datapce readout: What came out, failures included

Failures come first. All of it comes from one maintainer, not from outside readouts.

3 · What it taught so far

What came out, failures included?

Failures come first. All of it comes from my own data, not from outside readouts.

Deny rules replayed over past Bash calls

datapce-scan replayed two deny-rule sets over 48,982 past Bash calls of one operator, subagents included, up to 2026-10-21. It reads the logs only. It does not block, change or send anything.

  • A naive set of 6 prefix rules gave 269 denies. 233 of them had only temp-dir targets.
  • Its rm -rf / prefix rule fired 257 times. A rule for the root only fired 0 times.
  • The fixed set gave 20 denies, all git reset --hard. 10 of them had a temp-dir target.
  • A temp-dir target marks a likely false deny. Other denies were not checked by hand.

MEASURED 2026-10-22 · SCAN1: datapce-scan replay, own logs, read only. MEASURED 2026-10-22 · SCAN2: MIT, 97 tests, CI on Linux and macOS, Python 3.9 and 3.13. Code: codeberg.org/datapce/datapce-scan

Fig. 4 — change-chain links a commit to the agent session and the policy behind it. Before the hook, 0 % of 142 commits had a session link. After it, 95.0 % of 20. The same repo since 2026-10-01 shows 12.3 % of 163. The after-hook count is one repo over about 8 hours. It is not a benchmark, and the records are not signed. Links to a deployment and a review are not built. (coverage = commits with a session link ÷ commits) MEASURED 2026-10-20 and 2026-10-21 · chain.py coverage. The tool's 108 tests pass (python3 -m unittest, 2026-10-21). Next check PLANNED 2026-10-28: at least 90 % of new commits linked after 7 days. The link words (linked, inferred, missing) are the tool's own, not the page labels.
Text version of Fig. 4
Session and policy links, by population
PopulationCommitsSession linkPolicy linkDateSource
before the hook, one repo1420 %not given2026-10-20chain.py coverage
after the hook, about 8 h2095.0 %100 %2026-10-21chain.py coverage --since 2026-10-21
same repo since 2026-10-0116312.3 %12.9 %2026-10-21chain.py coverage --since 2026-10-01

Since 2026-10-21, on the repo of this site: 57 commits. Session link: 50 established (87.7 %), 7 inferred. Policy link: 57 (100 %). Deployment link: 5 (8.8 %). These are link words of the tool, not page labels. change-chain is not published yet. MEASURED 2026-10-22 · CC10: chain.py coverage --since 2026-10-21

Fig. 5 — Small models, negative results first. A 1.75M-parameter JEPA world model, trained for about 15k steps (25,311 in the dataset), failed gate G1 three times, and the stop rule fired. Of the 5 models tested for next-action prediction on one private dataset, logistic was best. Action shapes predict failed changes only weakly. The HMM gain is small, and its regime was a self-made proxy fix. No QLoRA run exists. (gain = (logistic − JEPA) ÷ logistic; Brier gain = chain − HMM) Bars start at 0, except the AUC panel, which starts at chance level. ESTABLISHED 2026-10-08 and 2026-10-18 · files named in each panel; PLANNED 2026-10-21. One operator, and private datasets; the shape probe uses public data (CC BY 4.0).
Text version of Fig. 5
Model results
ModelResultVerdictLabelDateSource
JEPA world model vs logistictest cross-entropy: logistic 1.681, Markov-2 1.812, JEPA 1.868, prior 2.120 nats/step; JEPA vs logistic -11.1 % [-11.7, -7.2]failed gate G1 three times; the stop rule firedESTABLISHED2026-10-082026-10-08-p6-g1-attempt-3.md
action shapes vs failed changesAUC 0.649 +/- 0.009 at n=12,000; rule 0.505weakESTABLISHED2026-10-18ml/swe_shape_probe.json
HMM regime vs Markov chainBrier: HMM 0.110, chain 0.116, prior 0.162; gain +0.006 [+0.002, +0.011]small gain; the regime was a self-made proxy fixESTABLISHED2026-10-082026-10-08-l1-regime-and-retry-ab.md
QLoRA fine-tune loop0 runs; code and tests built; 0 examples approvednot runPLANNED2026-10-21rag_cycle.jsonl

H1, a preregistered demand test of the old landing page, was stopped by the owner on day 2 of 21. It has no verdict. The study record shows its figure and rules.