Market

Each morning brings a changed market: new, removed, relisted, repriced, reworded, and rephotographed candidates. The useful view shows the whole field, including the reasons some listings cannot qualify. A shortlist alone would hide the evidence needed to correct it.

scavenger.observe_market@2 admits finite snapshots, normalizes and deduplicates listings, and preserves retrieval time and source claims. Scheduled observations coalesce rather than accumulate an endless backlog. The Scout profile fixes origins/paths, user agent, pacing, concurrency, bytes/pages, expiry, robots decision, selectors, sessions, CAPTCHA handling, fixtures, and kill switch. A bazos.sk adapter observes a source; the campaign decides what it means.

Finite feed observation

Scavenger revision 3 selects scavenger.observe_market@2 with the explicit bazos-pc-title-link@1 profile. It accepts one complete immutable CampaignSeed@1 and one exact public feed declaration. The seed pins subject/profile revision, concrete need, inventory, hard criteria, weighted preferences, budget, deadline, region, evidence standard, travel/delivery semantics, source/disclosure policy, learning mode and a Watch autonomy ceiling. Admission retains the seed before held Run publication; it does not execute plan_campaign.

This profile permits one bounded RSS Fetch/Extract and a local title/link projection. It admits no model call, qualification, ranking, item acquisition, seller contact, purchase, polling or preference learning. The source profile fixes the exact advertised Bazoš PC feed, the native RSS media/extraction profile and finite effect ceilings. A public declaration remains an authenticated operator assertion, not a technical proof of publicness or a reusable contact grant.

Initial seed encoding

CampaignSeed@1 requires every fact explicitly, including empty inventory, hard-criteria and preference arrays. subject_profile is Bazaar@1 or Technology@1; need is nonblank bounded text. Inventory, hard criteria and preferences each permit at most 16 entries. A preference carries nonblank criterion text and a nonnegative integer weight; this profile does not normalize, rank or learn from it. budget_minor is a nonnegative integer number of EUR cents or explicit null (not assessed), with currency: EUR. region is bounded nonblank text or explicit null; neither unknown establishes eligibility. deadline carries a timezone; admission checks its current validity without making historical seed decoding depend on wall time.

evidence_standard and travel_delivery are required nonblank descriptions, never grants. source_policy: admitted-public-feed-only@1, disclosure_policy: none, learning_mode: none and operating_mode: Watch close this initial profile. Need text permits 2048 UTF-8 bytes; individual inventory/criterion/preference/evidence/travel values permit 1024 bytes and region 256. The complete 16 KiB bound still applies. No complete seed is copied into Run or consent facts.

Whole field and local views

Scavenger owns immutable MarketObservation@1 records. The exact subject/request binds the canonical seed digest, Run and selected Composition/Pattern/manifest, source and acquisition profiles, purpose, retention profile and initial filter revision. Exact replay resolves the same admission/result; changed bindings refuse. The first profile permits one pending or retained observation per selected profile database. It does not create another Scout allowance store.

The whole field contains every item returned by the admitted bounded extraction, in source order, including unmatched, duplicate and unjudged occurrences. Each occurrence retains an internal identity, source ordinal, duplicate grouping, optional original plain title, inert original item link and source-record identity. Deduplication groups exact item locators without erasing their occurrences or query/path spelling. Missing values and unresolved facts remain visible; no eligibility or seller judgment is inferred. This field is the finite response, not the source's entire market.

Retain at most 100 observations, with titles at most 1024 UTF-8 bytes and links at most 4096 bytes, matching the qualified RSS mechanism. The complete seed is at most 16 KiB and the canonical application observation at most 256 KiB. Overflow refuses the complete result rather than hiding items or shortening fields. Preserve requested/contacted locator, retrieval time, Scout receipt, source/record/field digests, selected profiles, omissions, duplicate/missing counts and declared coverage. Source dates/GUIDs do not establish identity or freshness.

Only the admitted title/link field and necessary attribution survive in application custody. Feed bodies, channel text, summaries, images, publisher dates/identifiers and separate contact fields are discarded after transient consumption. Titles and links are deliberately retained source claims; the profile does not promise that source text contains no personal information. Run/Graph state, generic events/logs and Bridge history carry opaque references, safe counts/status and closed reasons, never item values or a copied campaign seed. Scout retains its independently owned contact attribution and positively selected extraction metadata, not this field.

An authorized read returns the complete field. A bounded literal, versioned case-folded substring filter over retained title/link values derives a local view of that same field digest. It is not regex, semantic ranking or qualification. The view reports total, visible, unmatched, duplicate and missing counts and keeps an all-observations view available. A new filter changes no field and performs no acquisition. Read and re-filter require current application retrieval authority; a receipt or opaque UUID supplies neither.

Retention, deletion and recovery

Retain this one field until its owner explicitly deletes it. Deletion is a distinct current-authority operation over the exact observation identity/digest, available only after acquisition and result disposition are proved terminal. Pending work must be cancelled through the Run owner first; unknown acquisition cannot be declared settled by deletion. The first profile supplies no automatic unknown-attempt reconciliation.

Deletion atomically removes the application seed, item payload and content projections. It keeps only a content-free tombstone and necessary immutable execution/Scout receipts; contact metadata in those receipts is not erased. A tombstone fences late writes and exact replay, so neither can recreate payload. Deleted is distinct from an empty market. Backup erasure is not implied. Reopening the application slot does not reset spent counters, unknown reservations or an issued claim. Replacing an observation is a visible delete followed by a fresh request and consent.

Before acquisition claim, the exact retained consent may resume under current authority and expiry. After claim or uncertain acknowledgement, lost transient material cannot be reconstructed or refetched. Commit the entire consumer field once, atomically under its request/claim identity, after current CONSUME and result-adoption checks. Settlement without a confirmed consumer commit returns material unavailable unless the same key proves a committed result. Lost commit acknowledgement reconciles that same key/digest; ambiguity stays unknown and creates no acquisition. A committed field supports fresh authorized reads/re-filters after restart; deletion remains final for that identity.

The first native profile uses one active Scout binding in a fresh profile database. Existing page budget rows refuse a page/RSS switch or coexistence without mutation. This is an explicit initial compatibility boundary, not a reset, migration or second budget store.

The Altar presents source/time/coverage, omissions, total/filtered counts, complete title/link rows, unknown/refused/deleted states and distinct new-fetch, local-filter and terminal-delete actions. Item links remain displayed data, with no automatic preview/contact. The native source and State must prove these operations before the interface can claim they are available.

Evidence before preference

scavenger.qualify_candidate@1 applies hard gates: false rejects, true admits, and unknown stays visible. No model may invent an address, fee, area, socket, clearance, firmware, condition, serial, stock, or seller fact. Only eligible candidates enter the proposed order from scavenger.rank_candidates@1:

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

These bounds assume finite, nonnegative configured weights with a positive total, and preference values normalized to [0, 1]. They apply only when the selected profile establishes those conditions. Here, unknown weight is the configured preference weight whose candidate value is still unknown. The lower bound counts only known utility; the upper bound shows the contribution those unresolved preferences could still make. Neither bound changes hard eligibility.

Price itself 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, development, or final transaction cost.

The whole candidate field is the review surface

The Magus can move a candidate, score it, assign contact, watch, ask, or reject, choose a reason, and explain a judgment. A marked image region keeps observable evidence separate from suspicion. Later seller answers or outcomes can revise the judgment.

Each action appends CandidateJudgment@1. RankingReview@1 seals the snapshot, original system order, reviewed order, judged subset, unjudged candidates, and judgment references. Hard-false candidates remain in a rejected lane; unknowns and unjudged material remain inspectable.

A derived PreferenceRevision@1 may contain scoped weights, rules, examples, exceptions, provenance, and confidence. It is Scavenger decision memory, separate from Archive and model weights. The next ranking pins the accepted revision and explains which evidence and preferences affected each score. The person can still correct any candidate or inferred reason.

Bazaar worked example: a daily vehicle market

A seven-day automotive campaign observes one finite Bazoš search each morning. The diagnostic view retains all changes from yesterday. The Magus reorders vehicles, scores them, marks image regions worth checking, and chooses what to contact, watch, ask, or reject.

A missing VIN, service history, condition detail, photo, or video can become an approved Bargain question. Its reply updates attributed evidence and may change qualification and order. It cannot authorize purchase or widen negotiation. Tomorrow's order uses the accepted preference revision; today's rejected and unjudged cars remain available for inspection. Technology, property, equipment, collectibles, and other finite markets use the same review contract with their own subject criteria.

Images, messages, and learned contact style

An image judgment binds source identity/digest, region, observed cue, interpretation, confidence, and correction. A model highlight proposes a place to inspect; it proves no damage, identity, condition, or value. Eligible public listing images may cross an exact source/egress policy to a named remote vision provider for bounded inference with provider/model provenance. Public availability grants neither training nor republication rights.

Already approved questions and truthful automation disclosures can be compared for decision-useful answers. Bargain still owns every send, limit, stop signal, uncertain 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

Take the first snapshot as an unaided baseline, collect intermediate corrections, and test the accepted revision on a held-out final snapshot. Measure order agreement, manual move/score correction effort, reason and image-region agreement with calibrated uncertainty, seller answers that change judgments, and human review time.

A representative seven-day pilot may target at least 50 explicit candidate judgments and 30 to 60 image-region corrections; the campaign declares its own minimum. Continue only when held-out ordering or decision time improves without hiding unknowns or spending more annotation effort than it saves. Otherwise retain the evidence, reject the revision, and change or stop the pilot.

School, work, family, health, routine, exact address, and phone data receive least-data treatment. Sources receive only the needed coarse region, opaque anchor, or derived observation. Listings and replies remain evidence; only admitted human judgments and accepted application revisions enter decision memory. Model-weight changes require a separate Training admission and promotion.

Continue with Bargain, or finish with the evidence-bound shortlist.