o 1jz@s,dZdZdZdZdjeedZdeZdS)aoThe version stamp that says which pricing logic wrote a row's prices. WHY THIS EXISTS --------------- On 2026-09-16, answering "which rows still carry pre-fix prices?", every timestamp and marker column in the deals table was checked and **not one of them dates a row's pricing**: * `last_seen_utc` and `source` are rewritten by ALL THREE paths - heavy (`smart_ingestor.py:580-581`), light (`:574-575`) and Stale Rescue (`:274-275`). `source` records who touched a row LAST, not who priced it, so a row priced by the heavy path on 09-10 reads `stale_rescue` today. * `Deal_found` is Keepa's `creationDate` (`stable_deals.py:83`), not a write time. * `last_update` is deliberately never populated (AGENTS.md 7.3). * `Inferred_Sale_Count` looks like it should work and does not. ASIN 0415009804 carries a count AND a pre-fix price: rows priced between 2026-09-11, when that column shipped, and 2026-09-12 16:50 UTC, when the price-association fix went live, have both. Confirmed against production. So the version stamp is the only way to tell a current price from a stale one. THE NULL RULE, AND WHY IT IS THE OPPOSITE OF THE Inferred_Sale_Count RULE ------------------------------------------------------------------------ **A row whose `Pricing_Logic_Version` is NULL, or lower than `PRICING_LOGIC_VERSION`, carries STALE PRICING and is due a heavy re-fetch.** NULL means "priced by unknown logic", which for scheduling purposes is the same answer as "priced by old logic". That is deliberately the reverse of the `Inferred_Sale_Count` rule in `Documentation/INFERRED_PRICE_LOGIC.md` 4.C, which states that a NULL count means "never computed" and **must never be read as zero and must never be used to hide a deal**. The two rules do not conflict, because they answer different questions: Inferred_Sale_Count -> may this deal be SHOWN to a subscriber? Pricing_Logic_Version -> does this row need WORK scheduled against it? Being unsure about the first must never hide a deal, because hiding a deal on a guess costs the subscriber a real opportunity. Being unsure about the second costs about 7 Keepa tokens and re-prices a row that may not have needed it. The asymmetry in the cost of being wrong is the whole reason the rules differ. Do not "harmonise" them. NO BACKFILL ----------- Every row that exists today lacks this column, and none of them may be given a value except by actually recomputing their prices. A backfill would have to guess which logic wrote a row, and guessing is precisely what this column exists to stop. BUMPING IT ---------- Raise `PRICING_LOGIC_VERSION` by one whenever a change alters the prices the pricing pipeline produces - `infer_sale_events`, `analyze_sales_performance`, `get_1yr_avg_sale_price` or anything they call. `repair_pricing.py` selects on `version IS NULL OR version < PRICING_LOGIC_VERSION`, so a bump is the entire mechanism for scheduling the next repair sweep: no new script, no new predicate. Do NOT bump it for a change that leaves prices identical (a refactor, a log line, a docs edit). A spurious bump schedules a full re-fetch of every row in the database. A bump also has an IMMEDIATE, VISIBLE effect as of September 2026: Prime Picks reads `CURRENT_PRICING_PREDICATE` below, so the moment a bump deploys, every row in the table is stale and Agent's Choice empties out - on the next run, and for read-time display straight away. It refills as the sweep repairs rows. That is the intended behaviour (an empty Agent's Choice is better than a recommended price the system knows is superseded), but it is not a subtle change, so expect it rather than diagnosing it. The main dashboard grid is NOT affected. zPricing Logic VersionPricing_Logic_Versionz(("{col}" IS NULL OR "{col}" < {version}))colversionzNOT {}N)__doc__PRICING_LOGIC_VERSIONPRICING_VERSION_HEADERPRICING_VERSION_COLUMNformatSTALE_PRICING_PREDICATECURRENT_PRICING_PREDICATEr r 6/var/www/agentarbitrage/keepa_deals/pricing_version.pysV