coulomb/glas-harness
5
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix: preserve unknown observed model evidence
All checks were successful
ci / validate (push) Successful in 4m11s
Details

Browse source 
Assistant: codex
Assistant-Model: gpt-6.1-sol
Assistant-Session: 01a10238-7ecc-78e2-903d-075f50e913cc
This commit is contained in:
tegwick 2026-10-03 17:34:40 +02:00
parent dc49415c9f
commit 5ce4f38aed
9 changed files with 114 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

38
docs/evidence/2026-10-03-emergent-session-readiness-review.md Normal file
View file

@ -0,0 +1,38 @@
# EMERGENT-WP-0002: session proof and readiness review
Knowledge 0.2.1; package digest d8ef41bacd2c6551ae3b87628004a5df6943928be1c93eb01e09efc6ad782767; emergent-security commit f9cba6fb35425f74e34401b1baf09f6057ace493.
Scoped technical review by delegated Codex worker, 2026-10-03. No profile or
credential activated and no production request made.
GLAS-WP-0008 is finished: blocked profiles refuse before sandbox creation;
unverified means a labeled bounded attempt; ready requires positive evidence.
GLAS-WP-0012 remains blocked on the owner delivery and real-provider acceptance
already recorded in T02–T05. A direct Codex/Claude/Grok CLI comprehension run
cannot close those gates or establish credential isolation.
Current contract supports start_session, dispatch_tool, end_session and cleanup;
it supplies a generated task file and opaque bounded rein run. There is no
provider resume, compaction or provider-switch event API in this boundary.
The generic task description can request orientation loading, but it is no
attestation that the provider loaded or retained knowledge. Session state in
the Python adapter is not a provider conversation/session identifier.
Local correction: configured model pins no longer fill missing resolved_model
in gateway or either adapter. Provider/rein-reported model remains evidence;
absence remains null. A rein report may still be an alias, not an exact model
revision. No session identifier or exact revision is fabricated from a profile.
Tests check missing reports and distinct observed models on both adapters.
Acceptance requirements retained in GLAS-WP-0012 T04: after owner readiness,
record profile/runtime/rein versions, provider-reported exact model and session
ID (unknown if unavailable), package digest and source status, startup retrieval
and comprehension, then provider-observed resume/compaction and repository /
provider transitions with post-event reload receipts and scoped comprehension.
Do not simulate an event and label it actual continuity. Preserve failed runs
and proof scope. Until native event evidence exists, EMERGENT T05/T07 remains
incomplete. These conditions extend existing proof requirements, not readiness.
Validation: focused gateway/adapter tests 68 passed; full `tests/` suite
139 passed with `.venv/bin/python -m pytest tests/ -q -p no:cacheprovider`.
`git diff --check` passed. These are deterministic regression tests, not
native provider continuity, readiness or isolation proof.

9
docs/harness-contract.md
View file

@ -178,3 +178,12 @@ of the returned sandbox. Errors preparing its reachability or SandboxHandle
fail at session_start, with destruction still attempted. Readiness here is the
owner lifecycle state; it does not promote a Glas profile's operational
readiness or independently prove runtime/provider health.
## Observed model evidence (2026-10-03)
`resolved_model` is supplied only by a rein execution report. A configured
model pin is a requested route, and does not fill absent observed metadata.
Unknown remains null, including failure before provider execution. A reported
alias does not attest an exact provider revision or conversation identifier.
See `docs/evidence/2026-10-03-emergent-session-readiness-review.md`.