Fase 1b, funn 1b. Den første levende kjøringen brant tolv runder på svar som
ikke lot seg parse til IR-formen; e371890 gjorde teksten synlig, dette fjerner
årsaken. generate_via_llm sender nå
options={"response_format": proposal_response_format()} på hvert
genererings-kall.
Formen er MÅLT, ikke valgt. ChatOptions.response_format tar
type[BaseModel] | Mapping, og begge profiler ærer den: LOCAL sender en Mapping
ordrett til Chat Completions, AZURE (FoundryChatClient -> RawFoundryChatClient
-> RawOpenAIChatClient) konverterer samme envelope til Responses-APIets
text.format. Klassen — det korteste svaret — er avvist på bevis: gitt en klasse
konverterer klienten med type_to_response_format_param, som emitterer
minimum/exclusiveMinimum/minItems/prefixItems og et assumptions-node hvis
additionalProperties er et skjema. Azures publiserte subset utelukker alle fire.
assumptions kan ikke bare droppes, og det er også en måling: validator
._monte_carlo faller tilbake på item.unit_cost for hver kode uten bånd, så uten
bånd er alle 512 samples identiske og P10 == P50 == P90. Den stokastiske
falsifisereren ville gått inert mens den fortsatt rapporterte persentiler.
Wire-en bærer derfor et array av navngitte entries som _parse_ir folder tilbake
til IR-ens map — additivt, aldri erstatning. Skjemaet deriveres fra
SavingsProposal; sanitiseren er fail-closed (StructuredOutputUnsupported).
Load-bearing målt mot hele suiten, seks mutasjoner alle røde, grønn kontroll
864/4: detach wiringen (1) · detach sanitiseren (3) · dropp assumptions fra
skjemaet (1) · fail-closed -> stille reparasjon (1) · detach normaliseringen
(3) · erstatning i stedet for tillegg (2, inkl. golden-transkriptet).
T3 ble skrevet vakuøs først og felt av sin egen mutasjon: den påsto å bli rød
når assumptions forsvant fra skjemaet, men den scriptede klienten ignorerer
skjemaet. Testen fikk en direkte assert på skjemaet.
Ærlighets-grense: ingen betalt kjøring gjort. Testene beviser konformitet med
det dokumenterte subsettet, ikke aksept fra det levende endepunktet.
859 -> 864 passed / 4 skipped. ruff + format + mypy rene.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EQNU4tfAhsBvdefT1jUhk
Fase 1b funn 1 lukket. generate._fetch_parsed kastet hvert uparsebart modellsvar i
except: continue, så prosjektets første levende kjøring brant tolv runder på formatfeil
og etterlot null tegn av det modellen faktisk sa. Enhver videre betalt kjøring ville
vært gjetning.
HVOR teksten overflates er avgjort av en måling, ikke av symmetri med Steg 5:
meter.tick_round() raiser BudgetExceeded INNE i _fetch_parsed, og uten mandat fanger
ingen den, så på nøyaktig den stien fangsten finnes for returnerer generate_via_llm
ingenting. Et felt på GenerationResult ville vært blindt for den, og et outbox-artefakt
skrevet etter kjøringen likeså. Sinken speiler meter: en kaller-eid akkumulator hvis
innhold kalleren holder uansett hvordan løkka endte. Artefaktet skrives fra en finally,
ikke except BudgetExceeded, og kun når noe faktisk feilet.
Iron Law: testfila rød ved collection FØR modulen fantes. Seks mutasjoner mot HELE
suiten, alle røde, hver med sin egen signatur; grønn kontroll 859 passed / 4 skipped
(fra 854). Den skarpeste er trunkering som BEHOLDER sentinelen: da faller kun
verbatim-asserten, som er det som beviser at den ene testen bærer den egenskapen.
Samme økt: mutasjonsmålingen økt 34 utsatte for de to Fase 5-gatene er kjørt. Fire
preflight-mutasjoner mot hele suiten, alle røde på riktig test og ingen annen (detach
fallbacken 2 røde, snu presedensen, presence i stedet for truthiness, avslaget navngir
kun vårt navn). To handover-mutasjoner kjørt MÅLRETTET mot egen testfil under tidspress,
ikke mot hele suiten — uttalt, ikke skjult (drop uv.lock, bygg fra arbeidstreet i stedet
for tracked files). De to DEPLOY.md-mutasjonene gjenstår: git archive leser HEAD, ikke
arbeidstreet, så de krever en midlertidig commit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WLQd8ojQ9xwxhB8vsETYBs
generate_via_llm consumed each validator Rejection internally (`last`), fed it into the
next attempt's prompt, and dropped it. So Step 5 was real but unobservable: a caller could
see THAT a proposal validated, never that it validated on attempt 2 after the deterministic
validator falsified attempt 1. It was the one step of the eight with no output to show.
The seam is a typed return value -- GenerationResult(outcome, refinements) -- rather than an
out-parameter or a callback: a returned value cannot be silently lost by a caller that forgets
to pass a collector, and mypy forces every call site to acknowledge it.
refinements carries ONLY rejections that were actually fed back. When the attempt budget runs
out the final rejection IS outcome; counting it here would be double-counting, and the bounded
control test goes red on the collect-everything implementation that gets this wrong.
The loop's bound is untouched: max_attempts and meter.tick_round stand, and `last` still drives
the prompt alone, so prompt growth is unchanged. run.py accumulates across _evaluate calls, so
_evaluate_mandate is untouched; RunResult.refinements defaults (the coverage precedent) and is
concatenated across approaches rather than keyed per approach -- stated as an honesty limit.
The simulation now shows it: the scripted proposer overclaims 250000, which the validator
falsifies against P90 = 90000, and the corrected 30000 validates. Only the overclaim is
scripted -- the rejection is computed. scripted_factory takes a per-role reply selector so this
needs no second scripted client body.
README records the two accuracy changes only (Step 5 is now inspectable; the simulation trace
shows the correction). The level-2 publishing claim stays deferred until after the demo (O4).
Load-bearing MEASURED against the full suite with a control, four mutations all red:
detach the returned history (4 tests) - collect-everything (control only) - detach the run
wiring (2 tests) - revert the simulation's proposer to a constant (the demo-protection test).
Control: 759 passed / 4 skipped; ruff, format and mypy clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017CcWFcREUi6YPjEpN3ACDP
Operator feedback: fagpersoner must be able to name the approaches a run shall
evaluate for a project, and/or ask the system for its own. Today the hypothesis
prompt is hardcoded ("Propose ONE concrete cost-saving measure") and the only
expert-facing lever, --dimension-config, FILTERS what may pass the scoping gate
rather than DIRECTING what is spent attempts on. This is the input that was
missing.
`mandate.py` is the typed commission + a fail-fast loader (mirrors
`load_dimension`/`load_goal_config`): missing or malformed refuses, because a run
must never proceed on a silently degraded commission — the coverage report would
then describe work nobody ordered. Stdlib + pydantic only, so it joins
`_MAF_FREE_MODULES` and can be mirrored to the D7 sibling.
Two refusals carry real defect classes: an EMPTY commission (no approaches and no
own proposals) is a caller error, not a result; and a duplicate approach id — or
one claiming the reserved OWN_PROPOSAL_ID — would collapse two coverage rows onto
one key (the S3.2 key-collision class), which is exactly the silence the coverage
report exists to prevent.
The numeric target is deliberately NOT duplicated here: it already lives in
GoalContract, and two copies of one number drift apart ((p) precedent). The
mandate carries intent; `announce` merely restates the figure.
`_build_messages(approach=...)` switches the opening instruction from *find one*
to *quantify THIS one*, carrying the expert's label and description VERBATIM —
the description is the reason the approach is worth trying, the one part the model
cannot infer from cost data. `approach=None` is byte-identical to the previous
prompt, so every existing run and golden is untouched.
The gate is unmoved: `validate_proposal` is called exactly as before. A
commissioned approach gets no discount — the expert directs what is EVALUATED,
never what is APPROVED.
Load-bearing MEASURED against the whole 695-test suite, four mutations all red:
detach the approach injection (2 red, control stayed green) · let a commissioned
approach bypass the validator · make the mandate loader tolerant · drop the
empty-commission refusal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ULCqjLF61rehj5cZmdUoR3
Every stage of validate_proposal reasoned only about numbers the proposal itself
supplied, so an internally-consistent hallucination cleared the whole gate (F3).
A new stage 0 reconciles each affected_item against the project's CostBaseline
before the CBC solve: an unknown cost code is rejected, and a real code carrying
a quantity/unit_cost outside the configured tolerance (5% default, relative to
the baseline value) is rejected. Validation, never repair.
The baseline argument is OPTIONAL (None = pre-S4.0 behaviour), but both run
paths set it: the road path projects project.cost_items, the bundle path loads
cost-baseline.json when the bundle ships one. Bundles written before the
amendment stay un-anchored, so the commons-owned goldens run byte-identically;
a baseline that exists but is malformed still raises on both loaders.
F8: the method-specific cap now comes from the METHOD_CAPS registry (measure
type -> fraction, injectable) instead of an energy_efficiency string comparison.
The baseline format and tolerance semantics were decided locally — the commons
amendment (D-A pt. 2) never arrived, exactly as in S3.2. D7 mirroring stays open.
Three portfolio fixtures quoted cost codes belonging to OTHER projects; the new
gate caught them. They now quote each project's own lines, and the two copied
REPLIES tables import the single source instead of drifting from it.
Load-bearing measured (tests/test_s40_cost_baseline_loadbearing.py), six
mutations all red: detach the reconciliation stage; detach the magnitude
tolerance; detach the road wiring; detach the bundle wiring; ignore the injected
cap registry; make the optional loader tolerant of malformed content. Control:
with the road wiring detached the repaired portfolio fixtures still pass, so
they are not masking the seam. 597 -> 612 tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JdwK7bQ4BZkWH4t8MRDKb4
Closes gap #5 (maalbilde §5/§7): generate_via_llm's outer max_attempts loop
built the prompt ONCE and re-sent it identically — a BLIND retry. The validator's
per-attempt Rejection.reason was captured in `last` but never reached the next
prompt, so the proposer re-answered the same question with no knowledge of WHY
the prior candidate failed. Step 5 routes that reason into the next attempt.
- generate.py: _build_messages() gains prior_rejection; when set it appends a
revision block carrying ONLY the falsification reason verbatim (never the
rejected proposal JSON). None -> the byte-identical base prompt, so attempt 1
is unchanged. generate_via_llm() rebuilds messages inside the outer loop with
prior_rejection=`last` (None on attempt 1); _fetch_parsed() takes messages as
an explicit parameter. `last` is overwritten each round -> only the most-recent
falsification ("forrige"), never an accumulated history. Bound unchanged:
max_attempts + meter.tick_round (no new loop; §6 — "improve until good enough"
without a ceiling stays impossible).
- Scope honesty: the only per-attempt falsifier here is the validator. The
checker is a run-level, one-shot signal (run.py, before generation); seeding
generation with the checker critique is separately scoped and NOT done here.
The boundary is written into the generate_via_llm docstring + README + CLAUDE.
Load-bearing (maalbilde §7): tests/test_step5_refine_loadbearing.py is a PAIR —
the positive test keys the proposer's flip on the validator REASON PAYLOAD (the
rejected claim value, derived from validate_proposal(bad).reason so test and SUT
share one source of truth), and asserts the reason reached attempt 2's prompt
VERBATIM (the green-but-dead guard). It goes RED on detach (build messages once):
the flip token never arrives, so the outcome never flips AND the verbatim
assertion fails — proven double-red. The bounded control proves a never-fixed
proposer exhausts exactly max_attempts and returns a Rejection.
Adversarial Plan agent hardened the design pre-implementation (flip on payload
not wrapper/call-count; derive flip-key from the validator reason; drive through
generate_via_llm directly; docstring honesty). Suite 136->138 passed, 4 skipped;
mypy + ruff check clean. New test ruff-formatted; pre-existing ruff-format drift
(budget/verdicts/test_contracts) left untouched for a surgical diff.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHR8iKxJRxDiDfNw8HZmWE
