The guard checked whether the installed SDK satisfied the pin. Nobody had
ever checked whether anyone had READ it. Those are different questions, and
the gap between them was a whole version range: pinned >=0.2.111,<0.3,
premises source-verified through 0.2.110, installed 0.2.120. Every build in
between was admissible and unexamined — `uv sync --upgrade` would have kept
806 tests green on an SDK no one had opened. Written red first: a guard
handed 0.2.140 returned it without complaint.
_VERIFIED_THROUGH is the ratchet. It records the newest build actually read
at source, and a newer one fails naming the five premises to re-check. The
pin is untouched and was never the defect — measurement dissolved the
premise that it needed lifting. It was not too narrow but too wide, and a
wider permission is not repaired by widening it further.
The premises themselves were prose the failure message recited. Nothing
tested them, so one that stopped being true would have surfaced on the one
live paid run (S10, D6). They are now a table introspected against the
installed package, with the printed prose derived from that same table so a
checked attribute cannot go unreported or a reported one unchecked. The
premise introspection structurally cannot see — that query() yields an
AssistantMessage then a closing ResultMessage — is named apart, and is the
honest reason the human reading still has to happen.
Value-proved, not merely named: disabling the ratchet reds 1 test, stubbing
the inventory to "no gaps" reds 3, re-hardcoding the prose reds 1, and
lowering _VERIFIED_THROUGH below the installed build reds the real
installed-version test rather than only a monkeypatched one.
0.2.139 read at source (0.2.120 -> 0.2.139, latest on PyPI today; STATE said
0.2.134, measured 08-09 and stale). The public query.py is byte-identical,
every premise field keeps its type and default, and the parser changes are
additive. One needed a look: 0.2.139 added a skills path defaulting
setting_sources to ["user", "project"], which would have undone the S10
isolation fix — it fires only on None, so the explicit [] is out of reach.
Prose carrying stale version claims moved with the reading, never ahead of
it: each was re-verified at 0.2.139 before being restated.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014dKDjVG7qrBh9NkAAxutqN
The repo has never had a tag, so this is version one. The number is chosen
from maturity, not from habit: the shared spec's V1 `generated` shape is
landed upstream but not adopted here (the golden is the library's emission,
so adoption is gated on the ingest pin swap), and the SDK pin reaches
further than the range whose premises are source-verified. Both gaps are
held open by tests on purpose. 1.0.0 would claim a settled surface this
implementation does not have.
pyproject.toml already declares 0.1.0 and needs no change; README carries
three badges, none of them versioned, and src/ declares no __version__ —
enumerated, not assumed, so nothing else can drift from the tag.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014eYPddPVYPMxc5L7a4nxvA
The advisory finding: provenance.py/artifacts.py stamped no SDK version, while
the SDK's total_cost_usd is a client-side ESTIMATE computed against a price
table frozen when the SDK was built. An untraceable estimate is a figure nobody
can check later, so the run now records which build produced it.
Provenance gains sdk_version: str | None. The value comes from the PRODUCING
CLIENT — getattr(client, "sdk_version", None) — exactly as model and cost_usd
already do, never from importlib.metadata at stamp time. That distinction is
the seam: a run driven by the scripted stand-in used no SDK at all, and
stamping the installed version there would attribute a build to a run that
never touched it (§1). SdkModelClient reads the installed build once from
package metadata (offline: no key, no network); every other client reports
null. A blank string is refused by the schema — null is the one way to say
"not produced by the SDK".
Scope note: this traceability covers OUR run cost only. The savings the
framework recommends are settled by the deterministic validator against the
golden suite, and no SDK estimate touches them.
Two seams, both detach-proven RED:
- make the stamp read importlib.metadata instead of the client → a scripted run
claims a build it never used → red
- back-fill runs/s10/provenance.json → red
That second guard is the point of the change as much as the first. runs/s10/ is
the byte-frozen record of the ONE live run (2026-07-03), executed before this
field existed; the suite reads it nowhere else, so nothing would have caught a
retro-stamp. Adding a build id to it now would be a guess presented as
provenance. It stays without one, and the README says why.
run_s10.py is deliberately untouched (byte-frozen fasit script), and the field
defaults to None, so every existing caller and artifact shape is unchanged.
603 passed · ruff clean · mypy strict clean · runs/s10/ byte-identical.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MQu2xxwedckjU56byu1aUG
The last ungated build session: the operator now drives the whole build from the
command line, and the documents claim exactly what the code does (§1).
run.py becomes the collecting entrance. Exactly one of --bundle (one project) or
--portfolio (N projects from a schema-validated reference config, with
--verdict-dir as the portfolio-level expert inbox) is required; both and neither
are refused. --goals loads a goal contract and checks it against --ledger's
realized sum BEFORE the first model call: the §8 caps bound spend, the goal bounds
achievement, so a hard target the book already meets stops the run at exit 4
without constructing a client. A soft target reached is a flag and the run
continues; an absent ledger is an empty book, so the goal is still evaluated,
never skipped. The one declared goal also drives --value-report's goal progress —
one contract, never two figures that can disagree.
The portfolio path persists nothing (K3 returns typed results; the outbox names
pairs by run_id, which a portfolio pass has none of). Rather than accept
--out/--outbox/--run-id/--value-report/--inbox/--live-dry-run there and silently
ignore them, the entrance refuses them and says why. run_portfolio is imported
lazily — portfolio.py imports this module, so a module-level import is circular.
Three seams, each detach-proven RED:
- unwire the goal check → the run proceeds and spends → red
- unwire the portfolio branch → the configured projects never run → red
- document a flag no CLI offers → the README honesty grep goes red
That last one is the doc-sync made load-bearing: the test reads README.md,
collects every --flag it documents (excluding third-party dev-tooling lines) and
asserts each exists in the --help of a CLI the README names. The drift it exists
to close was real — README claimed 562 tests, CHANGELOG claimed 265, actual 597.
Docs synced to the code: README gains an operator-CLI section and honest goal/
portfolio descriptions, CHANGELOG is rewritten to what actually shipped, and
docs/oppskrift-kunnskapsbase.md delivers D-H point 1 — the documented team
process for building a knowledge base, with the honest 1–2 week expectation and
every factory-dependent step (verdict translation, demo path) marked NOT BUILT.
597 passed · ruff clean · mypy strict clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MQu2xxwedckjU56byu1aUG
