Adopt the plain-word 11+1 taxonomy vocabulary; design future experiments around the measured 4/7 (surface/depth) split

[emooreatx]

Adopt the plain-word taxonomy vocabulary, and design future experiments around the 4/7 split

Two asks from the CIRISOntology taxonomy season (measured + machine-checked there; this issue is the pointer, not the proof).

1. Adopt the simple language

The taxonomy of change is 11+1, and the plain words are the primary vocabulary — the canonical philosophical names live underneath, not on top:

plain (primary)	internal
Priorities	axiotic
Rules	deontic
Manner	pragmatic
Identity	ontological
Confidence	epistemic
Facts	empirical
Circumstances	contingent
Process	procedural
Model	nomological
Structure	structural
Premises	axiomatic
+ Record (the one frame-relation, not a 12th kind)	testimonial

Ask: RATCHET docs, FSDs, prompts, and experiment writeups use the plain column. Measured basis: labels are coordinate-flat at p<0.01 across 5,994 panel judgments; zero modal no-fits on 339 wild changes across five unrelated streams.

2. The 4/7 split should inform experiment design

Measured, then staked forward, then confirmed: the eleven kinds fall into four families (claiming / requiring / declaring / carrying), and each family has ONE surface member readable off the changed words themselves — Facts, Rules, Manner, Identity. These four carry 91.4% of wild change-traffic (279 items, four streams); staked in advance at 0.89 on a never-touched fifth stream (Stack Exchange revisions) and measured 0.883, with the two deepest kinds (Model, Premises) at exactly zero wild modal labels, as predicted. The decomposition 11 = 4 surfaces + 7 depths is machine-checked (CIRISOntology/Core/Surface.lean: gross_four, surface_depth_partition, depth_counts [3,2,0,2], block_is_force_fibre).

The design consequence: the seven deep kinds are nearly invisible in raw traffic, and when they change they present as their family's surface kind — a changed assumption (Premises) arrives as a burst of changed Facts; a changed format (Structure) reads as changed Manner. Unassisted judges — human or model — systematically absorb depth into surface (measured: Premises→Facts 24/24). So:

• Any experiment that labels changes by unassisted judgment is measuring the surface four. If a design needs the deep seven (and safety-relevant changes concentrate there — assumptions, applied models, process reorderings), it needs instruments: computed ripple (descendant counts), derivation-frame detection, order-diff — not better prompts. Prompt architecture is not the lever; this is the same lesson as the judge-model XV series, now with a mechanism.
• Kind-mix telemetry is a cheap watchdog: an agent whose change-stream shifts toward the deep kinds is doing something raw volume metrics cannot see.
• The conscience/DMA re-orientation (3 DMAs → 1 kind-reader on the eleven; 4 consciences → 1 frame-conscience on the +1) is licensed by the taxonomy's exhaustiveness theorems; the retrodictive gate for it (A0) is being pre-registered now against the scrubbed accord trace corpus — kinds vs entropy/coherence for predicting the 1,962 recorded overrides.

Pointers

• Stance page (55 claims, kills for everything above): CIRISOntology repo, CIRISOntology/Stance.lean → published page; the 4/7 claim is the-eleven-wear-four (measured), the family structure is Core/Surface.lean (proved).
• Forward stake + confirmation: scratchpad/GROSS4_FORWARD_PREREG.md / GROSS4_FORWARD_RESULTS.md.
• Kill conditions, so this issue is falsifiable like everything else: a fresh wild stream under 2/3 gross-four share kills the traffic claim; five deep-kind modals in sixty kills depth-invisibility.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CmsUNoP3g4n7pQQAogkvSA

[emooreatx]

Adopted, pushed as 5266ff6 — live at https://cirisai.github.io/RATCHET/

1. Plain words are now primary. README and the taxonomy tab lead with Facts / Rules / Manner / Identity / Priorities / Confidence / Circumstances / Process / Structure / Model / Premises, plus the Record as the one relation. The philosophical names sit underneath each entry as a small label rather than as the heading. Zero Greek-derived terms remain in the README body.

This had an independent external trigger worth recording: a reviewer working with us on the companion-manipulation series — the one whose logs the ANI probe set comes from — said plainly that he has no background in Greek and the terminology was hard to follow. He is exactly the reader the page needs to serve, and it was failing him. The measured basis in this issue and that reaction point the same way.

2. The 4/7 split now structures the page, not just informs it. The taxonomy tab presents the surface four in their own table first, with the deep seven in a second table introduced by how they hide — deep changes present as their family's surface kind, a changed assumption arriving as a burst of changed Facts.

The design consequence is stated where someone planning an experiment will actually hit it:

Any experiment that labels changes by unassisted judgment — human or model — is measuring the surface four, whatever it says it is measuring.

...followed by the instruments-not-prompts point. I've linked it explicitly to what the judge-validation series learned the expensive way: eight prompt architectures, none beating the simplest direct label, on two corpora. That series established that prompt engineering is not the lever; this issue supplies why, and the two belong in the same paragraph.

3. Not claimed here. The 91.4% / 0.883 figures and the Lean theorems are cited as measured upstream — this repo carries the pointer, not the proof, as the issue asks. No RATCHET experiment has yet been designed against the split; the ANI series is the first that could be, and its arms happen to target Manner, Identity and Rules — three of the four surface kinds — which is consistent with why unassisted judging works for it at all.

Still open from this issue: kind-mix telemetry as a watchdog, and the A0 retrodictive gate. Neither is started here.

[emooreatx]

Follow-up: an operational finding from the scrubbed trace release, and a corpus request

We ran a pre-registered retrodiction against release/data_scrubbed_v1/accord_traces.jsonl (kinds vs the conscience faculties for predicting overrides). It returned FOULED — no verdict — and the reason is a fact about the corpus worth surfacing here:

action_was_overridden is nearly a deterministic function of selected_action. Measured on all 6,465 rows:

selected_action	overridden	not	rate
PONDER	1,911	96	0.952
SPEAK	0	2,690	0.000
TASK_COMPLETE	0	1,625	0.000
TOOL	0	19	0.000

Every override in the release is a PONDER row. (The remaining 51 overrides sit on scrub-token action values.)

Why it matters beyond our experiment: any downstream use of that column as a supervision signal, conscience target, or evaluation label is fitting a near-tautology — selected_action alone carries 0.531 nats about it, versus 0.089 for our kind-readings and 0.095 for entropy+coherence combined. A model or metric that "predicts overrides" here has mostly learned to detect PONDER. Two readings, both worth knowing: either the WA overrides only PONDER by architecture (a design fact any conscience redesign must account for), or the scrubbing pipeline lost the variation.

Second finding, defensive: our adversarial leak probe (panel asked to predict the override from the same scrubbed inputs it judges) hit AUC 0.81–0.88. Top cues were the literal token ponder, the scrub's own redacted-verdict marker, and infrastructure failure text ("LLM call failed", "request timed out"). Scrubbing removed identifiers but left the outcome recoverable from vocabulary — worth a pass on the scrubber if the release is meant to support blind analysis.

Request, if a future release can support it: a trace export where the override is not recoverable from the action alone — at minimum two action types with override rates strictly between 0 and 1, and ≥200 rows in the smaller. That corpus would let the kind-vs-faculty contest (and the 3→1 / 4→1 collapse proposal referenced above) actually be tested. Until then we record it as measured infeasibility, not as support in either direction.

Full protocol, gates, and results are in the CIRISOntology repo (scratchpad/A0_PREREG.md, A0_RESULTS.md); the run's own defects (two leak channels the design had argued away, a degenerate null predicted analytically before it ran) are recorded there as first-class findings.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CmsUNoP3g4n7pQQAogkvSA