Skip to content

Ekbasis-Skopos-27B: the hosted switch on the release day, and the pre-launch results appended after the release - #63

Merged
caiovicentino merged 3 commits into
mainfrom
skopos-followup
Oct 16, 2026
Merged

caiovicentino merged 3 commits into
mainfrom
skopos-followup

Conversation

@caiovicentino

Copy link
Copy Markdown
Member

Follow-up to #62 (Ekbasis-Skopos-27B).

The hosted API changed on the release day. It serves ekbasis-skopos-27b-int4-v1.3 since 2026-10-16 14:29 UTC, with
less notice than the 14 days the model-change policy promises: the owner's decision, recorded in
results_skopos/prelaunch/RELEASE_DECISION.md ("Second decision"). Ekbasis-27B's weights stay public. The CHANGELOG has
its own entry for the switch, and the release entry points to it.

Pre-launch results appended after the release, as RELEASE_EVAL_SKOPOS.md listed them at release
(results_skopos/prelaunch/compare/PENDING.json):

  • P5: speed pairing: Ekbasis-27B bf16 re-run on this machine (phase B); P5_partial.json stays as released (P5.json)
  • P6: item pairing: Ekbasis-27B bf16 re-run (phase B; the release run kept no per-item answers); P6_partial.json stays as released (P6.json)
  • P1: official identity re-run of Ekbasis-27B bf16 (phase B); P1.json unchanged (P1_identity.json)
  • P8: official identity re-run of Ekbasis-27B bf16 (phase B); the early re-run is already in P8.json / P8_same_machine.json (P8_identity_phaseB.json)
  • P12: preflight SUBSTITUTE, bf16 pairing (needs the phase B Ekbasis-27B bf16 replay); reported separately, never with published numbers (P12_preflight_substitute_bf16.json)

None of these files holds a release blocker; none of them could block under the rule.

Not run, each with its reason:

  • P12, intent replay same-machine control (Ekbasis-27B INT4 served on the evaluation machine, phase E); the rule's reference, the hosted published run, is in P12_intent.json: the evaluation machine was retired on 2026-10-16
  • P12, preflight SUBSTITUTE, INT4 pairing (needs Ekbasis-27B INT4 on the same requests, phase E); Skopos INT4's replay exists (prelaunch_out/P12/preflight_substitute/skopos_int4.jsonl) but has no Ekbasis-27B pair: the evaluation machine was retired on 2026-10-16
  • P11, GPU memory caps (phase F): the evaluation machine was retired on 2026-10-16
  • P9, Ekbasis-27B re-run on this machine (seconds, puzzle identity); optional: the evaluation machine was retired on 2026-10-16

RELEASE_EVAL_SKOPOS.md is rebuilt from results_skopos/ with every pre-launch item final or reported as not run; the
battery's outputs of phase B are added to results_skopos/prelaunch/out/.

Not in this pull request: A7 of prereg_skopos/PREREG_ADDENDUM.md (confident errors with every option's probability, on
the two INT4 builds), which runs after the switch; its results come in a later pull request.

🤖 Generated with Claude Code

Caio Vicentino and others added 3 commits October 16, 2026 11:11
…nd decision

- CHANGELOG: a "Hosted API" entry: the hosted API serves ekbasis-skopos-27b-int4-v1.3 since 2026-10-16 ⟨HH:MM⟩ UTC,
  on release day, with less notice than the 14 days the model-change policy promises (owner decision,
  results_skopos/prelaunch/RELEASE_DECISION.md); Ekbasis-27B's weights stay public; calibrated.temperature 1.67 instead
  of 1.55. The release entry's hosted sentence now points to it instead of the 2026-10-30 date. ⟨HH:MM⟩ stays until the
  switch's time is set (SWITCH_TIME).
- results_skopos/prelaunch/RELEASE_DECISION.md: the record as updated with the second decision.
- write_release_eval_skopos.py: a full --final build lists what was appended after the release and what was not run
  (PENDING.json's "not_run", with its reason).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…nd F not run

RELEASE_EVAL_SKOPOS.md is now built with --final (no release mode): every pre-launch file PENDING.json lists is here.
Appended after the release (phase B, on the evaluation machine before it was retired): P5's speed pairing (P5.json),
P6's item pairing (P6.json), the identity re-runs of Ekbasis-27B for P1 and P8 (their rule outcomes stay in P1.json
and P8.json), and the P12 preflight substitute on bf16 (not under the rule, never compared with published numbers).
Not run, the evaluation machine having been retired on 2026-10-16: P12's same-machine intent control, the preflight
substitute's INT4 pairing, P11 and P9's Ekbasis-27B re-run. No new blocker: P8's paraphrased wording remains the only
one, declared. Phase B's outputs are added to results_skopos/prelaunch/out/ and the as-run compare scripts updated.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant