Indlela Yokusebenzisa i-Project Lighthouse Ukuhlaziya Imibhalo Yesayensi Yasendulo nge-AI

Okwenziwa yi-Project Lighthouse

I-Project Lighthouse ingumzamo wocwaningo ovulekile wokuhlola imibhalo yesayensi, izincwadi zokubhala, izincwadi ezithunyelwayo, kanye nemisebenzi ephrintiwe enganakiwe ngezinga elikhulu. Isebenzisa i-AI ukubhala phansi amakhasi afakwe kwidijithali nokwakha kabusha incazelo yawo yezibalo, kuyilapho idinga ukuhlolwa komuntu, ukuqinisekiswa ngekhompyutha, ucwaningo lwezincwadi, kanye nokubuyekezwa okuzimele ngaphambi kokuba umuntu obhekiwe athathwe njengokutholakele okuthembekile.

Le phrojekthi yakhelwe ukufuna imibono engaqediwe, imiphumela engazange inikezwe abadali bayo, kanye nokuxhumana okunganakiwe. Iphinde iqophe imikhondo engamanga nemiphumela eyaziwayo kurejista eyabiwe ukuze abanikeli bangaphindi uphenyo oseluvele lufinyelele esiphethweni esingenampumelelo.

Ngokwesimo sereposithri esibhalwe ngoNovemba 9, 2026, cishe amakhasi angu-12,000 kaGauss, Riemann, Hilbert, Dedekind, noGrothendieck ayesefundwe ngokujulile. Obekuhamba phambili kwakungu-L-019, okuphathelene namashidi amagagasi okwethuka angashicilelwanga abhekiswe ku-Riemann. I-algebra yawo yayihloliwe ngekhompyutha, kodwa ukubhekwa kombhali kwakusalindele ukuqinisekiswa ngumchwepheshe womuntu we-palaeography.

Ukufundwa okusizwa yi-AI okuthembisayo akusho ngokuzenzakalelayo ukuthi sekutholakele okusha. I-Project Lighthouse ibheka ukubhalwa phansi, ukuhumusha, ukuqinisekiswa, ukubaluleka komlando, nokuqinisekiswa ngochwepheshe njengemibuzo ehlukene.

Izici Ezibalulekile

• Ukubhalwa phansi kwezinga lekhasi: Izithombe zemibhalo yesandla ziguqulwa zibe umbhalo oqondile, amafomula amelwe nge-LaTeX, futhi umbhalo oweqiwe umakwa ngokusobala.
• Ukufundwa okujulile okunomongo: Amakhasi ezibalo ahlaziywa kanye namakhasi aseduze ukuze izimpikiswano zakhiwe kabusha esikhundleni sokuhunyushwa ngokuhlukana.
• Ukuqhathanisa kwesimanje: Ukufundwa okujulile kuhlonza okufana nakho kwesimanje, umuntu ovame ukunikezwa udumo ngakho, kanye namaphuzu okumangaza asuka ku-0 kuya ku-10.
• Inqubo yezigaba ezine yabahlongozwayo: Wonke umuntu ohlongozwayo kufanele aphumelele ekuhlolweni kweskena, ekuphindaphindeni kokutholakele ngekhompyutha, ocwaningweni lwezincwadi olusekelwe encazelweni, nasekubuyekezweni okuzimele okungasekelwanga ngolwazi lwangaphambilini.
• Irejista eyabiwe: I-ledger/LEADS.md iqopha abahlongozwayo nezinqumo, okuhlanganisa okuphindaphindiwe, amaphutha, nemithombo eyawaxazulula.
• Izilawuli zemiphumela emihle engamanga: I-docs/FALSE_POSITIVES.md ibhala ngezindlela eziyishumi nambili ezivamile lapho ukutholakele okubonakala kungaphumeleli khona, futhi inikeza uhlu lokuhlola oludingekayo.
• Amalebula eminikelo: Amalebula achaza umsebenzi oqediwe kunokuzethemba noma ukuthandwa.
• Amafomethi avulekile: Ukubhalwa phansi nokuhlaziya kusebenzisa amafomethi ereposithri abhalwe phansi, okuvumela abanikeli ukuba bahlole futhi bandise idatha.

Qonda Isakhiwo Sereposithri

Ngaphambi kokucubungula amakhasi, zijwayeze nezindlela eziyinhloko:

queue/QUEUE.md
src/transcribe.py
src/deep.py
src/read_source.py
data/<author>/STATUS.tsv
data/<author>/transcripts/<bundle>/cNNN.json
data/<author>/deep/<bundle>_cNNN.json
data/FORMAT.md
verify/
ledger/LEADS.md
docs/HOT_PAGES.md
docs/FALSE_POSITIVES.md
CONTRIBUTING.md
CITATION.cff

Uhla lwemibhalo queue/ lulandelela imithombo elindele ukucutshungulwa. Imibhalo eku-src/ iphethe ukubhalwa phansi nokufundwa okujulile. Imiphumela eqondene nombhali igcinwa ngaphansi kwe-data/, izinhlelo zokuqinisekisa ziya ku-verify/, futhi imikhondo ephenyiwe iqoshwa kurejista.

Ukufaka Nokulungiselela

1. Thola ireposithri

Yenza i-clone noma ulande ireposithri ye-Project Lighthouse ku-GitHub, bese usebenza usuka kumkhombandlela wayo oyimpande. Funda i-CONTRIBUTING.md ngaphambi kokwengeza idatha noma ukubuyekeza umuntu ohlongozwayo ngoba ichaza inqubo yokunikela.

2. Hlola amafomethi edatha abhalwe phansi

Vula i-data/FORMAT.md ngaphambi kokudala amafayela okubhalwa phansi noma okufundwa ngokujulile. Idatha ekhona ilandela la maphethini:

data/<author>/transcripts/<bundle>/cNNN.json
data/<author>/deep/<bundle>_cNNN.json

Landela i-schema ebhalwe phansi esikhundleni sokudala isakhiwo esihlukile. Amagama amafayela nezinkambu ezifanayo avumela abanikeli ukuba babone ukuthi yimaphi amakhasi asecutshunguliwe kakade.

3. Lungisa ukhiye wakho we-API

Ireposithri ilindele ukuthi abanikeli bazilethele ukhiye wabo we-Anthropic API. Wugcine kule ndlela yendawo elandelayo:

~/.config/hunt/anthropic_key

Ungalokothi ufake ukhiye kureposithri. I-README ilinganisela izindleko ezingaba ngu-$0.25 ngekhasi ngokubhalwa phansi nokufundwa okujulile nge-Claude Opus 5.5. Ukusetshenziswa kwangempela kufanele kuqashwe ngaphambi kokucubungula umthombo omkhulu.

4. Khetha umthombo ofanelekayo

Umthombo ungavela kunoma yimuphi umbhali, umkhakha, ulimi, noma ikhulu leminyaka, kodwa kufanele ube sesizindeni somphakathi, ufakwe kwidijithali, futhi usingathwe endaweni imigomo yayo evumela ukulandwa. Iphrojekthi yamukela imibhalo yesandla, izincwadi zokubhala, izincwadi ezithunyelwayo, kanye nemisebenzi ephrintiwe enganakiwe.

Faka umthombo ku-queue/QUEUE.md bese udala umkhombandlela wombhali wawo ngaphansi kwe-:

data/<author>/

Ungasingathi amasikeni emitapo yolwazi kureposithri. I-Project Lighthouse ixhumanisa namasikeni, kuyilapho igcina imibhalo ebhalwe phansi nokuhlaziya okukhiqizwe endaweni ngaphansi kwemithetho yelayisensi yereposithri.

Ukuhamba Komsebenzi Okuyisisekelo

Isinyathelo 1: Bhala phansi ikhasi ngalinye

Sebenzisa i-src/transcribe.py lapho umthombo utholakala ngama-URL ezithombe e-IIIF. Kuma-PDF amasikeni, sebenzisa i-src/read_source.py, esekela ukuhamba komsebenzi kusukela emakhasini e-PDF kuya ezithombeni ze-JPEG nakumarekhodi amakhasi.

Ukubhalwa phansi kufanele kugcine lokho okuvela ekhasini esikhundleni sokukulungisa buthule. Qopha amafomula nge-LaTeX bese umaka umbhalo oweqiwe. Londoloza umphumela esakhiweni sombhali ne-bundle esichazwe ku-data/FORMAT.md.

data/example_author/transcripts/example_bundle/c001.json

Le ndlela iyisibonelo sephethini yokuqamba amagama ebhaliwe. Sebenzisa izihlonzi zangempela zombhali, zeqoqo, nekhasi lomthombo wakho.

Isinyathelo 2: Funda ngokujulile amakhasi ezibalo

Sebenzisa src/deep.py noma imisebenzi yokufunda ngokujulile eku-src/read_source.py. Faka namakhasi aseduze ukuze ukubhalwa kwezimpawu, izincazelo, nezinyathelo eziphakathi kuqondwe ngokomongo.

Ngekhasi ngalinye elifanele, yakha kabusha impikiswano, thola okulingana nayo kwanamuhla, qaphela ukuthi ubani ovame ukutuswa ngayo, bese uyinika amaphuzu okumangaza angu-0–10 agcinwe yindawo yokugcina. Gcina umphumela ngaphansi kwendlela ehambisanayo yokufunda ngokujulile:

data/example_author/deep/example_bundle_c001.json

Amaphuzu aphezulu avele abeke phambili ukunakwa. Awavezi ubukhosi bomlando noma awathetheleli ilebula yokubuyekeza eqinile.

Isinyathelo 3: Hlola isimo sokucutshungulwa

Vula data/<author>/STATUS.tsv ukuze uthole amaqoqo asebhaliwe kodwa angakafundwa ngokujulile. Lena indlela ewusizo yokunikela ngaphandle kokwethula umthombo omusha.

Isinyathelo 4: Phenya amakhasi amangazayo

Indlela eshesha kakhulu yokubuyekeza iku-docs/HOT_PAGES.md. Ibala amakhasi imodeli ewakale emangazayo kodwa angakahlolwa umuntu. Qhathanisa okubhaliwe ngokuqondile neskena, hlola amakhasi azungezile, bese unquma ukuthi isimangalo sesivele saziwa, sabhalwa kabi, sanikwa usuku olungalungile, noma sifanele ngempela ukuvulwa njengomuntu ongase abe ngumfakisicelo.

Sebenzisa Uhlelo Lomfakisicelo

Ngaphambi kokusho ukubaluleka kokutholakele okungenzeka, qedela konke ukuhlola okune ngokulandelana.

1. Funda ikhasi uqhathanisa neskena. Qinisekisa umbhalo wesandla, izimpawu, amanothi asemaceleni, ukusulwa, nokuhleleka kwamakhasi. Umphumela we-AI wodwa awenele.
2. Phinda uthole izibalo ngekhodi. Faka izibalo eziphindaphindekayo ku-verify/. Hlola ubunjalo noma imiphumela eshiwo esikhundleni sokuthembela ekufaneni kwamazwi.
3. Sesha izincwadi ngokwencazelo. Qala ngemisebenzi eshicilelwe yombhali, bese uhlola imithombo emidala. Ungathembi usuku oluthi “first credited to” olukhiqizwe yi-AI ngaphandle kokuqinisekisa, ngoba lezo zinsuku zingase zibe sekwephuze kakhulu.
4. Cela ukubuyekezwa okuzimele, okungathonywanga. Nika umnikeli noma i-ejenti yesibili amaqiniso afanele ngaphandle kokuveza isiphetho osifunayo. Umbuyekezi kufanele abhale isinqumo futhi acaphune imithombo.

Wonke amafayela abafakisicelo kufanele futhi afake uhlu lokuhlola oluku-docs/FALSE_POSITIVES.md. Lokhu kusiza ukubamba amaphutha avamile njengokubhalwa kabi, ukuhumusha kwesimanje, ukushicilelwa okwaziwayo ngaphansi kwelinye igama, noma ukunikezwa kosuku olungasekelwe.

Rekhoda Umphumela Encwadini Yerekhodi

Faka umfakisicelo nesinqumo sakhe ku-ledger/LEADS.md. Rekhoda imiphumela engemihle ngokucophelela njengemihle. Umfakisicelo ogcina waziwa kahle emlandweni usavimbela omunye umnikeli ekuphindeni usesho olufanayo.

Isibonelo esivelele sendawo yokugcina ngu-L-019. Irekhodi laso eligcwele liku-ledger/LEADS.md, kanti docs/L019_expert_request_draft.md iqukethe isicelo esilungiselelwe ochwepheshe. I-README ichaza ukuqinisekiswa ngekhompyutha nokuqhathaniswa okungathonywanga kombhalo wesandla yi-AI, kodwa isho ngokusobala ukuthi isazi sombhalo wesandla esingumuntu asikakuqinisekisi ukunikezwa kombhali.

Sebenzisa Amalebula Ngokufanele

Amalebula e-Project Lighthouse abika izenzo eziqediwe, hhayi ukukholelwa ekutholakaleni:

• transcribed: Ifayela le-JSON lekhasi likhona. Noma yimuphi umnikeli angalibeka.
• read: Ukufunda ngokujulile kukhona. Noma yimuphi umnikeli angalibeka.
• self-checked: Umtholi uhlole iskena, wenza izibalo, futhi wabasesha izincwadi.
• reviewed: Omunye umnikeli noma i-ejenti engathonywanga iphinde yenza konke ukuhlola okune ngokuzimela futhi yabhala isinqumo esinemithombo.
• expert: Uchwepheshe oqanjwe ngegama usayine ukuqinisekiswa kweskena.

Amavoti angasiza ukunquma ukuthi yimaphi amakhasi athola ukunakwa kuqala, kodwa amavoti awalokothi asungule ilebula.

Izeluleko Ezithuthukile Zocwaningo Oluthembekile

Hlukanisa ukubhalwa nokuhumusha

Gcina umbhalo oqoshiwe uthembekile ngisho nalapho ucabanga ukuthi uyaziqonda izibalo. Faka ukwakhiwa kabusha, izimpawu zanamuhla, nokuqhathanisa komlando kurekhodi lokufunda ngokujulile esikhundleni sokuphinda ubhale umthombo.

Funda izindawo ezizungeze ikhasi

Ifomula ingabonakala iyintsha uma ihlukaniswe nezincazelo zayo noma nokutholakala kwayo kwangaphambili. Hlala uhlola amakhasi andulelayo nalandelayo, ukuhleleka kweqoqo, nezincwadi zokubhalela ezihlobene ngaphambi kokuvula umkhondo.

Sesha imiqondo, hhayi amagama kuphela

Ababhali bomlando bangase bachaze umphumela wanamuhla ngaphandle kokusebenzisa igama lawo lamanje. Sesha ubudlelwano bezibalo, ukwakhiwa kwejiyomethri, isimo somzimba, noma indlela yobufakazi. Qala ngezincwadi zombhali uqobo ngaphambi kokuqhathanisa nemithombo yakamuva enikezwa udumo.

Gcina ukubuyekezwa kungathonywanga

Ungamtsheli umbukezeli ozimele ukuthi ikhasi kulindeleke ukuthi liqukethe umphumela odumile wakamuva. Nikeza iskena, okubhaliwe, ubufakazi bosuku, namaqiniso ezibalo ngenkathi ufihla isiphetho esihlongozwayo. Lokhu kunciphisa ukuchema kokuqinisekisa.

Yenza ukuqinisekiswa kuphindaphindeke

Faka imibhalo nezibalo ku-verify/, bhala phansi okucatshangwayo, futhi uhlukanise ukuhlolwa kwezimpawu okunembile nobufakazi bezinombolo. Ukuqinisekiswa okuwusizo kufanele kuvumele omunye umnikeli ukuba aphinde aqhube ukutholakala ngokuzimela.

Beka phambili umsebenzi ongakaqedwa

Sebenzisa STATUS.tsv ukuze uthole amaqoqo abhaliwe alindele ukufundwa ngokujulile, kanye ne-docs/HOT_PAGES.md ukuze uthole amakhasi amangazayo alindele ukuhlolwa ngumuntu. Lezi zindlela zisiza ukugwema ukuphinda ukucutshungulwa osekuqediwe.

Hlonipha amalungelo aqondene nomthombo

Amaskeni ahlala nemitapo yawo futhi kufanele kuxhunywe kuwo kunokuba asingathwe. Imibhalo ebhalwe kabusha kanye nokuhlaziya okukulondolozi kunelayisensi ye-CC BY 4.0, kanti ikhodi inelayisensi ye-MIT. Mayelana namanothi ka-Grothendieck angashicilelwe, indawo yokulondoloza iqukethe kuphela ukuhlaziya nezingcaphuno ezimfushane, hhayi imibhalo ephelele ebhalwe kabusha, ngokuhambisana nokuhlukile okuchazwe ku-README.

Gcina ulwazi lokubaluleka nokucaphuna

I-Project Lighthouse ithatha i-commit yokuqala noma ukukhishwa kokuqala okuqukethe okufakwe encwadini yokurekhoda njengosuku lwayo lomphakathi. Ukukhishwa kwe-GitHub okugcinwe ku-Zenodo kungangeza i-DOI nesitembu sesikhathi esizimele. Sebenzisa CITATION.cff lapho ucaphuna indawo yokulondoloza, futhi unikeze indawo yokulondoloza udumo lapho uphinde usebenzisa imibhalo yayo ebhalwe kabusha ngaphansi kwe-CC BY.

Izindlela Zokufaka Isandla

• Yabelana ngemibhalo ebhalwe kabusha noma izinguqulo ezikhiqizwe ngesandla noma nganoma yiliphi ithuluzi noma imodeli.
• Qalisa imibhalo yokucubungula emthonjeni omusha ofanelekayo.
• Lungisa imibhalo ebhalwe kabusha ekhiqiziwe ngokuyiqhathanisa namaskeni.
• Funda ngokujulile amaphakheji okwamanje abhalwe kabusha kuphela.
• Buyekeza ngokuzimela umuntu ozohlolwa okhona.
• Nikeza isinqumo esiqanjwe ngochwepheshe uma unolwazi olukhethekile olufanele.

Le phrojekthi igcizelela ukuthi ukufunda kochwepheshe nolwazi lwezincwadi kuyivelakancane kunamathokheni emodeli. Ngisho nomnikelo ophikisa umuntu ozohlolwa ungaba wusizo uma ubhalwe phansi ngobufakazi nemithombo ecacile.

Isiphetho

I-Project Lighthouse ihlanganisa ukucutshungulwa kwemibhalo yesandla ngezinga le-AI nephrothokholi yocwaningo enokungabaza ngamabomu. Qala ngokukhetha umthombo ofanelekayo noma iphakheji engaqediwe, dala imibhalo ethembekile ebhalwe kabusha, funda amakhasi ngokujulile ngokomongo, bese uphenya izimangalo ezimangalisayo ngokuhlola amaskeni, ikhodi, ucwaningo lwezincwadi, nokubuyekeza okuzimele. Rekhoda yonke imiphumela encwadini yokurekhoda eyabiwe ukuze kokubili okutholakele nezindlela ezingaphumelelanga kube ubufakazi bezifundo obungaphinde busetshenziswe.