Skip to content

CHANGELOG.md

Latest commit

 

History

History
183 lines (163 loc) · 10.5 KB

CHANGELOG.md

File metadata and controls

183 lines (163 loc) · 10.5 KB
Raw

Changelog

Releases are tagged on GitHub and archived on Zenodo (concept DOI 10.5281/zenodo.21804210). The second half of this file holds the operational history that used to live as comments beside the constants it explains — kept because an annotation on the archive can only be read back with it, moved here so the code states rules and the changelog tells stories.

1.1.0 — 2026-09-02

Pre-publication release accompanying When AI Meets the Archive.

Provenance

  • Every dcterms:subject / dcterms:spatial link written by NER or reference indexing carries an iwac:nerModel value annotation (property 314) naming the model that proposed it. Links already on an item are never re-stamped. The write steps read the model from the checkpoint the AI step left beside its CSV; --model overrides, and they ask when neither applies.
  • AI_ocr_extraction/02_mistral_ocr_processor.py runs the pinned mistral-ocr-4-1 through the shared common/mistral_ocr.py client instead of a copy of it addressed to the rolling mistral-ocr-latest alias. A test forbids the alias in any pipeline script.
  • AI_summary_issue/03 asks which model consolidated the index rather than stamping the registry default silently.

Python only

  • The .claude/ agents and skills are gone. Reference-indexing step 2 is AI_reference_indexing/02_enrich_references.py (structured output through the shared provider, resumable, keyword summary); the magazine pipeline is its Gemini and Mistral scripts.

Write safety

  • Every step that writes to Omeka now shares one gate (common/write_guard.py): --dry-run, --yes, --backup-dir, --no-backup, the same confirmation panel, and a pre-write dump of every item before its PATCH. Three text write steps and the sentiment panel had no backup at all; the magazine TOC step also had no --yes and ignored failures.
  • AI_publication_extraction/04's "unchanged" check compared against Omeka's echoed value objects and never matched, so every rerun re-PATCHed.

Shared code

  • OmekaClient.get_items() takes any /api/items filter (class, template, modified_after …), follows Omeka-S-Total-Results, and warns if the count moves during the walk; iter_items() streams, count_items() reads the header, get_items_by_ids() batch-fetches, list_page() serves samplers, and the client remembers Omeka-S-Version. Five hand-rolled paging loops are gone. search_items_by_property() raises on a transport failure instead of answering "no match". Session backoff is 1.0. upsert_property_value() is deprecated.
  • common/link_update_cli.py and common/reconciliation_cli.py hold the write and reconciliation runs that NER and reference indexing had each copied; the four scripts are entry points. The two PDF downloaders are one run_cli() with --item-set-id.
  • common/reconciliation.py prepares each string once and gates the expensive SequenceMatcher.ratio() behind its cheap upper bounds.
  • common/llm_provider.py records tokens (and cost where the provider states it) on every client as client.usage; LLMConfig.service_tier reaches OpenAI (AI_summary/02 --service-tier flex); OPENROUTER_ZDR=1 adds OpenRouter's zero-retention routing.
  • OCR correction no longer picks a thinking level by model name.

Interfaces

  • --prompt N on the audio and video transcribers, --no-split on the audio one, --item-set-id / --item-id on the three downloaders that only asked; the audio downloader stops producing filenames the transcriber then refuses. Every main() returns an exit code. One repo-root idiom.
  • Input folders (PDF/, Audio/, video/, TXT/, ALTO/) ship empty with a .gitkeep; output, cache, backup and serving/work/ are git-ignored.

Documentation

  • Around thirty README statements corrected against the code, including flags argparse rejected, a wrong expansion of IWAC, a sentiment annotation that is not written, and the audio pipeline's provenance claim.

1.0.0 — 2026-08-05

First archived release.

Operational history

What the constants used to explain in place. Dates are when the archive changed.

AI model authority items (common/iwac_config.AI_MODEL_ITEMS)

  • 2026-07-27 / 07-31 — Item 79608 "Gemini 3.6 flash" was a duplicate of 79611 and item 79609 "GPT-5.6 Luna" was deleted upstream and replaced by 79610; both are in RETIRED_AI_MODEL_ITEM_IDS and must not be annotated with.
  • 2026-07-31 — The Gemini key was gemini-flash, the registry key of the rolling gemini-flash-latest, which reports its version as literally "Gemini Flash Latest"; annotating such a run as "Gemini 3.6 flash" asserted a version the run could not confirm. Re-keyed to the pinned gemini-3.6-flash.
  • 2026-08 — gemini-flash-lite was the same bug and outlived the fix: the key of the rolling gemini-flash-lite-latest claiming item 78631 "Gemini 3.1 flash lite", which resolved to 3.5 once 3.5 Flash-Lite shipped. Both Flash-Lite generations now have pinned keys and their own items; the rolling key is deliberately absent. gemini-pro was the third case, found by the guard added for the second (test_no_registry_key_is_a_rolling_alias), and is re-keyed to gemini-3.1-pro; the rolling gemini-pro stays in the registry for pipelines that stamp nothing.
  • 2026-08-07 — claude-opus split into claude-opus-4.6 (item 78528) and claude-opus-5 (79615), so an operator asserting which Claude read a magazine could not name last year's release by accident. Since 2026-09-02 no pipeline uses a Claude agent; the keys stay so the annotations resolve.
  • 2026-08-14 — Three display_title values corrected to the live item titles ("Gemini 3.1 Pro", "Gemini 3.6 Flash", "Gemini 3.1 Flash Lite"). Omeka regenerates the key on read, so nothing stored was wrong; it is what the confirmation panel shows.
  • 2026-08-18 — mistral-ocr-4-1 (item 111889) created for AI_publication_extraction; the 425 bibo:content values already on the reference corpus carry no provenance annotation.
  • 2026-08-25 — qwen3.8-27b-selfhosted (item 111933): keyed on the self-hosted route because that is what produced the annotations; the OpenRouter twin has written nothing and would need its own item.
  • 2026-08-27 — gemini-3.5-transcribe (item 113077) for AI_audio_summary/02c; the audio pipeline stamped no transcription provenance before it.

Model registry (common/llm_registry.py)

  • Gemma 4 on two routes. gemma-4 (Gemini API) is free of charge because it is served on the free tier, and Google's pricing page states that free-tier content is used to improve its products — the thing data_collection: "deny" exists to prevent when whole archival articles are sent. gemma-4-openrouter is the key for anything that sends archive text; gemma-4 stays for multimodal document work the chat API cannot do. Gemma has two thinking levels, MINIMAL and HIGH; a request for "medium" lands on high.
  • Qwen3.8 27B on two routes. Same weights, and the route is half of what a provenance record claims. Both carry temperature 1.0, Qwen's thinking-mode recipe (the model card's 0.7 is the non-thinking one). The self-hosted default is low, not the vendor's xhigh: an unconfigured bulk run reasoning as hard as it can on shared university GPUs is the accident to design against. The hosted twin (qwen3.8-27b-openrouter) is in no tier and exists to measure one route against the other.
  • DeepSeek V4 Flash. The dated deepseek-v4-flash-0731 is the shared text default; the April preview deepseek-v4-flash is in no tier and survives only so annotations it wrote still resolve.
  • 2026-08-14 — gemini-flash-latest rolled onto Gemini 3.7 Flash, which dropped the MINIMAL thinking level, and every pipeline that hardcoded minimal began failing with a 400. supported_thinking_levels and clamp_thinking_level() date from that day.

Sentiment panel (AI_sentiment_analysis/sentiment_core.PANEL)

  • 2026-07-31 — Generation 2 goes live: six properties per member named for the model, never the vendor, because vendor-keyed generation 1 could not be attributed without git archaeology. The Gemini slot moved off gemini-3.6-flash the same day: at $1.50/$7.50 per 1M it cost five to seventeen times the others, so any disagreement read as "the expensive model knows better". Nothing was ever written to iwac:gemini36Flash*.
  • 2026-08-05 — Qwen3.5 122B-A10B dropped before annotating anything: with require_parameters OpenRouter left four endpoints and a median call of 104 s against 4–6 s for the rest. iwac:qwen35A10b* holds zero values.
  • 2026-08-07 — Generation 1 (iwac:gemini*, iwac:chatgpt*, iwac:mistral*, 12,286 items each) and the April DeepSeek preview (11,482) deleted from Omeka after confirmation on the Hugging Face full mirror. The Hub keeps generation 1 frozen with omeka_prefix=None; that freeze is the archive.
  • 2026-08-14 — The Google slot became Gemma 4 31B, replacing gemini-3.5-flash-lite, which had held it since 07-31 and annotated nothing (verified at 0 items). Gemma replaces rather than joins Gemini: same lab and pretraining family, so running both buys correlated annotator error. It is routed through OpenRouter (see the registry note); the Gemini route is also capped at 16,000 input tokens per minute for this model. OpenRouter lists 19 endpoints, 16 with structured outputs, so it does not repeat Qwen3.5's starvation. Its graduated thinking levels collapse to on/off across third-party backends: measured 2026-08-14, medium and high are indistinguishable in latency and reasoning length.
  • 2026-08-25 — Qwen3.8 27B joined as a fifth voice, annotated on university hardware: the first member whose requested medium is a rung the model has. Its coverage is 12,098 of 12,251 and stays there: 153 articles were attempted four times and retired, 145 failing the schema's cross-field rule because the model declines subjectivité when Islam is peripheral where the prompt licenses declining only when it is absent. Recorded, not repaired; its missing subjectivité is not missing at random.
  • Reasoning depth, verified 2026-07-29/31 and 08-14 — GPT-5.6 Luna accepts none/low/medium/high/xhigh/max; Gemma 4 only MINIMAL/HIGH; DeepSeek 0731 only low/high/max; Mistral Small 4 only none/high. Only Luna and Qwen3.8 sit at a genuine middle; the others are rounded up to high, which any write-up of the panel must state.