Market

scavenger.observe_market@1 admits finite source snapshots, normalizes and deduplicates listings, and retains retrieval time and source claims. Scheduled observations coalesce instead of building an unbounded backlog.

Each Scout profile fixes permitted origins and paths, user agent, pacing, concurrency, byte and page caps, expiry, robots decision, selectors, session needs, CAPTCHA handling, fixtures, and a kill switch. A bazos.sk adapter can observe a listing; it cannot choose the subject, pass a hard gate, bargain by default, or own the campaign.

Evidence before preference

scavenger.qualify_candidate@1 applies profile hard gates first: false rejects, true admits, and unknown remains visible. A model cannot invent an address, fee, area, socket, clearance, firmware, condition, serial, stock state, or seller fact.

scavenger.rank_candidates@1 then orders only eligible candidates from declared weights and dated evidence:

known utility = Σ(known weight × normalized preference)
lower bound   = known utility / total configured weight
upper bound   = (known utility + unknown weight) / total configured weight

The whole candidate field is the review surface

The ordinary market view can teach preference without asking the Magus to leave the work and fill an abstract training form. The whole normalized snapshot remains inspectable. Hard-false candidates stay in an attributable rejected lane, unknowns remain visible, and eligible candidates receive one proposed order.

The Magus may then:

• move a candidate higher or lower;
• assign a score and an application-specific state such as contact, watch, ask, or reject;
• choose a structured reason and add a short explanation;
• mark an exact region in a listing image, state what is observable there, and distinguish evidence from suspicion; and
• revise the judgment when a seller answer or later outcome changes the evidence.

Each explicit action appends a CandidateJudgment@1 rather than mutating chat history. A completed RankingReview@1 pins the source snapshot, the system's initial order, the reviewed order, the judged subset, unjudged candidates, and all judgment references. Scavenger may derive a PreferenceRevision@1 containing scoped weights, rules, examples, exceptions, provenance, and confidence. That revision is Scavenger-owned decision memory, not general Archive memory or a silent model-weight update. The next ranking pins the exact accepted preference revision it used and explains which evidence and preference affected each score.

Repeated review therefore stays reversible. A later revision may improve the proposed order, but the full candidate field remains available for inspection and correction every time.

Bazaar worked example: a daily vehicle market

A seven-day automotive Bazaar campaign can observe one finite Bazoš search each morning. The new snapshot records newly listed, removed, relisted, repriced, reworded, and rephotographed vehicles against the previous day. The diagnostic view shows the complete field rather than only a shortlist.

The Magus reorders vehicles, scores them, assigns contact, watch, ask, or reject, and records the deciding reason. On a photograph, the Magus can mark the exact region that deserves attention and explain what should be checked. Missing VIN, service history, condition detail, photograph, or video can become an approved question through Bargain. A reply updates attributed evidence and may change qualification, score, state, and order; it never becomes permission to buy or negotiate outside the campaign envelope.

The next morning Scavenger proposes a new full-list order from the accepted preference revision. The person can still move any vehicle, correct the inferred reason, or leave it unjudged. The same review contract applies to technology, property, equipment, collectibles, and other finite markets; the subject profile supplies the domain criteria.

Images, messages, and learned contact style

An image judgment keeps source identity, digest, exact region, observed cue, interpretation, confidence, and correction. A model-highlighted region is a proposal, not proof of damage, condition, identity, or value. If source and egress policy permit it, a public listing image may be sent to a named remote vision provider for bounded inference with provider and model provenance. Public availability does not make the image model-training material or grant a right to republish it.

Scavenger may also compare which already approved questions and truthful automation disclosures produce decision-useful answers. Bargain still owns every send, limit, stop signal, unknown acknowledgement, and seller thread. Phone numbers, private replies, credentials, and private anchors never enter a preference revision as ambient context.

Prove that preference learning earns its cost

The first snapshot establishes an unaided ranking baseline. Intermediate snapshots collect human corrections. A final held-out snapshot measures whether the accepted preference revision improves:

• agreement between the proposed and reviewed order;
• the number and size of manual moves and score corrections;
• agreement on stated reasons and image regions, including calibrated uncertainty;
• the share of seller answers that materially change a candidate judgment; and
• human review time compared with the baseline.

A representative seven-day pilot can target at least 50 explicit candidate judgments and 30 to 60 image-region corrections, while each campaign declares its own evidence minimum. Continue only when held-out ordering or decision time improves without hiding unknowns or increasing annotation cost beyond the saved work. Otherwise retain the evidence, reject the preference revision, and change or stop the pilot.

Price remains an attributable interval across condition, warranty, age, bundle value, fees, and uncertainty. A route observation proves access at one time, not title, structural condition, financing, insurance, future development, or final transaction cost.

School, work, family, health, routine, exact address, and phone data receive least-data handling. Sources see only the coarse region, opaque anchor, or derived observation required by the admitted step. Listings and seller claims are evidence, not preference memory or training material by default. Only admitted human judgments and accepted application revisions enter decision memory; changing provider model weights is a separate, explicit Training workflow outside this Pattern.

Continue with Bargain or end with an evidence-bound shortlist.