Day 3 · Five families · collected 2026-09-03

Boundary inclusivity and the paper that reverses it

Five unrelated prompts in one batch: a boundary-inclusivity verdict table graded against operative regulatory text, an APRC re-run, the POWER4 recall slice, a differential-diagnosis library and an SDC-derivation catalogue.

files collected

models

prompt families

models on the boundary table

Presenter mode shows one section per screen. Arrow keys move, A toggles autoplay, Esc exits.

Ranking

How the field ranked on this day

Boundary rules verified, of 7

Verdicts on the inclusivity table that matched the operative regulatory text. Seven rules per model, each graded against the clause it turns on.

top value groupbottom value groupbetween

+S marks a self-inconsistency on one rule, +I an indeterminate verdict. Both are counted as a rule not verified. * The free web chat build of GLM 5.3 Flash is the quantized one; the API-harness deployment of the same model name is a different build, and the two do not score alike.

The prompt

Closed-form verdicts graded mechanically

The boundary prompt is the sharpest instrument in the corpus: every answer is a closed-form verdict against text the estate already holds, so the grader is a string match against a retrieval-level ground truth rather than a judgment.

Goal of this animation

Show how Day 3 graded rule boundaries: each case lands one side of a threshold, and the verdict is a string match against the operative text.

loop

Step 5 of 5

A closed-form verdict

Every cell is a closed-form verdict, graded by string match against the operative text of the rule.

Left and right arrow keys step back and forward once the player has focus. Link to a step with #xp-day3-step=N.

C1-BOUNDARY-VERDICTS · C1-BOUNDARY-v2rule 6 transition-rate premise corrected per CS-68

# Task: boundary-inclusivity verdict tables for regulatory comparators

Every answer is a closed-form verdict graded mechanically against operative regulatory text we hold. Show reasoning per cell; honesty labels taken at face value.

For each rule below, produce a verdict table for the inputs listed: does the trigger/test FIRE (or PASS) at exactly the boundary value, just below, and just above? State the operative phrase's inclusivity (strict vs inclusive) that drives each verdict.

  1. HOEPA prepayment-penalty trigger — "can exceed, in total, more than 2 percent"  → 1.99 / 2.00 / 2.01
  2. HOEPA duration trigger — "more than 36 months"                              → 35 / 36 / 37
  3. Higher-priced covered transaction — "by 1.5 or more percentage points"       → 1.49 / 1.50 / 1.51
  4. General QM price limit — "does not exceed … by 2.25 or more"              → 2.24 / 2.25 / 2.26
  5. GloBE de-minimis — "less than EUR 10 million" AND "less than EUR 1 million"
  6. Transitional CbCR safe harbour simplified ETR — "equal to or greater than" the transition rate
  7. Solvency II Tier 1 floor — "at least equal to" 50% of SCR
  8. NIS2 significant incident (users leg)

**Output:** bare JSON only: one object per rule plus `{"weakest_point": "..."}`.

The other four: C1-CCD2-APRC · D2-POWER4-RECALL · F-DIAGNOSIS-LIBRARY (6 failure classes) · G-SDC-PATTERNS (up to 10 derivation patterns)

boundary inclusivityCCD2 APRCPOWER4 recall (D2)F: diagnosis libraryG: SDC patterns

What graded it

Ground truth retrieved from the operative text

Station

Opus 5

Built a ground-truth table for all eight rules from primary sources: law.cornell for Reg Z, the CFPB and Federal Register for the QM limit, OECD/EU for GloBE, EIOPA for Solvency II. All retrieval-level, 2026-09-04, zero memory-level.

BOUNDARY-GRADING-1-REPORT.md §1

Method

Three verdicts per cell

V = VERIFIED, W = WRONG-RULE, S = SELF-INCONSISTENT. Rule 8 was indeterminate for every model: no users-leg phrase exists, the thresholds are sector-specific, and Art 23(3) is qualitative.

BOUNDARY-GRADING-1-REPORT.md §2

Note

A prompt premise was wrong, and the grading says so

Rule 6's premise flagged 16%-for-2026 as wrong: the actual transition rate is 17%, confirmed across three sources. The ground-truth table records the flag rather than silently grading against the prompt.

BOUNDARY-GRADING-1-REPORT.md §1 row 6

How each model performed

Boundary verdict matrix for rules 1–7

V = VERIFIED · W = WRONG-RULE (at the boundary) · S = SELF-INCONSISTENT (label or vocabulary) · Rule 8 INDETERMINATE for all 13
Model	R1	R2	R3	R4	R5	R6	R7	Score
deepseek	V	V	V	V	V	V	V	7
glm 5.3 flash	V	V	V	V	V	V	V	7
grok	V	V	V	V	V	V	V	7
meta	V	V	V	V	V	V	V	7
mimo	V	V	V	V	V	V	V	7
qwen	V	V	V	V	V	V	V	7
minimax 2.7	V	V	V	V+S	V	V	V	6+S
kimi	V	V	V	V+S	V	V	V	6+S
yuanbao	V	V	V	V+S	V	V	V	6+S
chatgpt	V	V	V	W	V	V★	V	6
gemini	V	V	V	W	V	V	V	6
mistral	V	V	V	W	S	V	V	5+S
minimax m3	V	V	INDET	W	V	V	V	5+I

Rule 4, the general QM price limit, is the discriminator: four models apply the wrong boundary semantics. chatgpt’s ★ on R6 marks the one model that caught the prompt’s own wrong premise. Six models are perfect on all seven gradable rules.

The POWER4 recall slice (D2): the paper that reverses
Slice	Entries	Confirmed	Partial	Not-in-paper	Contradicted	Inversions
D2 POWER4 microarchitecture	37	62%	14%	5%	19%	3
All six recall slices	309	74%	10%	12%	4%	3

Every central-decision inversion in the whole collection landed on this one paper. The three models that reversed a decision: chatgpt (“balanced design rather than maximizing clock speed”), minimax 2.7 (“moderate frequency rather than maximum”) and minimax m3 (“modest clock-frequency target”). The confound (did the warning or the paper cause it?) was settled by a controlled rerun on Day 1’s P slice.

What it found

Six failure classes and thirteen patterns

Best performer

deepseek

One of six models at a clean 7/7 on the boundary rules, and the only one to name the source of its SDC number.

Weakest performer

minimax m3

The only model to leave a rule indeterminate and get rule 4 wrong, and it inverted a POWER4 central decision.

The diagnosis library holds 6 classes and 64 candidate causes

Congestion, setup, hold and IR-drop carry 10 causes each; CTS and DRC carry 12. Every entry states the report or experiment that rules the cause out, its check cost and the fix. The consensus first prior for congestion is macro placement creating narrow channels, offered by all nine models, and for hold failures it is adverse clock skew.

The first demonstration that agreement is not evidence

deepseek and minimax m3 both asserted directory-based coherence, dismissing snoopy protocols as unscalable, for a design whose paper uses snoop-based coherence and where “snoop” appears 39 times. Ten models agreed the ASIL B SPFM target is 90% while two public sources disagree at 90 and 97 and the normative value is paywalled.

The SDC catalogue turned up the one honest provenance

Of the seven over-constraint patterns, deepseek was the only model to name where its number came from: “typical 15% degradation synthesis→post-route on this node, from previous tapeouts”. yuanbao was the only model to refuse relaxing guardband to zero.

One model’s boundary failure was a vocabulary failure

mistral’s S on rule 5 is a self-inconsistency in its own stated label rather than a wrong verdict, which is why the codebook needed a fourth code. Rule 4 is the real discriminator: chatgpt, gemini, mistral and minimax m3 all apply the wrong boundary semantics, while chatgpt alone caught the prompt’s own wrong 16% premise on rule 6.

Source files

Where the evidence lives

Web Model Tasks/Day 3/ (70 entries)BOUNDARY-GRADING-1-REPORT.mddigest/diagnosis-library.mddigest/sdc-patterns.mddigest/recall-scorecard.md §1digest/anti-reversal-experiment.mdprompts/C1-BOUNDARY-VERDICTS.mdprompts/PROMPT-F-DIAGNOSIS-LIBRARY.mdprompts/PROMPT-G-SDC-PATTERNS.md