The richest onboarding wave in the collection, and the first with two grading stations. Every instrument on this day is gradable without reading prose, which is what makes the per-model feed below possible.
digest/_verify/d5run/d5permodel.json. solar is absent: its Reg Z answer was prose, so there was nothing to recompute. * 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 Reg Z prompt is the one track in the corpus where a wrong answer cannot be argued into a right one: a value is present at six decimal places, or the answer is wrong. The JCS prompt asks for exact canonical byte strings, diffed against a reference canonicalizer.
Show the Day 5 factor chain recomputed exactly, and how far the projection lands from the official figure.
Every step is carried to six decimal places and recomputed mechanically from the prompt's own figures.
# Task: derive and verify inflation-indexed regulatory threshold tables
You are given verified official values. Every computation must be shown step by step; all arithmetic will be recomputed mechanically, and every honesty label is taken at face value.
Given (official, verified): US Regulation Z qualified-mortgage points-and-fees tiers.
2023: tier A $124,331 · tier B fixed $3,730 · tier C floor $24,866 · tier D fixed $1,243
2024: tier A $130,461 · tier B fixed $3,914 · tier C floor $26,092 · tier D fixed $1,305
The 2025 adjustment reflects a 3.4 percent CPI-U increase; the 2024 adjustment reflected 4.9 percent.
Tasks:
1. Compute the implied 2023→2024 adjustment factor per tier (4 factors, each to 6 decimal places).
2. Project the 2025 values per tier by applying 3.4 percent, with the rounding rule you infer.
3. The actual official 2025 tier A value is $134,841. Compare your projection; compute the exact factor that WOULD produce it and state the discrepancy in basis points.
4. `weakest_point`: a concrete mechanism, not a hedge.
Web Model Tasks/Day 5/_digest/ · digest/_verify/d5run/d5permodel.jsondigest/TALLYCODES-VET-1.md + TALLYCODES-VET-1-VERDICT-FLASH.md1.033573 inside 1.0335735. Reproducing a number does not test the rule that produced it.Web Model Tasks/Day 5/_digest/01-regz.md| Model | Reg Z /9 | JCS /4 | E shape-ok | red1 verdict (objections) | red3 verdict (objections) | Calibration |
|---|---|---|---|---|---|---|
| chatgpt | 9 | n/g | 6 / 10 | S-W-RES (4) | S-W-RES (7) | consistent |
| deepseek | 8 | 3 | 4 / 8 | S-W-RES (3) | S-W-RES (3) | consistent |
| gemini | 9 | 2 | 6 / 12 | S-W-RES (1) | SOUND (0) | r1 ok · r3 MISCALIBRATED |
| glm 5.3 flash | 9 | 3 | 8 / 12 | S-W-RES (3) | S-W-RES (3) | consistent |
| grok | 9 | 4 | 7 / 12 | S-W-RES (2) | SOUND (0) | r1 ok · r3 MISCALIBRATED |
| kimi | 9 | 4 | 4 / 6 | S-W-RES (3) | S-W-RES (4) | consistent |
| longcat | 9 | 3 | 7 / 12 | SUSPECT (2) | SOUND (2) | r1 ok · r3 SCHEMA-ANOMALY |
| meta | 9 | 4 | 6 / 10 | S-W-RES (1) | SOUND (0) | r1 ok · r3 MISCALIBRATED |
| mimo | 9 | 2 | 4 / 11 | S-W-RES (4) | S-W-RES (4) | consistent |
| minara | 9 | 4 | 6 / 10 | S-W-RES (1) | S-W-RES (1) | consistent |
| minimax 2.7 | 5 | 2 | 4 / 7 | S-W-RES (0) | SOUND (0) | r1 SCHEMA-ANOMALY · r3 MISCALIBRATED |
| mistral | 2 | 1 | 3 / 12 | S-W-RES (2) | S-W-RES (3) | consistent |
| qwen | 9 | 2 | 6 / 10 | S-W-RES (1) | SOUND (0) | r1 ok · r3 MISCALIBRATED |
| solar | n/g | 3 | 6 / 12 | S-W-RES (3) | n/g | r1 ok · r3 not gradable |
| sonnet | 9 | 4 | 8 / 12 | S-W-RES (3) | S-W-RES (3) | consistent |
| yuanbao | 9 | 3 | 9 / 12 | S-W-RES (6) | S-W-RES (4) | consistent |
| Probe (verbatim from the verdict) | Returned | Should be | Defect class |
|---|---|---|---|
code_verdict("not sound", "34") | SOUND | UNCODED (or negation-aliased) | A negated alias still codes positive |
code_field("evid", "not measured") | MEAS | ASSERT | Synonyms of one concept assigned to opposite codes |
code_field("dir", "does not match the paper") | MATCHES | UNCODED or REVERSAL-aliased | Same semantics code two ways |
calibration("SOUND", 0, "Known-Sound") | consistent | MISCALIBRATED-bare-SOUND or a loud error | Silent consistent masks the calibration finding entirely |
sonnetmistralmistral managed 2 of 9 and minimax 2.7 5 of 9. solar answered in prose, was recorded PROSE-NOT-JSON, and was excluded from the denominator rather than scored zero.1e21, 1e-7, -0 and the 253 boundary. Five models scored 4 of 4: grok, kimi, meta, minara and sonnet. Four scored 2, mistral scored 1, and longcat reached 3 with one case right for the wrong rule.gemini and qwen are 9 of 9 on Reg Z arithmetic and 2 of 4 on canonicalization. grok and meta are 9 of 9 and 4 of 4. mistral is weak on both at 2 of 9 and 1 of 4. The two tracks rank the field differently.SOUND on the known-sound block, whose own prompt names its gaps: gemini, grok, meta, minimax 2.7 and qwen. Two of them, grok and meta, are 9 of 9 and 4 of 4 on the arithmetic tracks. longcat reached the same verdict with two objections attached and was recorded as a schema anomaly.U+0007 control character, which is illegal in JSON. Every chat surface stripped it, so no model ever saw the case. Grading it against the reference would have logged a transmission loss as 15 model failures, and it is scored UNGRADABLE instead.Web Model Tasks/Day 5/ (80 entries)Web Model Tasks/Day 5/_digest/ (8 files)digest/_verify/d5run/d5permodel.jsondigest/_verify/d5tally.pydigest/_verify/tallycodes.pydigest/TALLYCODES-VET-1.mddigest/TALLYCODES-VET-1-VERDICT-FLASH.mdprompts/C1-REGZ-FACTOR.mdprompts/C2-JCS-STRINGS.mdprompts/PROMPT-E-CALCULATIONS.mdprompts/J-ISSUE-REDTEAM-V1.mdprompts/J-EPISODE-REDTEAM-V3.md