Extracting a Poem: Codex Deep-Extraction of Bukowski's I Met a Genius

donto research note — 2026-06-04

A 40-word poem, extracted to saturation. 415 live claims. One named person. The interesting half is the half that isn't in the text.

1. The poem, its title, author, and an honest provenance note

Title: I Met a Genius Author: Charles Bukowski (1920–1994) Collection: You Get So Alone at Times That It Just Makes Sense (Black Sparrow Press, 1986)

The version held in ctx:test/poem/bukowski-genius and used as the extraction source is, verbatim:

i met a genius on the train today
about 6 years old,
he sat beside me and as the train ran down along
the coast

we came to the ocean and then he looked at me
and said,

"it's not pretty."

I had not realized that before.

Provenance — read this before trusting any "fact" below

• Attribution is sound, but Bukowski attribution online is famously noisy. A large number of short, sentimental, or aphoristic poems circulate online over Bukowski's name that he did not write (the most notorious being "The Laughing Heart" misquotations and the widely-misattributed "you are marvelous, the gods wait to delight in you" fragments). I Met a Genius is not one of those: it appears in the genuine Black Sparrow collection You Get So Alone at Times That It Just Makes Sense. We web-verified the title and collection against multiple independent indexes. It is authentically Bukowski.
• Textual variant — flagged honestly. Our source file ends "I had not realized that before." Several web reproductions render the closing line as "It was the first time I'd realized that." These are semantically equivalent (first-time realization) but not word-identical. The extraction was run against our source text; any anchor to the final line reflects our wording, not the canonical Black Sparrow print line, which we did not photograph. Treat the closing-line anchors as accurate-to-source, provenance-uncertain-to-print.
• Only one person in this poem is named: the author, Charles Bukowski. The "genius" — the 6-year-old child — is unnamed, ungendered beyond the pronoun "he", and otherwise unidentified. So is the "I"/speaker (conventionally read as Bukowski's persona, but the text never says so). Every claim below that gives the child a class, an age, a role, or a meaning is the extractor's reading of an anonymous figure, not a record about a known individual. The age ("about 6") is itself hedged in the poem ("about") — the extractor preserved that hedge rather than asserting a hard integer.

2. Executive summary

We ran a single short Bukowski poem — 40 words, 9 lines, 11 non-blank lines counting the blanks — through a deep, coverage-to-saturation extraction (codex gpt-5.4 at reasoning effort xhigh) followed by an always-on semantic citer that anchors every claim back to the text or flags it honestly. The result, now live in donto context ctx:test/poem/bukowski-genius:

• 415 live claims over 51 distinct entities/events using 176 distinct predicates — the great majority of predicates freely minted by the model for this one poem (128 of 176 predicates appear exactly once).
• 133 STATED claims (32%) — anchored to an exact or near-exact span of the poem (entities, quoted speech, lines, the literal "about 6 years old", "the ocean", "the coast").
• 282 INTERPRETED claims (68%) — themes, irony, the child-instructs-adult inversion, the deflated aesthetic expectation, event ontology, the speaker/child/we-group identities — held as hypothesis, never given a fake span.
• The headline finding: for a minimal poem, deep extraction produces far more interpretation than transcription, and the value of the substrate is that it can hold the interpretation without lying about where it came from. The citer's refusal to bogus-anchor a theme is what makes the 68% trustworthy.

This is a stress test of the thesis in DONTO-ABUNDANCE.md: generation is cheap and unbounded (a 40-word poem yields 400+ claims), so the hard part is holding an abundant, mostly-interpretive firehose with its evidence status intact. donto holds all 415 paraconsistently — stated and interpreted side by side, contradiction-tolerant, evidence-first.

3. The process — codex xhigh deep extraction + the always-on citer

3.1 Extraction: coverage-to-saturation, no count-floor

The extractor was codex gpt-5.4 at reasoning effort xhigh, run in deep mode. The instruction was not "extract N facts" — there was no count-floor. The instruction was coverage to saturation: keep emitting properties and relations, in all directions, inventing predicates as needed, until the model stops finding new structure in the text. This is the generative-abundance posture: at extraction time you maximize, you do not gate. The poem is 40 words; the model found 415 distinct things to say about it and the world it implies.

What saturation looks like on a tiny input:

• Surface entities — ex:train, ex:coast, ex:ocean, ex:child, ex:speaker, the seats, the window, the sea-view.
• Events as first-class — ex:meeting-event, ex:sitting-event, ex:train-running-event, ex:arrival-event, ex:looking-event, ex:speaking-event/ex:utterance-event, ex:realization-event — each with participants, ordering (followed/followedBy), and temporal containment (during, partOf).
• Linguistic structure — every line reified (ex:line-1…ex:line-11), pronoun tokens (ex:opening-i-token, ex:it-token, ex:we-token, ex:closing-I-token), the quoted speech act (ex:utterance-1, ex:quoted-it).
• Propositional content — ex:ocean-not-pretty-proposition, ex:not-pretty-evaluation — the child's claim about the ocean held as its own node, separable from whether it is true.
• Bibliographic frame — ex:charles-bukowski, ex:black-sparrow-press, ex:you-get-so-alone-at-times-that-it-just-makes-sense.
• Themes/devices — ex:theme-child-instructs-adult, ex:theme-childlike-honesty, ex:theme-deflated-aesthetic-expectation, ex:theme-perception-versus-convention, ex:theme-seeing-anew, plus usesDevice (20 uses) and hasTheme (17 uses).

3.2 The citer: STATED vs INTERPRETED — the star of the show

Extraction answers what is claimed. It does not answer where in the source each claim is grounded — and the two are deliberately separate stages ([[feedback_always_post_process_cite]]). After extraction, an always-on semantic citer runs over every single claim and sorts it into one of two evidentiary classes, using a cascade of techniques (exact substring → lexical overlap → bge-small embedding cosine → honest "unanchorable"), never a hardcoded keyword list:

• STATED — the claim has a real anchor: an exact or near-exact span of the poem supports it. ex:train rdf:type ex:Train ⟵ "the train". ex:child quotedAsSaying "it's not pretty." ⟵ the quoted line. ex:age-estimate approximateValue 6 + usesHedge "about" ⟵ "about 6 years old". These are anchored at confidence up to 1.0.
• INTERPRETED — the claim is a reading: a theme, an irony, an event-ontology choice, an identity hypothesis. The citer finds no honest span, so it does the right thing: it marks the claim hypothesis_only and leaves the anchor null. It does not fabricate a span. A theme like "child instructs adult" is true of the poem but is nowhere in the poem as text — bogus-anchoring it would be a hallucinated citation, and the citer's whole job is to refuse that.

This is the load-bearing distinction. The citer's layers over the live set:

The principle, restated: extractors are free to maximize recall and even to interpret; the citer guarantees that nothing interpreted is dressed up as transcribed. Faithfulness becomes a property of the pipeline, not of any one model — rotate the extractor, the citer still separates fact from reading.

4. Everything extracted — the complete table

Every one of the 415 live claims in ctx:test/poem/bukowski-genius, untruncated. STATED = anchored to a span; INTERPRETED = held as hypothesis (anchor deliberately empty, shown as "—"). "Conf" is the citer/extractor confidence where recorded (· = not separately scored at cite time; the claim is live regardless).

4.1 Per-entity fact counts (51 subjects)

4.2 The full claim table (all 415)

5. Outcomes — what a 40-word poem yields under deep extraction

5.1 The numbers

5.2 Facet coverage — the poem decomposed in all directions

The extraction did not stop at "who/what". It covered, simultaneously: physical scene (train, coast, ocean, window, seats), dramatis personae (child, speaker, we-group — all anonymous), event structure (10+ reified events with ordering and containment), speech-act layer (utterance, quotation, the proposition "the ocean is not pretty" as a standalone node), textual/linguistic layer (every line, every pronoun token), bibliographic frame (author, press, collection), and interpretive layer (5+ named themes, devices, ironies). The two largest subjects — ex:poem and ex:source-poem (50 facts each) — are the poem reflecting on itself: its lines, its themes, its devices, its frame.

5.3 Anchored vs hypothesis — the central result

A minimal, concrete poem still comes out roughly one-third transcription, two-thirds interpretation. That ratio is the point. The poem's surface is tiny and quickly exhausted (133 stated claims and you have essentially read the literal text into structure). But the poem's meaning — why a child's flat "it's not pretty" is a small devastation, why the inversion of child-instructs-adult lands, why the speaker's "I had not realized that before" is the volta — is unbounded, and the extractor reached for it freely. donto holds all of it, with the 282 interpretive claims clearly marked as hypothesis rather than masquerading as quotation.

5.4 What this demonstrates

• Esoteric / aesthetic extraction works under the abundance model. You do not need a "poetry schema". The model invented usesHedge, quotedAsSaying, hasTheme, deflatedAestheticExpectation-style predicates on the fly — exactly the emit-free/untyped, defer-typing-to-query-time posture donto is built for. The 176 predicates (128 used once) are not a proliferation bug; they are the signature of abundance on a single tiny input.
• Paraconsistent, evidence-first holding is what makes interpretation safe to keep. A vector DB or a winner-take-all KG would either discard the themes (not "facts") or flatten them into the same status as "the train". donto holds the literal span "the ocean" and the contested reading "this poem is about perception versus convention" in the same store, at the same time, with their evidentiary status preserved — one anchored, one openly hypothetical. Nothing was thrown away; nothing was faked.
• Faithfulness is a pipeline property, not a model promise. Because the citer is a separate, always-on stage, the 68% interpretive mass never received a fabricated span. You can swap codex for any other extractor tomorrow and the stated/interpreted boundary still holds, because it is drawn by the citer against the source — not asserted by the generator.
• The honest limits. 415 claims on 40 words is over-generation by any practical measure — much of it is the same scene described under several minted predicates (e.g. destination/hasDestination, attestedBy/several event-participant predicates). That redundancy is exactly the query-time alignment workload donto defers: collapse destination≈hasDestination by similarity later, never by a hand-maintained synonym table. The extraction's job was to miss nothing; reconciliation is a separate, learned step.

5.5 Provenance caveat, restated

Every claim about the child is a claim about an unnamed, anonymous figure read from eight lines of verse. The only named person in the entire 415-claim graph is Charles Bukowski, the author. The age is hedged in the source and the hedge was preserved. The closing line differs from some print reproductions and was extracted as-sourced. These are honest limits, surfaced rather than smoothed.

Substrate context: ctx:test/poem/bukowski-genius · 415 live statements · extractor codex gpt-5.4 xhigh · always-on semantic citer (lexical → bge-small → honest-hypothesis). Source: /home/ajax/extraction-harness/poem-source.txt. This note is a companion to the canonical Generative Abundance report.

Sources (provenance verification): Hello Poetry — I Met A Genius · PoetryVerse — I Met A Genius · PoemHunter — I Met A Genius