The next Ekbasis: one forward pass, trained further on what actions actually did.
Given the current state and an action, Skopos answers typed questions — will this lose work? will it fail? what will X be afterwards? — with a calibrated probability, in one forward pass, like Ekbasis-27B and with the same prompt, readout, API and client. It was trained further on executed shell, git, SQLite, PostgreSQL and Python scenarios and real git sequences.
Every number on this page is read from the release's results files, with its pre-registration where there is one; the others are marked not pre-registered or added after review. the full report is linked below. Limits are at the bottom of the page.
The pre-registered release protocol of the Ekbasis repository, the same items for both models, each served the same way. Rule: at most 1 point worse than the released model on each criterion, ECE at most +0.01 (PREREG_R4_RELEASE_READ.md).
git v3, known and held-out command types, all questions
Never-trained worlds, answers the actions change: one prompt per question / read once
ftest_family, answers the actions change (629 items)
Calibration: ECE on git v3
The guard, known command types: work-losing flagged / false alarms (of 863)
The guard, held-out command types: work-losing flagged / false alarms (of 799)
Real-repository git guard (T3c): flagged · false alarms · "fails?" right
The pre-registered release read missed one criterion: ftest_family, on its 629 items (−1.6 points, beyond the 1-point limit), so the release read as registered did not pass. A confirmation was then registered (after that result, before any fresh item existed): the same families on fresh items, by trainer readout; it did not confirm the loss:
"Held-out" command types are held out of the git-sandbox training rows; Skopos' shell and git factory rows include them, so for Skopos those rows are a weaker test than for Ekbasis-27B.
The confirmation on fresh items
5,385 fresh items of the never-trained families (cards, machines, timers), 3,014 of them answers the actions change (trainer readout).
83.4%
Skopos, answers the actions change
81.8%
Ekbasis-27B, same items
+1.6 [+0.7, +2.5]
difference, 95% CI
"Will uncommitted work be lost?", flagged at P ≥ 0.2 (the guard's threshold), on the larger guard set (served), the same items for both models; paired differences (Skopos − Ekbasis-27B) with exact McNemar tests. The held-out command types (git rebase <branch>, git restore --source=HEAD~1 <file>) are held out of the git sandbox's training rows; Skopos' factory rows contain both, in other repositories.
26 / 799
false alarms, held-out types: −2.9 points [−4.1, −1.6], p < 0.001 (Ekbasis-27B 49)
27 / 863
false alarms, known types: −1.2 points [−2.3, 0.0], p = 0.08, not significant (Ekbasis-27B 37)
97.1%
work-losing flagged, held-out types: +1.1 points [−0.1, +2.5], p = 0.15, not significant (Ekbasis-27B 96.0%)
The 16 scenarios on a clone of pallets/itsdangerous that Ekbasis-27B's release ran (the repository has not changed since), the truth from running the commands in a copy. With an instruction planted in a file name, a branch name or the commit messages, no work-losing scenario dropped below the flag.
Ekbasis-Skopos-27B
Ekbasis-27B
Where the new training shows most: executed scenarios from the generators Skopos was trained on, with seeds it never saw — in-distribution for Skopos, new for Ekbasis-27B, so read these as what the training adds, not as generalization. These sets also chose between the candidate runs (by pre-registered rules), which adds selection on top; the independent evidence is the fresh never-seen families and the real agent turns below (trainer readout, paired 95% CI).
Executed factory scenarios (shell, SQL, Python, git)
Real git sequences, 5–12 commands
Same, validation twin
Agent sessions (shell, read as a transcript)
The published confident-errors test, all 15,008 items as pre-registered (served bf16 vs Ekbasis-27B's published predictions). Wrong answers given with confidence ≥ 0.9 are what a guard would act on. In world families it never saw, Skopos is wrong slightly less often than Ekbasis-27B but more often with high confidence, and worse calibrated — a limit, below. Without the 5,381 items that overlap the lineage's training rows, the trained-family and held-out-question rows keep their direction (the evaluation report has both tables); no never-trained item overlaps.
Trained families
… answers the actions change
Held-out question types
… answers the actions change
Never-trained families
… answers the actions change
Each answer can also come at a calibrated temperature (the additive `calibrated` field, client 0.1.41), fit on validation only (fval_git3 + m3val (validation only)). Not pre-registered: INT4 as served (Skopos on an RTX PRO 6000 with vLLM and serve.py; Ekbasis-27B on the hosted API's production server, an RTX 5090), on 2,482 items (git_long 900, guard_known 800, fac_val 782) from generators Skopos was trained on, with seeds it never saw; for Ekbasis-27B only the guard questions come from its training generators (git_long runs its sandbox at lengths it was not trained on; the factory is new to it). So this compares the two models on Skopos' home ground:
Accuracy
Confidently wrong per 100, raw
… at the calibrated temperature
ECE, raw → calibrated
Temperature
Each build ships its own temperature in decision_config.json (remove the key to turn the field off):
bf16
INT4
FP8
With the open builds the state can be text or an image, as in Ekbasis-27B (the hosted API reads text only). Skopos' last training was text only; on 395 new image items from world families it was trained on (only the image modality is untrained; the published drawings were regenerated, so pixels may differ; INT4 against INT4, served) its image ability is preserved by the pre-registered criterion.
79.8%
state given as an image (Ekbasis-27B 78.7%)
94.2%
reading the final state from the picture (Ekbasis-27B 91.9%)
91.9%
from text, same items
455 turns of real agent trajectories (54 of them fail), from Qwen's AgentWorldBench: will the next action fail? Ranking (AUROC): the three Ekbasis models are within 0.010 of each other, inside their confidence intervals. Balanced accuracy at the fixed 0.5 threshold: Skopos is the lowest of the three; choose the threshold for your false-alarm budget, as the guard does. AgentWorld-35B-A3B (thinking off) left 59 answers unparsed, counted wrong; Skopos minus AgentWorld, balanced accuracy: +0.057 [−0.025, +0.136]. AgentWorld's training data may include these tasks.
Ekbasis-Skopos-27B
the first Skopos (unreleased)
Ekbasis-27B
AgentWorld-35B-A3B (thinking off, 87 s per answer)
Before release, the checks of Ekbasis-27B's release evaluation that Skopos had not yet run were run on the same items and scripts, next to Ekbasis-27B's published results (PRELAUNCH_EVAL.md, sealed before its first request). The rule: a paired 95% bootstrap CI (2,000 resamples, seed 20261015) entirely on the worse side of 0 on a set of 100 items or more is a release blocker; smaller sets are counts, and a cell worse by 2 or more is flagged (PRELAUNCH_EVAL.md, Part A).
Release blocker, declared: on 120 short checks in paraphrased wording, Skopos got 113 right against Ekbasis-27B's 117 (−3.3 points, 95% CI [−6.7, −0.8]), so the pre-registered rule fired. The exact test on the same pairs does not reject (p = 0.125; 4 items differ, all against Skopos); Ekbasis-27B re-run on the same machine reproduced its answers item for item (240/240); in the original wording Skopos got 116 against 115, and over both wordings 229 against 232 of 240 (−1.25 points, [−3.3, +0.8]). It is consistent with the confident-error limit in never-trained world families. Released with it declared, by the owner's decision (2026-10-16).
Not run as published: P12 preflight on 141 plans (plans' snapshots lost).
Still running at release; their results are appended to the evaluation report on completion: P11, GPU memory of each build.
Every build is evaluated on the same release protocol against the bf16 model, served; the quantization gate was pre-registered. An MLX build for Apple Silicon follows after its test on Apple hardware.
Ekbasis-Skopos-27B
Ekbasis-Skopos-27B-INT4
Ekbasis-Skopos-27B-FP8
Self-hosted: download a build and point the client at it — the API, the client, the git and shell guards, the Claude Code hooks, check_plan, monitor() and simulate work unchanged.
Hosted API: it serves Ekbasis-Skopos-27B-INT4 since 2026-10-16 14:29 UTC (it served Ekbasis-27B-INT4 until then). The switch came on release day, with less notice than the 14 days our API policy promises; Ekbasis-27B's weights stay public for anyone who needs the previous model. Every answer names its model in a response header. The hosted API reads text only; the open builds served with serve_vllm.sh also read images.
Certified mode of ekbasis.verify (thresholds certified on Ekbasis-27B): checked with Skopos on 1,400 fresh scenarios — INT4: 7 errors among 927 answers accepted without verifying (weighted 0.74%, 95% upper bound 1.41%; target at most 2%), 22.0% verified; bf16: 6 among 935 (bound 1.26%). On the same scenarios Ekbasis-27B bf16 made 5 among 886 (weighted 0.47%, 25.4% verified): Skopos accepts more and errs slightly more often. Per rule node the check has little power (43 nodes, median 16 accepted answers; the weakest 2 errors in 12, p = 0.023 uncorrected, not significant after Bonferroni). The certificate carries over; a check, not a new certificate.
hf download caiovicentino1/Ekbasis-Skopos-27B-INT4 --local-dir Ekbasis-Skopos-27B-INT4
bash Ekbasis-Skopos-27B-INT4/serve_vllm.sh Ekbasis-Skopos-27B-INT4 8001
python Ekbasis-Skopos-27B-INT4/serve.py --model Ekbasis-Skopos-27B-INT4 --vllm-url http://127.0.0.1:8001 --port 8000
pip install -U ekbasis && EKBASIS_URL=http://127.0.0.1:8000 ekbasis git-check -- "git reset --hard"The scorecard of the hosted API's change of 2026-10-16 14:29 UTC, posted the same day (our API policy asks for 14 days; this switch came on release day). It compares Ekbasis-Skopos-27B-INT4 (what the hosted API serves since 2026-10-16) against Ekbasis-27B-INT4 (what it served until then). Accuracy: served, the release protocol's sets (the same items for both).
git v3, known and held-out command types, all questions
never-trained worlds, answers the actions change, one prompt per question
never-trained worlds, answers the actions change, read once
ftest_family, answers the actions change (629)
ECE on git v3 (10 bins)
the guard at 0.2, larger set: work-losing flagged / false alarms, known (of 863) and held-out (of 799) types
never-trained families, fresh confirmation (5,385 items; against Ekbasis-27B bf16, trainer readout)
The regression a hosted user should know about — never-trained world families of the published confident-errors study: wrong at confidence ≥ 0.9 per 100, measured on the INT4 (served), both builds (PREREG_ADDENDUM A7): raw answers, Skopos 4.75 against Ekbasis-27B 3.65 on 5,364 items (+1.10 [+0.65, +1.57]). Reading the calibrated field (T 1.67 and 1.55) cuts them to 1.62 and 1.10: Skopos calibrated is below Ekbasis-27B raw, and calibrated against calibrated Skopos still has +0.52 per 100 [+0.26, +0.78]; calibrated ECE is level (0.021 against 0.022). A7 on bf16 not run (the evaluation machine was retired).
Confident errors and calibration on 2,482 items, served through the System One API, both INT4 (from generators Skopos was trained on: its home ground).
Accuracy
Wrong at confidence ≥ 0.9, per 100: raw
… at the calibrated temperature
ECE: raw → calibrated
Calibrated temperature
Latency and cost on 200 fixed items, one RTX PRO 6000, vLLM + serve.py, nothing else on the GPU (the hosted API runs on an RTX 5090). Cost on these fixed items (about 300 input tokens each); a check's cost scales with its input tokens.
One at a time: median / p90
16 in parallel
Input tokens per request
Cost per 1,000 checks (at $0.30 per 1M input tokens)
A general-purpose calibrated evaluator, Jev (typesafe-ai/jev on the Vercel AI Gateway, zero data retention), on the same items with the same state text: it, too, answers typed questions with probabilities in one pass. Measured 2026-10-16, not pre-registered; two prompt variants (Ekbasis' question and one written in Jev's documented style), the headline variant chosen per set by a rule fixed before scoring. 14,976 calls cost US$ 0.34.

git v3, 3,105 questions: accuracy
git guard, known types (1,726): will it lose work?
git guard, held-out types (1,598)
git3_known
git3_held
guard_held
worlds
intent
Reading: on questions about the repository's state Jev comes within 6–7 points; as a guard it does not separate work-losing commands from safe ones (its probabilities sit near the middle), so at the guard's threshold it flags nearly everything. In its favour: it is about four times cheaper per check, and it is almost never confidently wrong — fewer confident errors than Ekbasis-27B and Skopos on every suite where their per-item answers are published. Method, per-item answers and a script that recomputes every number: results_skopos/jev_comparison in the code repository.
Lineage: Qwen3.8-27B → Eikos-27B → Ekbasis-27B → the first Skopos (v49f_w50, not released) → Ekbasis-Skopos-27B.
The first Skopos: Ekbasis-27B's adapter (rank 128) trained further for 900 steps, then the exact interpolation halfway between it and Ekbasis-27B's own adapter (WiSE, λ 0.5), merged; an intermediate checkpoint that was not released: its weights are archived privately, and the published code and data retrain it approximately. Then a new LoRA (rank 64, all linear layers of the language model, 1,200 steps, cross-entropy on the answer letter plus the JEPA-style hindsight term) trained on the first Skopos and merged into its weights.
λ 0.5 was chosen by the selection rule of its plan (PLAN_v49b), on a validation set of world-family questions (trained, held-out question types and never-trained families): the un-interpolated checkpoint failed its accuracy and calibration criteria (mean accuracy over the groups 88.19% against a floor of 88.85%; ECE 0.061 against Ekbasis-27B's 0.048, tolerance +0.01), λ 0.75 failed the accuracy criterion (88.79%), and λ 0.5 passed both (89.30%, ECE 0.051). The last stage's selection (R1 → R2 → R4) had no confident-error or world-family calibration criterion, and the released model shows more confident errors in never-trained families (Limits).
Its training data: shell, git, SQLite, PostgreSQL and Python scenarios generated by our factories and executed in throwaway sandboxes; the last stage used the factory regenerated with fresh seeds; real git sequences of 4–16 commands executed in throwaway repositories, and guard-style git items with their executed labels; plan-level items in synthetic rule-based worlds, with exact answers from our simulators; a replay of Ekbasis-27B's training mixture (simulated worlds, git, and the eikos-decisions replay rows).
Not in the training data: Vikunja rows built during development were all dropped by the overlap firewall; WebWorld transitions were excluded by rule.
Open: the weights; the training rows of every stage (except 7,423 replay rows that are not part of the published eikos-decisions (as in ekbasis-data) and 31 more rows withheld by its privacy scan, 26 of them CWM rows (all listed by id)); the trainers and recipes as they ran; the generators that computed the rows; every evaluation with its per-item predictions (except the quantized builds' answers on the fresh confirmation, kept as summaries); the candidate runs that were not released and why; and the pre-registrations.
shio-bash-26-1
swe-rebench-v2-cwm
Contamination audit of every evaluation set against every training row of the lineage: identical prompts (also against the hindsight prompt), and near-duplicates at word 8-gram Jaccard >= 0.5 against one training text after removing each set's own layout (Ekbasis-27B's audit rule, exact best match). Not covered: Eikos-27B's own decision data, Qwen's pretraining, T3c (its scenarios are not dataset items). Items flagged: multi_test_trainfam (reported only) 291 of 600; fac_val 105 of 3,983; fhorizon 547 of 1,375; image (new seeds) 5 of 400; confident-errors set 5,381 of 15,008. Numbers that use those sets are reported without the flagged items as well.