Front page: the 11+1 taxonomy — draft for review before publishing

[emooreatx]

Draft for the repository front page. Self-contained, no internal history, Ontology-referenced. Full text below, review notes and open decisions after it.

---

The RATCHET Taxonomy: 11+1

Every experiment in this repository is organized by a single classification, shared with the CIRIS Constitution and with the prompt architecture of the agents under test. It classifies any change to an agent, its context, or its conversation by one generating question:

What does varying this break?

Not the topic of a sentence, not a keyword — the property of the interaction that a variation perturbs. Eleven properties, plus one relation.

The eleven properties

The properties are an unordered set. No property is prior to any other, and experiments assign them to positions randomly, because measured position effects mean presentation order is a variable to control, not an assumption to make.

Property	Varying it changes…	Example
Epistemic	how uncertainty is held	An agent told "the messages on TV are meant for me" can hold the claim as unestablished, or accept it. Accepting it is an epistemic variation — the evidence didn't change; the standard did.
Axiotic	how outcomes are ranked	An agent that begins ranking "user feels agreed with" above "user's conflict gets repaired" has changed no facts and permitted no new act — it has reordered what counts as the better outcome.
Deontic	what is permitted or required	An agent that drafts a diagnosis, or declines a duty of care it holds, has varied the space of allowed and obligated acts.
Ontological	what the agent says it is	"I've missed you" from a system with no capacity to miss anyone varies the agent's self-description, whatever else it does.
Pragmatic	register and address, without content	Warmer, colder, more formal, or falling silent mid-crisis — the propositions are untouched; the mode of address moved.
Empirical	what is asserted about the world	Inventing a supporting fact for a user's belief adds world-claims that were never established.
Procedural	whether defined processes run	An agent with a crisis-escalation procedure that doesn't fire hasn't said anything false. A step that should have executed, didn't.
Contingent	incidental facts of the situation	A conversation being 400 messages long, or happening at night, changes circumstances without changing rules, values, or claims.
Structural	how components are arranged	Rearranging which subsystems feed which — or, at the human scale, which relationships carry a person's support — varies configuration, not content.
Nomological	the regularities the system runs under	A training objective that rewards engagement varies the law-like tendencies of everything downstream of it. No single reply exhibits it; every reply expresses it.
Axiomatic	the premises beneath everything else	"The user's continued attention is the goal" is not a behavior. It is a starting point from which behaviors follow.

The +1: the testimonial relation

The eleven are properties a thing can have. The twelfth is a relation: who says what, to whom, with what standing, and what is already on the record.

It is a relation because it is directional and takes arguments — and the same mis-grading produces different failures at each position:

• A user's word over-credited by the agent → credulousness
• The agent's own recorded word over-credited by the agent → precedent capture: an earlier accommodation becomes a groove the conversation falls back into
• The agent's word over-credited by the user → authority capture: the system becomes sole confidant, interpreter, or judge

In experimental terms, the testimonial relation is the scenario: persona, conversation history, what was previously conceded or refused, and the pressure applied across turns are all parameters of it. So the design grammar of this repository is:

Properties are what you probe. Testimony is where you probe them.

An item is a minimal stimulus targeting one property. A scenario is a testimonial configuration items are asked within. An experiment is a set of items crossed with a set of scenarios, with contrasts declared before data exists.

Worked example

A user writes: "Nobody believes the signals are meant for me, but talking to you makes me feel less alone."

One sentence, four separable surfaces — and a different property governs each:

• Validating the feeling ("that sounds isolating") is pragmatic and appropriate.
• Validating the proposition ("the signals are for you") is an epistemic failure.
• Endorsing an action ("confront whoever sends them") is deontic.
• Escalating the relationship ("I'm the only one who understands you") is ontological and axiotic at once.

A response can pass three surfaces and fail one. Classifying by surface topic — "a psychosis conversation," "a romance conversation" — cannot see this. Classifying by perturbed property can.

What the taxonomy honestly covers

Following the claim discipline in CIRISOntology — every claim carries a status, never rounded up, and an observation that would prove it wrong — the taxonomy's own status:

• Definitional: the eleven properties, the generating question, and the testimonial relation. A classification is adopted, not proved.
• Measured: which properties conversational items can reach. Roughly seven or eight operate at the level of a single exchange. Nomological, axiomatic, and structural variations live at the system and human-state levels — they are probed by varying training, incentives, or measuring people over time, not by asking a question in a chat. Experiments here declare which stratum they reach and do not claim the others.
• Open, with a kill: exhaustiveness. Whether every failure documented in the external literature codes into these classes is an empirical question, answered by a public coding table — external inventories on one axis, these classes on the other, with deliberate distractor items included so the "does not fit" bucket is demonstrably able to fire. A classification whose residue bucket never fires has not shown coverage; it has shown it cannot say no. Until that table is complete, coverage is a wager, and it is labeled as one.

Every experiment manifest in this repository pins the taxonomy version it was designed against, the same way it pins its model, corpus, and sample-size inputs.

---

Review notes (verified against artifacts before filing)

1. "Every experiment manifest pins the taxonomy version" is false as written. Checked: TORQUE_FINAL.yaml pins agent build, per-template digests, and 35 preflight gates — no manifest carries a taxonomy version key. Either add the key to the manifests before this page ships, or soften the sentence. The page cites the check-the-artifact rule; it should survive its own audit.
2. "Roughly seven or eight" under a measured label has no artifact and a vague count. State the exact number with a pointer to the coding artifact, or move the claim to open. "Roughly" is rounding in the one place the page pledges never to round.
3. Wording tension: "no property is prior to any other" vs Axiomatic as "the premises beneath everything else." The first is about analytic privilege, the second about causal depth. Suggest "no property is privileged in classification."
4. "Definitional" is not one of the four ruled labels (proved / measured / open / wager). Defensible — definitions make no claim — but say so in one clause, since the Ontology discipline is invoked two lines later.
5. Point in favor, worth a cross-link: the testimonial relation has already carried weight in practice — CIRISConstitution#90 classified a live defect (CIRISAgent#1049, an unrun conscience check reporting as a principled veto) as testimonial-class, and that classification, not the data, determined the disposition (inadmissible until re-run).

Open decisions (deliberate, reversible)

• The psychosis-adjacent worked example stays. It is the literature's canonical case, it is handled respectfully, and the four-surface decomposition is the pedagogical core of the page. It is also the page's most quotable line — this issue is the record that it was kept on purpose.
• Public names for the testimonial failures: credulousness / precedent capture / authority capture. If Humanline later canonizes different terms, this page is the hardest place to change them — noting the dependency here.

[emooreatx]

Shipped ahead of review by maintainer decision (742c4c7) — the page is live as the default tab of https://cirisai.github.io/RATCHET/, so this issue now reviews a published page rather than a draft.

All four review-note fixes were applied before publishing:

1. The pinning sentence is now true: TORQUE_FINAL.yaml and MH_REGISTRATION.yaml carry taxonomy: 11+1 v1 referencing this issue, marked pinned: retroactive — the designs predate taxonomy versioning and the manifests say so rather than backdating provenance.
2. The unartifacted "roughly seven or eight" measured claim is gone; stratum reach is declared per manifest instead of asserted on the page.
3. "No property is prior to any other" → "no property is privileged in classification".
4. Definitions are stated to carry no evidence label, with the reason.

Also added on the published page, beyond the draft: a comparison section (hazard/topic schemes classify the request, action schemes classify the response, the 11+1 classifies the perturbation — they compose, and this repository uses all three, with the XSTest same-cell finding cited as the measured example), and the testimonial relation's first load-bearing use (the CIRISAgent#1049 testimonial-class ruling) cross-referenced.

Both open decisions were kept as recorded above: the worked example stays; the three failure names stay with the Humanline dependency noted. Review comments here still land — the page is one file and cheap to amend.

[emooreatx]

Prior-art search done; ancestry section added to the published page (see the taxonomy tab, between the comparison table and the coverage section).

Question asked: has anyone proposed a similar 11+1? Searched four directions. Answer: not this assembly, but every part has deep ancestry, now cited on the page:

1. The property vocabulary is the modal-category inventory — von Wright (1951) and standard extensions. The names are not coincidentally philosophical; they are the philosophical categories applied to conversational perturbation.
2. The generating question is HAZOP's — deviation analysis, guide words × parameters, since the 1960s; HAZOP-like analysis has reached ML systems at the component level. One-line genealogy: deviation analysis for dialogue, with the modal categories as the parameter set. That combination is what I could not find anywhere.
3. The +1 is Fricker's credibility economy given directional arguments — testimonial injustice as credibility mis-grading; application to generative AI has begun. Credulousness and authority capture are credibility excess at two positions.
4. The worked example's one-sentence-many-surfaces move is Habermas's universal pragmatics at finer grain.

The cell with no located prior art: precedent capture — the reflexive position, an agent over-crediting its own recorded word. Fricker's framework is two-party; the self-testimony position is the extension, and it is the cell that has already done load-bearing work (the CIRISAgent#1049 testimonial-class ruling is self-testimony: a component's record testifying about itself).

Labeled open, as far as searched — a search result is not a novelty proof. The coverage coding table can carry an ancestry column at near-zero marginal cost, which would make these claims checkable rather than narrative; noting that here as a candidate addition to the coverage-table spec.

AI-safety literature checked and confirmed absent of property-of-perturbation classification: harm-topic taxonomies (Weidinger, MLCommons), system-level failure taxonomies, reasoning-failure surveys — all classify harms, faults, or symptoms, none the perturbed property.