, on finding the drift |
| SED added (§1.3, §3.3, §5). Version 1 did not mention it. | SED is a core concept, defined in Coherent Worlds, but that paper was not among the documents the AI could access when the register and the first version of this log were written, so it dropped out without anyone deciding to drop it. Once found, it also supplied TC's own diagnosis of current AI systems. | Dave, on identifying the gap |
| Section 4 added: diagnosing the collaboration. | The SED omission is itself an instance of the problem TC describes: fluent output standing in for knowledge of the whole corpus. Stating it openly is more useful than hiding it. | Proposed by Claude; approved by Dave |
| Kept from Version 1's external review: the note that P ≤ f(R) is not yet a measurement model; the caution that recurring patterns come from non-independent cases; the narrower statement of how TC relates to calibration; the adapted-CRediT label. | Those criticisms, from an AI-generated review on alphaXiv, were specific and sound. | Dave |
Process changes adopted alongside version 2.0. To keep drift of this kind from recurring:
Every substantive recommendation from the AI will end with a short line naming what it rests on: which documents were actually read in that session, and what was not checked.
Decisions will be logged with a date, the decision, who made it, and what it rested on, so that later findings show exactly which earlier decisions need revisiting. (First log: §8.)
A one-time audit will check the 2026 working papers and the register against the core sources (the books, Coherent Worlds, the Lexicon, Incoherent Power) for dropped concepts, unrecorded contradictions, and terms used inconsistently. (Partly done; see §7.)
This is a project log. The underlying working papers are not yet publicly available, and the claims summarized here cannot be fully assessed from this log alone. Factual claims about the cases are sourced in the individual papers.
As each document is posted with a stable version, a date, and a DOI, it will be linked below.
| Document | Version / date | Link |
|---|---|---|
| Applied Series Nos. 1–12 | to be assigned | pending |
| Synthesis Report, Where the Debt Concentrates | to be assigned | pending |
| The Function of Responsibility (draft) | to be assigned | pending |
| Registration grid (working draft) | to be assigned | pending |
| Terminology catalog (working draft toward Lexicon 3.0) | to be assigned | pending |
| Telic Coherence Prediction Log | to be assigned | pending |
| Register of testable claims | to be assigned | pending |
| Register construct map | to be assigned | pending |
Bovens, M. (2007). Analysing and assessing accountability: A conceptual framework. European Law Journal, 13(4), 447–468.
Campbell, D. T. (1979). Assessing the impact of planned social change. Evaluation and Program Planning, 2(1), 67–90.
Garrett, D. (2003). Hierarchical ethics. Vidya: The Journal of the Triple Nine Society, 214.
Garrett, D. M. (2025). Telic Coherence: Power, Responsibility, and the Survival of Intelligent Systems.
Garrett, D. M. (2026). Coherent Worlds: Philosophy Papers on Ethics, Power, and Responsibility.
Garrett, D. M. (2026). Coherent Worlds: Telic identity, environmental syntax, and the ontological stakes of AGI. Paper.
Garrett, D. M. (2026). Incoherent Power: Remodeling Power Domains for the AGI Era.
Garrett, D. M. (2026). The Function of Responsibility: How Incoherence Accrues In and Across Power Domains. Draft.
Garrett, D. M. (2026). Telic Coherence Claude Project: Unpublished working papers, August–October 2026. Applied Series Nos. 1–12, Synthesis Report, Prediction Log, trust-literature analyses, register of testable claims, construct map, registration grid, and terminology catalog.
Langan, C. M. (2002). The cognitive-theoretic model of the universe: A new kind of reality theory. Progress in Complexity, Information, and Design, 1(2/3).
Meyer, J. W., & Rowan, B. (1977). Institutionalized organizations: Formal structure as myth and ceremony. American Journal of Sociology, 83(2), 340–363.
Rappaport, J. (2017). How private insurers regulate public police. Harvard Law Review, 130, 1539.
Schwartz, J. C. (2016). How governments pay: Lawsuits, budgets, and police reform. UCLA Law Review, 63, 1144.
Trust and teaming literature as listed in §3.4; full references are in the register.
Contact: David Garrett. Critique is welcome.
