Bring the Python side to the ADR: the judge scores 0 to 10 with the cap at 4 and migration 33 clears every judgement, summarise reads the top 30 of the want and the match order, a hidden company's roles are never in the feed, and the company switch is called All locations.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
WolfandClaude Opus 5.5 committed 2026-10-04 21:40:32 +02:00
1 parent 6cad9713ac
commit 6d82287402
27 files changed
+390 -247

No files matched your search

+3 -2
View File
@@ -22,8 +22,9 @@ it filled none, `step` passes over all three. Extract reads the cut; the
score after it applies the fetch gate and the cut again with the columns
extract filled, so a role can drop out or enter (one that enters is
extracted on the next pull). Judge reads each agent's cut as recomputed,
against its CV. Summarise runs last, on the top of each agent's feed by
final score and the roles liked or saved: it is no part of any score.
against its CV. Summarise runs last, on the top of each agent's feed in
the want order and in the match order, and the roles liked or saved: it is
no part of any score.
A batch that failed, or a stage somebody killed, does not stop the chain: the
work comes round on the next pull, and a rescore of what did arrive is worth
+30 -21
View File
@@ -2,15 +2,17 @@
A `ModelTask` beside `summarise`, after the `/rank` of
github.com/MadsLorentzen/ai-job-search: the model reads a posting against a
written profile and scores skills, experience and career 0 to 100, with one
to three strengths and gaps. The total is `app.models.cv_score`, the three
numbers under the agent's `cv_*` weights, worked out when a role is served and
not stored, so the weights are no part of `input_key`; it is what the feed's
CV match sorts by. The model also reports a verdict on each of the
written profile and scores skills, experience and career 0 to 10, with one
to three strengths and gaps. A model's rating is an estimate, so the scale is
coarse. The total is `app.models.cv_score`, the three numbers under the
agent's `cv_*` weights times 10, so 0 to 100, worked out when a role is served
and not stored, so the weights are no part of `input_key`; it is what the
feed's CV match sorts by. The model also reports a verdict on each of the
agent's must-haves (one per line of its must text, numbered M1..Mn in the
prompt); it does not apply them. The rule is code's: a "broken" one makes
career count as at most `app.models.BROKEN_CAP`, when the role is served. Culture and location are left out: the first
is guessed from ad text, the second is what the fetch gate does.
career count as at most `app.models.BROKEN_CAP` (4), when the role is
served. Culture and location are left out: the first is guessed from ad text,
the second is what the fetch gate does.
The work is pairs, `(agent, role)`: an enabled agent with a CV and a prompt,
and an open role in its cut (`scores.cut`, see `app.jobs.score`: the best by
@@ -63,7 +65,8 @@ EXCERPT = 8000
#: 2: the verdict on each numbered must-have (`musts`), the cap in code.
#: 3: what the person likes (`like_text`) beside what they do not, and the
#: career rubric reading both.
FORMAT = 3
#: 4: the three numbers 0 to 10 rather than 0 to 100.
FORMAT = 4
#: How many strengths and gaps are kept, and how long each may be.
MOST, LONGEST = 3, 300
@@ -75,20 +78,20 @@ what they do not like (the work, the workplace, situations), and their
must-haves and deal-breakers when they gave any. Each item is one posting with a key, its title, company, place
and text. Answer one judgement per item.
Score three dimensions, each 0 to 100:
- skills: 80-100 the core requirements are the person's primary skills; 60-79
most requirements match, with 1-2 gaps that are learnable; 40-59 a partial
match, significant upskilling needed; 0-39 a fundamental mismatch.
Score three dimensions, each a whole number from 0 to 10:
- skills: 8-10 the core requirements are the person's primary skills; 6-7
most requirements match, with 1-2 gaps that are learnable; 4-5 a partial
match, significant upskilling needed; 0-3 a fundamental mismatch.
- experience: match on the function and nature of the work performed, not the
literal job title, and on its level. 80-100 direct experience in the same
domain and role type; 60-79 related, with transferable skills clear; 40-59
adjacent, the person would need to make the case; 0-39 unrelated.
literal job title, and on its level. 8-10 direct experience in the same
domain and role type; 6-7 related, with transferable skills clear; 4-5
adjacent, the person would need to make the case; 0-3 unrelated.
- career: against what the person wants (the prompts), what they like and
what they do not like. 80-100 strongly
aligned with their direction, with clear growth; 60-79 a good role, partly
aligned; 40-59 a decent job that does not build toward their goals; 0-39 a
dead end or a step backwards. Judge whether the tasks and the setting would
energise them, not only whether they could do them.
what they do not like. 8-10 strongly aligned with their direction, with
clear growth; 6-7 a good role, partly aligned; 4-5 a decent job that does
not build toward their goals; 0-3 a dead end or a step backwards. Judge
whether the tasks and the setting would energise them, not only whether
they could do them.
Their must-haves and deal-breakers are numbered M1, M2 and so on. Give a
verdict on every one of them, in `musts`, for each item:
@@ -269,10 +272,16 @@ def build_batch(config: Config, store: Store, limit: int = 5,
# -------------------------------------------------------------------- answer
#: The top of the judge's scale.
TOP = 10
def _mark(value: Any) -> int | None:
"""One of the three numbers, rounded and held to 0..`TOP`; None for
anything not a number."""
if isinstance(value, bool) or not isinstance(value, int | float):
return None
return max(0, min(100, round(value)))
return max(0, min(TOP, round(value)))
def _lines(value: Any) -> list[str]:
+9 -9
View File
@@ -10,7 +10,7 @@ on. Computed once per pull (`plan`) and handed to each board's batch as the
empty, or holds any other member (a function, a remote value other than
`remote`, a place the gazetteer does not know), can take a role from
anywhere, so the base reach is then everywhere.
- **Wide reach, for the companies set to All countries**
- **Wide reach, for the companies set to All locations**
(`companies.all_countries`, global): the union over the enabled agents of
their `require` reach, the intersection of each one's `require` gates on
country and place, Europe today, expanded to its countries through the
@@ -272,7 +272,7 @@ class Plan:
def why(self, sub: Subscription) -> tuple[Reach, str]:
"""The reach `sub` is read with, and why: `override`, `small`,
`unknown` (a board never read whole), `countries` (its company is
set to All countries) or `agents`."""
set to All locations) or `agents`."""
if OVERRIDE_KEYS & set(sub.params or {}):
return EVERYWHERE, "override"
size = self.sizes.get(sub.target)
@@ -296,7 +296,7 @@ class Plan:
def for_subscription(self, sub: Subscription) -> dict[str, Any] | None:
"""The `reach` a board's batch hands `ask_for`; None for none. With
fetch terms, and unless its company is set to All roles, a big board
carries them, and `wide` when its company is set to All countries, for
carries them, and `wide` when its company is set to All locations, for
`ask_for` to decide which side gives way over `MAX_SERIES` (see the
module docstring). A board with no size carries
`unsized` and `small`, and the reach of a big board unless its params
@@ -360,7 +360,7 @@ def plan(store: Store, subs: Iterable[Subscription] | None = None) -> Plan:
base = wide = EVERYWHERE
subs = list(store.subscriptions(source="ats") if subs is None else subs)
rows = store.companies()
# All countries never reads less than the board would without it: the
# All locations never reads less than the board would without it: the
# base's remote, which a `require` of Europe does not name, is kept.
return Plan(base, wide | base, sizes(store, subs),
all_countries=frozenset(c.key for c in rows if c.all_countries),
@@ -380,7 +380,7 @@ def reaches(store: Store, subs: Iterable[Subscription]) -> Mapping[str, dict[str
#: Why a board is read whole, as the Companies list says it.
_WHOLE = {"small": f"under {SMALL} roles", "unknown": "size not known yet",
"override": "set here", "countries": "all countries (everywhere)",
"override": "set here", "countries": "all locations (everywhere)",
"agents": "an agent takes any place"}
@@ -452,8 +452,8 @@ def line(found: Plan, sub: Subscription) -> str:
"follows agents: DK, SE, remote · follows agents' terms: engineer,
product", "follows agents: DK, SE, remote · all roles" (a company set to
All roles, or a reader without a query, or terms over `MAX_SERIES`),
"all countries (Europe) · follows agents' terms: engineer, product,
data", "all countries (Europe) · all roles", "set here: DE · all roles",
"all locations (Europe) · follows agents' terms: engineer, product,
data", "all locations (Europe) · all roles", "set here: DE · all roles",
"all places, too many countries to ask · follows agents' terms:
engineer", or, for a board read whole, why: "under 300 roles: read
whole", "Greenhouse cannot filter: read whole". A board with no size yet
@@ -485,13 +485,13 @@ def line(found: Plan, sub: Subscription) -> str:
else "")
if why == "countries":
named = found.wide_names or f"{len(found.wide.countries)} countries"
return f"all countries ({named}){unasked} · {roles}"
return f"all locations ({named}){unasked} · {roles}"
if why == "override":
return f"set here: {_places(ask)}{unasked} · {roles}"
return f"follows agents: {_places(ask)}{unasked} · {roles}"
if ask.terms:
if given.everywhere:
what = "all countries (everywhere)" if why == "countries" else _WHOLE[why]
what = "all locations (everywhere)" if why == "countries" else _WHOLE[why]
return f"{what} · {roles}"
return f"all places, too many countries to ask · {roles}"
if why in ("agents", "countries") and given.everywhere:
+4 -4
View File
@@ -10,7 +10,7 @@ good for a given search. Each agent first applies its fetch gate:
any gates one group, the place part: a role that matches none of
them, and has the column filled for at least one, is
gated ("none of country DK / remote remote"); skipped for
a company set to All countries
a company set to All locations
exclude gates a role that matches one of the values keeps its row, with
score 0 and the value ("title account executive"); those
on `title` are skipped for a company set to All roles
@@ -42,8 +42,8 @@ all there is, when it is gated as "no vector". An agent with no votes scores
every role without `taste`; the direction is newsdesk's
`centroid(liked) - centroid(disliked)`. An agent with no prompt that has
text has no `fit` either, so nothing of it enters the cut by fit and its
scores do not decide what summarise reads (`Store._summarise_top`); a run
names such agents in one notice.
want order does not decide what summarise reads (`Store._summarise_top`); a
run names such agents in one notice.
Beside the score, `fit` is stored as it is, and for an agent with a CV,
`cv_fit`: how close the role is to the CV, the best cosine between the role
@@ -181,7 +181,7 @@ def failed_gates(candidate: Candidate, gates: Sequence[Gate]) -> list[str]:
it fails, every exclude gate it matches, and the `any` group when it
matches none of them. A gate with an empty column (state -1) neither
passes nor fails, so the group fails only when some member had evidence.
A company set to All countries skips the `any` group, the place part,
A company set to All locations skips the `any` group, the place part,
and one set to All roles the exclude gates on `title`, the titles to
leave out; the require gates, the region, apply to every company."""
states = list(zip(gates, candidate.gates, strict=True))
+4 -2
View File
@@ -7,8 +7,10 @@ answer), and running it is `desk.modeltask`'s.
A role is waiting while it has no summary, has some text, is open and no copy,
and is either in the top of an enabled agent's feed, the best `summarise_top`
final scores among the roles its fetch gate passes (its `limits`), or one the
owner cares about whatever its rank: saved (it has an application, the star)
(its `limits`, 30 by default) by the want feed score and the best
`summarise_top` by the match order (CV match, then CV closeness) among the
roles its fetch gate passes, so at most twice that many, or one the owner
cares about whatever its rank: saved (it has an application, the star)
or voted up by an enabled agent. Summarise is no part of any score and runs
last in the pull, once the final scores are set. The limit keeps the cost
flat: a pull summarises the roles somebody will look at, not the 30k that are