Joakimpalm-Zen/Xyntetik-Typed-Decisions-record

Xyntetik typed decisions: TD-000 to TD-002 record

The record of the typed-decisions track's first two studies: TD-000, a measurement stage of three models' /v1/decide probabilities, and TD-001, a preregistered test of a log-loss LoRA fine-tune of Xyntetik-Kvist-14B against a free per-class temperature map. TD-001 is INCONCLUSIVE by its registered rule. A dated update (2026-10-15) adds TD-002, the same adapter on a fresh held-out sample: also INCONCLUSIVE, from its analysis file; its report is not yet written.

Research evidence dataset. No model weights. Part of the collection Xyntetik Kvist: distilled agents on this account. The model studied is Xyntetik-Kvist-14B. This dataset holds the track's two preregistrations with their amendments, the two reports, the question files, every scored row, the analyses, the harness and its tests: TD-000 (run 2026-10-09/10) and TD-001 (run 2026-10-13), with TD-001's descriptive addendum (2026-10-13/14). The dated update of 2026-10-15 adds TD-002's preregistration and analysis file (run 2026-10-14/15).

Dataset summary

Question tested. A model is given a state and a fixed set of answer strings, and Xyntetik Runner's POST /v1/decide returns the exact probability of each answer: the option strings tokenised in context and renormalised over the options. Are those probabilities usable as decisions? Concretely: do they stay put when the options are reordered or the question is rephrased; are they calibrated on held-out questions; how much of the miscalibration a free post-hoc map (one inverse temperature per class) removes; and can a training objective do better than that map?

• TD-000 measured. It tested nothing; by rules fixed before its data, it produced the inputs TD-001 was designed from (per-class targets, margins and sample sizes).
• TD-001 asked one question: does a LoRA fine-tune of Kvist-14B, trained with the exact decide log loss, reach a held-out log loss on the ledger class at least 0.05 nats below the base model's per-class temperature map?

The subjects are pretrained language models and the data carries human judgement. This track is not part of the machine-only (GENESIS) work, and nothing here is read as a result about it.

Models involved.

• Studied and fine-tuned: Xyntetik-Kvist-14B. Scoring used the Q8_0 GGUF (Joakimpalm-Zen/Xyntetik-Kvist-14B-GGUF at 2fd41906); TD-001's frozen training base was the BF16 release (Joakimpalm-Zen/Xyntetik-Kvist-14B at f7434073). TD-001's adapter is not distributed: it was never committed to the record.
• Measured beside it in TD-000: Muse-Glimmer-30B (Runner's Q4_K of the BF16 parent, Kvist's parent) and Qwen3.5-4B (unsloth/Qwen3.5-4B-MTP-GGUF at 86835bf9, Q4_K_M).
• Comparators in TD-001: the base model raw (B) and the base model after the per-class map fitted on the training side (B+map, the arm the primary compares against). Both are TD-000's Kvist rows, checked for reproduction on a tenth before training.
• Inside the data: the proposed edits in the ledger questions are Hermes-4-14B samples from an earlier study or, in some of slice 1's, the upstream fix.
• The instrument: Xyntetik Runner v1.1.4 (release binary sha256 89d2e07e…) and its scripts/decide-calibrate.py (sha256 c427d28f…), both unchanged and pinned.

Method. Each study followed the same order: a preregistration with its kill or outcome rule written before its data, a harness with unit tests, instrument checks, the runs, an analysis computed by a pinned script, and a report in the order manipulated, observed, survived, failed. Amendments are recorded in each preregistration with their timing against the data.

• The questions. Slice 1 has 1,303 questions: ledger 464 (does a proposed code edit make the failing tests pass; labels from running the library's own tests), synthetic 809 (Python program states, answered by execution) and a 30-question placeholder that is read as nothing. Slice 2b has 8,044: ledger 544 (236 items at their real base rate, and 36 items whose target is a measured pass rate), NLI 3,900 (WANLI) and intent classification 3,600 (banking77 and MASSIVE). Every item appears in two or three variants: option orders or question phrasings.
• Splits. Held out by split group, never by row. One joint split over both files was fixed in TD-000 before any data: 583 of 2,917 ids held out.
• The readout. Rendering raw-v1, answer prefix "Answer: ", context 16,384, one slot, on one MIG 1g.24gb slice of an NVIDIA RTX PRO 6000 GPU.
• TD-001's treatment. One LoRA (rank 16, α 32) on the seven projections the runner serves, in all 52 blocks, trained on the joint split's training side with the class-balanced decide log loss for 2 epochs, one seed, no checkpoint selection, served through runner v1.1.4 --lora.
• TD-001's rule. Δ = LL(F, raw) − LL(B, mapped) on the pooled ledger held-out questions (32 units, 174 questions), with a paired bootstrap over units (10,000 draws, seed 20261015). PASS if the upper bound U < −0.05 and the guardrails hold; KILL if Δ̂ ≥ 0 or the lower bound L > −0.05; INCONCLUSIVE otherwise, and on any failed instrument check. The smallest attainable one-sided p is 1/10,001.

What each file contains.

• README.md: this card.
• PUBLIC-COPY-CHANGES.md: every change this public copy makes to the record: three excluded share-alike files, five paraphrases of the owner's words, and 53 machine, session and commit-identity mentions replaced by role names or placeholders.
• THIRD-PARTY-NOTICES.md: the attributions, licences and notices of the third-party material (section "Third-party data and licences" below).
• typed-decisions/: the shade branch typed-decisions at f34e8d672, research/evaluation/typed-decisions/, with the changes above (238 files):
  • README.md (the question shape and the slices, with the slice 2 sections kept as written) and TRACK.md (the track's question, label and standing rules);
  • prereg/TD-000.md and prereg/TD-001.md with their amendments; report/TD-000.md and report/TD-001.md, the latter with the descriptive addendum at its end;
  • data/: questions-slice1.jsonl, questions-slice2b.jsonl with its manifest (the drawn WANLI ids), and sources/ (the WANLI test split, the banking77 and MASSIVE rows, and SOURCES.md);
  • evidence/td000/: one directory per run (I1-M, I1-Q, K1, K2, DK, M2, DM, Q2, DQ): every scored row with its probabilities and log-probabilities (rows.jsonl), every request's time, tokens and state digest (requests.jsonl), the instrument's aggregate, the server and record logs, and the chain log;
  • evidence/td000-analysis/: the registered readouts per run, the determinism checks, and TD-001's inputs and references;
  • evidence/td001/: per phase (I-B, I-Z, T3, T4, F, I-A, S-held, S-dev) the rows, logs, training history and timings; the plan; analysis.json (the outcome); descriptive-addendum.json; the gate and chain logs;
  • evidence/td001-build/: the harness build's tests, CPU smoke, plan and the BF16 fetch log;
  • the two baseline rows of 2026-09-22 (evidence/muse30b-slice1-decide-calibrate.*, evidence/qwen35-4b-slice1-decide-calibrate.*) and the contamination audits of slices 2 and 2b;
  • scripts/: the question builders, the contamination audit, TD-000's chain, recorder and analysis, and TD-001's harness (scripts/td001/, with HARNESS.sha256 and three test files).
• Not included: slice 2 (data/questions-slice2.jsonl) and its SNLI and MultiNLI source rows, because they are share-alike. Neither study reads them. TD-001's adapter and the models' weights are not here.

Reproduction. Every number in both reports is computed from the committed rows on CPU with Python 3 and numpy. The pinned instrument is the one file needed from outside this dataset:

curl -L -o decide-calibrate.py \
  https://raw.githubusercontent.com/Joakimpalm-Zen/xyntetik-runner/78a469876cc550c5097b30fa3df1367dfd913bd9/scripts/decide-calibrate.py
sha256sum decide-calibrate.py   # c427d28f461363de3acb27fccea889490ea7f5ab3ac41bd8531b63593fbe7053
TD=typed-decisions
# Dated update, 2026-10-15: TD-002, INCONCLUSIVE

TD-002's report is not yet written. The numbers below come from its committed analysis file,
`typed-decisions/evidence/td002/analysis.json`, which the preregistered analysis wrote at the end of the run
(2026-10-14/15). This update adds that file and the preregistration, `typed-decisions/prereg/TD-002.md`.

**The question.** Does TD-001's adapter, frozen, beat the base model's per-class temperature map (B+map) on a fresh
held-out ledger sample by the same 0.05-nat margin? The fresh questions came from more commits mined from permissive
libraries, with the proposed edits regenerated by Hermes-4-14B at Q8_0: 96 units and 728 rows. The primary Δ weights the
strata by the instrument's own ledger composition (`pair` 0.460, `cond` 0.468, `rate` 0.071), fixed before the data,
with a paired bootstrap over units (10,000 draws). The outcome rules are TD-001's, unchanged.

**The unflattering readings first** (secondaries, descriptive):
- on the `pair` rows the adapter was worse than B+map: +0.012 nats [+0.002, +0.021];
- the gain came from the `cond` rows, −0.062 [−0.092, −0.031], and from the 12 rate units, −0.632 [−0.744, −0.481].
  The rate units carry 0.071 of the weight but −0.045 of the primary's −0.069;
- unstratified, the difference was −0.042 [−0.060, −0.023].

**TD-002: INCONCLUSIVE.** Δ̂ = **−0.0686 nats**, 95% paired bootstrap **[−0.0863, −0.0493]**. PASS needed the upper bound
below −0.05; it fell short by 0.0007. KILL needed Δ̂ ≥ 0 or the lower bound above −0.05; neither held. Every instrument
check, data check and guardrail passed. The run used 5.9 h of slice time against its 16 h cap. Like TD-001's, the
interval excludes "no difference" but does not decide the margin.

**Not in this update:** TD-002's harness, its fresh question file (`data/questions-slice3-ledger.jsonl`) and its run
evidence. The question file reproduces upstream commits from click, glom, more-itertools, packaging, rich and one library
this record does not yet cover, parse (MIT). `THIRD-PARTY-NOTICES.md` has to be extended for them before the file is
published. The track's `README.md` and `TRACK.md` stay as published on 2026-10-14; on the research repository their
TD-002 lines still describe the run as armed.

# TD-001: the outcome, and the descriptive addendum
python $TD/scripts/td001/analyze.py --instrument decide-calibrate.py --run $TD/evidence/td001 --out analysis.json
python $TD/scripts/td001/descriptive_addendum.py --instrument decide-calibrate.py --run $TD/evidence/td001 --out descriptive-addendum.json
# TD-000: TD-001's registered inputs
python $TD/scripts/td000_analyze.py --instrument decide-calibrate.py \
  --instrument-sha256 c427d28f461363de3acb27fccea889490ea7f5ab3ac41bd8531b63593fbe7053 td001 \
  --questions $TD/data/questions-slice1.jsonl $TD/data/questions-slice2b.jsonl \
  --rows $TD/evidence/td000/K1/rows.jsonl $TD/evidence/td000/K2/rows.jsonl --tags kvist-slice1 kvist-slice2b --out td001-inputs.json

On this copy (Python 3.12, numpy 2.3.5) each output is byte-identical to the committed file (evidence/td001/analysis.json, evidence/td001/descriptive-addendum.json, evidence/td000-analysis/td001-inputs.json), as are TD-000's K2 readouts and I1-Q's instrument check. The other analyses are listed in evidence/td000-analysis/analysis.log. What this copy cannot re-run: the GPU phases (they need the runner binary, the Kvist weights and, for TD-001's scoring, the adapter), the question builders (they read an earlier study's records and a local task bank, outside this directory), the contamination audits (they need the distillation corpora), and anything that reads slice 2. TD-001's phase 0 also refuses on two harness files that this copy edits (PUBLIC-COPY-CHANGES.md, "Hash pins").

Result. In the order the track's rules set: what was manipulated, what was observed, what survived, what failed, with the unflattering results first.

TD-001: INCONCLUSIVE. On the pooled ledger held-out (32 units, 174 questions), F against B+map gave Δ̂ = −0.074 nats, 95% paired bootstrap [−0.128, −0.037], against a margin of 0.05. PASS needed U < −0.05 and KILL needed Δ̂ ≥ 0 or L > −0.05; neither held. The interval lies wholly below zero, so it excludes "no difference"; it does not exclude a gain smaller than the margin, and it does not exclude a gain of the margin or more. The registered rule has no outcome for that band, and none is added here.

• Manipulated (TD-001): the adapter, that is, the training; one seed. Held fixed: the runner, the instrument, the recorder, the rendering and prefix, the Kvist Q8_0 file, the joint split and the dev subset.
• Observed, the unflattering results first:
  • No gain on slice 1's ledger. F against B+map: +0.004 [−0.016, +0.023]. F still answers "yes" to all 84 of slice 1's held-out ledger questions, at 50% accuracy and log loss 0.698, above ln 2. The pooled gain comes from slice 2b alone: −0.148 [−0.233, −0.078] on its 90 questions.
  • Few "no" answers. F answered "no" on 12 of the 174 pooled ledger questions (6.9%; B: none). All 12 lie on units of the k=64 pass-rate items, and 8 of them are the rate-labelled rows. The registered expectation of at least 20% did not hold.
  • The interval was wide. The bootstrap SE was 0.023, the top of the design's range (0.010-0.024); at that SE a PASS needed an observed gain of about 0.095 nats.
  • Raw F's ECE is worse than B+map's in four of the five classes (only classify is better). The primary is log loss; ECE decides nothing here.
  • Gains elsewhere, descriptive. On slice 2b NLI the argmax changed across a group's variants in 11.1% of held-out groups, against B's 63.7%, and accuracy rose from 42% to 66% while AUROC moved from 0.59 to 0.60. Per-class log loss of F against B+map: NLI −0.229, classify −0.779, synthetic −0.515. These are secondaries; no decision rests on them.
• Survived: every instrument check passed (I0, base reproduction I-B and zero-adapter I-Z bit-identical, render and mask I-R, VRAM I-V, training health I-T, adapter transfer I-A, completeness I2, envelope I4, aggregate I7), and so did every guardrail: fidelity N1 (mean KL to the base 0.0040 nats, top-1 agreement 97.50% over 65,536 positions), accuracy N2 and invariance N3. The run took 4.25 h of its 12 h slice cap.
• Failed: the primary did not decide. The registered 1.5 GB VRAM margin did not hold in training (430 MiB left; I-V passed). The pinned analysis did not produce some registered secondaries: the exposure log never existed, and the reliability tables, abstention, AURC and the S1 and S4 intervals were missing until the addendum.

TD-001's descriptive addendum (descriptive; no verdict). By the owner's decision, the missing secondaries were computed afterwards on CPU from the committed rows, with the primary's bootstrap. They change nothing: TD-001 stays INCONCLUSIVE and its successor stays open. Unflattering readings first: F's ECE is above B+map's on slice 1 ledger (+0.118 [+0.107, +0.180]) and on NLI (+0.062 [+0.000, +0.102]). F's AURC is below B+map's in every class, with every interval below zero, the pooled ledger's too (−0.209 [−0.375, −0.050]). On slice 1's ledger, where every arm answers "yes" at 50% accuracy, F's AURC is 0.352 against B+map's 0.593, from the order of its confidences alone. Whether that ranking is usable was not defined before the data and is not decided.

TD-000: the measurement stage, ended at its slice cap.

• Manipulated: the model (Kvist-14B Q8_0, Muse-Glimmer-30B Q4_K, Qwen3.5-4B Q4_K_M) and the question file (slice 1, slice 2b).
• Missing, first: three runs produced no rows (Muse and the 4B on slice 1 under the common setting, and the 4B's determinism subset). The runs held the slice for 17.7 h against a 16 h cap, because the slice counter restarted at a resume; the completed runs cost 2.9 times their central estimate. The chain's analysis stage never ran, and the readouts were computed afterwards from the committed rows. Muse's and the 4B's slice 2b rows are inconclusive by the completeness check (one 11.4 k-token item timed out), so they are references only.
• Observed, the worst first:
  • The ledger decisions carry no information. Kvist answered "yes" to every ledger question, 464 of 464 on slice 1 and 544 of 544 on slice 2b; Muse answered "yes" to all 542 of its scored slice 2b ledger questions. Ledger accuracy equals the share of "yes" labels.
  • Order sensitivity on NLI. On slice 2b NLI the argmax changed across a group's variants in 68.9% (Kvist), 74.4% (Muse) and 98.1% (the 4B) of held-out groups. A temperature map cannot move an argmax.
  • Little probability lands on the options. Kvist's median option mass is 0.14 on yes/no, 0.010 on the NLI words and under 0.001 on the class names.
  • The free map. Fitted per class on the training side, it took Kvist's slice 2b NLI log loss from 1.836 to 1.083 (ln 3 is 1.099) and classify from 3.149 to 1.138; on slice 1's ledger its inverse temperature sat at its bound and made the answer uniform (log loss 0.694, ln 2).
• Survived: both instrument reproductions of the 2026-09-22 baseline rows (I1-M, and I1-Q across host and build), determinism for Kvist and Muse (every log-probability bit-identical on a fresh server), the aggregate check (I7) for every analysed run, and the envelope check (I4) on every completed request.
• What TD-001 took from it: pooled ledger log loss as the primary, because the map already drives two ECE targets below zero and the registered sample sizes for ECE were 7 to 140 times the held-out groups that exist; and the 0.05-nat margin, TD-000's decision 6, fixed before its data.

Links.

• Model card: https://huggingface.co/Joakimpalm-Zen/Xyntetik-Kvist-14B (GGUF: https://huggingface.co/Joakimpalm-Zen/Xyntetik-Kvist-14B-GGUF)
• Kvist-14B's training record: https://huggingface.co/datasets/Joakimpalm-Zen/Xyntetik-Kvist-14B-training-record
• Xyntetik Runner, the /v1/decide endpoint and decide-calibrate.py: https://github.com/Joakimpalm-Zen/xyntetik-runner
• Collection: Xyntetik Kvist: distilled agents
• Start with typed-decisions/report/TD-001.md, then typed-decisions/prereg/TD-001.md; TD-000's are beside them.

Third-party data and licences

The record's own content (code, preregistrations, reports, logs, measurements, the synthetic questions and the question templates) is Apache-2.0. The third-party material keeps its own licence; THIRD-PARTY-NOTICES.md carries the full attributions and every licence and notice text. Each licence was read on 2026-10-14 from the publisher's Hub card and from the publisher's own repository, page or paper. Where a mirror's card and the publisher disagree, the publisher's licence applies.

source	in this dataset	licence	attribution and changes
WANLI, alisawuffles/WANLI at 61c95318 (test split)	data/sources/wanli-test.jsonl (the Hub file, unchanged); 1,300 pairs rendered into slice 2b	CC BY 4.0 (the first author's Hub card; the paper and the GitHub repository state no licence)	Liu, Swayamdipta, Smith and Choi, "WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation", Findings of EMNLP 2022, arXiv:2201.05955. Changed: rendered as premise and hypothesis, labels mapped to options, a class-balanced draw. The pairs were written by GPT-3 and labelled by crowdworkers
banking77 (PolyAI), read through mteb/banking77 at 18072d26	data/sources/banking77.jsonl (900 test rows, verbatim); 600 rendered into slice 2b	CC BY 4.0 (publisher card PolyAI/banking77 and the PolyAI repository; the mirror's card says MIT)	Casanueva, Temčinas, Gerz, Henderson and Vulić, "Efficient Intent Detection with Dual Sentence Encoders", ACL 2020 NLP4ConvAI workshop, arXiv:2003.04807. Changed: rendered as an utterance, intent names with spaces, four seeded distractors
MASSIVE (Amazon), read through mteb/amazon_massive_intent at 940fd47a	data/sources/massive.jsonl (900 English test rows, verbatim); 600 rendered into slice 2b	CC BY 4.0, Copyright Amazon.com Inc. or its affiliates (publisher card AmazonScience/massive, its LICENSE and NOTICE; the mirror's card says Apache-2.0); English text from SLURP, CC BY 4.0	FitzGerald et al., "MASSIVE: A 1M-Example Multilingual Natural Language Understanding Dataset with 51 Typologically-Diverse Languages", arXiv:2204.08582; SLURP, Bastianelli et al., EMNLP 2020. Changed: as for banking77
20 public Python libraries (186 upstream commits): attrs, boltons, cachetools, click, dateutil, deepdiff, glom, humanize, itsdangerous, jinja, marshmallow, more-itertools, packaging, pyparsing, python-tabulate, rich, structlog, tenacity, tomli, toolz	the ledger questions' states: commit messages, tests and function source, verbatim, in both question files	MIT, BSD-2-Clause or BSD-3-Clause, Apache-2.0, read from each repository's licence files at the commits used	every licence and notice text, with the repositories and commits, in THIRD-PARTY-NOTICES.md. The commit messages carry upstream contributors' names and addresses as the repositories publish them
Hermes-4-14B (Nous Research) outputs	the proposed edits inside the ledger states	Apache-2.0 (base Qwen3-14B, Apache-2.0); no terms on outputs	sampled at temperature 0.8 in an earlier study
Kvist-14B, Muse-Glimmer-30B and Qwen3.5-4B readouts	every row under evidence/ (numbers only, no generated text)	Apache-2.0 each; Muse-Glimmer-30B also has Meta's usage policy	none required beyond the models' names

Excluded for licence reasons: SNLI (CC BY-SA 4.0) and MultiNLI (mixed; its FICTION genre includes CC BY-SA 3.0 text) rows, and slice 2, the question file built from them. They were the NLI class of slice 2 until 2026-10-08, when slice 2b replaced them with WANLI to keep share-alike out of anything published.

Public copy

This dataset is a public copy of the shade record, not the record itself. PUBLIC-COPY-CHANGES.md lists every change: the three excluded files (X1-X3), the owner's words paraphrased (5), machine and session names replaced by role names such as "the GPU host", "the existence host" and "the runner maintainers' session" (51), and the local commit identity replaced by <commit-identity> (2). Every edit keeps the file's line count, and the file gives the hash pins the edits move. Third-party text inside the data is kept verbatim. That includes the upstream contributors' names and e-mail addresses in the ledger questions' commit messages and test code, as those repositories publish them. Removing them would change both question files' pinned sha256 and every state digest under evidence/, and the reported numbers would no longer reproduce from this copy.

About Xyntetik

This repository is published by Xyntetik, the Zenova AB program for running capable language models on hardware you own and proving what they did. Its released engine is Xyntetik Runner: a free, single-binary GGUF engine in C for CPU, CUDA and Metal that serves, scores and trains through the quantized file it deploys and records every run as a receipt. The measurements behind the program's claims are collected at xyntetik.com/evidence. The program's organization on Hugging Face is Xyntetik.