JDs and the IAC profile¶

Paste a URL or text; the platform extracts an axis-weighted skill profile + an iterator/amplifier/connector taxonomy + a 2-sentence summary. Use this to drive readiness, scheduling, and reflect context.

What gets parsed¶

app/services/jds.py extracts three things via Claude (with a keyword fallback when no API key is available):

Field	Type	Drives
`axis_weights`	`dict[axis, float]` (10 axes, each 0–1)	Scheduler bias + readiness math
`signal_tags`	`list[str]` from a curated vocabulary	Problem matching via `problems.jd_signal_tags`
`iac_profile`	`{iterator, amplifier, connector}` floats, ~sum to 1.0	Fit alignment on the JD detail page
`iac_evidence_md`	Short markdown citing the JD language	Shown alongside the IAC scores
`summary`	2-sentence English	JD detail page header

The iterator/amplifier/connector taxonomy¶

From the LessWrong piece Talent Needs of Technical AI Safety Teams (Larsen/Hubinger et al.). Three archetypes:

Iterator — pushes an existing research agenda forward. Strong at: executing experiments, debugging models, shipping incremental wins. Signals in a JD: "run experiments", "iterate on training runs", "debug model behavior", "ship results".
Amplifier — multiplies others' research output. Strong at: tooling, infra, evals, internal codebases, knowledge management. Signals: "build internal tooling", "eval infrastructure", "developer experience for researchers".
Connector — synthesizes across disciplines / people. Strong at: setting new agendas, bridging fields, distillation. Signals: "propose new research directions", "cross-team collaboration", "external communication".

Most roles want a blend. The three IAC scores roughly sum to 1.0 — they represent role mix, not independent traits. If the JD is silent on this, the parser defaults to {0.5, 0.3, 0.2} (iterator-leaning research engineering).

Your own IAC profile¶

When you upload a CV via /credentials, the same iac_profile shape gets extracted and stored on users.settings_json.iac_profile. The JD detail page then renders a side-by-side "this role wants / you bring" block plus a 1-line fit interpretation from a small Claude call (cached on the JD row in jds.iac_fit_md).

If you don't have a CV uploaded, the alignment block prompts you to add one.

Active JD¶

Only one JD is active at a time. Setting one active:

POST /jds/<id>/activate

is what the Activate button on the JD detail page does. The active id lives at users.settings_json.active_jd_id. The scheduler reads it, readiness reads it, and the home page's Today header references it.

Adding a JD¶

Three ways:

Paste URL: /jds → paste a job-board link. The server fetches the page, strips it to text, extracts metadata (og:title, organization), runs Claude. If the fetch fails (auth-walled boards, SPA-rendered pages), the URL-slug fallback kicks in to derive a clean title.
Paste text: /jds → paste the full JD body. Same pipeline minus the fetch.
Manual title + text: optional override fields on the form.

Where the data goes¶

jds table: axis_weights (JSONB), signal_tags (JSONB), summary, iac_profile + iac_evidence_md (JSONB), iac_fit_md (cache).
The JD parsed_json drives scheduler._score() (in app/services/scheduler.py) when a JD is active — adds a small bonus for problems whose domain_tags and jd_signal_tags overlap with the JD.

Title rendering¶

Stored titles that look like URLs (legacy data) are healed at render time by the jd_title Jinja filter. The filter detects a URL-shaped title and derives a clean form via jds.title_from_url(). Applied on home.html, jd_detail.html, jds_index.html, readiness.html.