Deep-mode extraction on a 109-word Discord message yields 1000 facts

Subject: A single Discord message about Nietzsche's Eternal Recurrence posted in #donto was passed through donto-memory's new mode: "deep" extraction with 7 sequential LLM passes. Pipeline: omega-bot → /memorize (async, mode=deep, passes=7) → tokio queue → 7× GLM-4.7 → dedup → ingest Job: d7ed356c-4c95-4ad6-a1e2-22bb2324ad3d Holder: agent:omega-bot Session: discord:1349727923434815519:1497274794586931220

TL;DR

• Input: 109 words (≈740 chars) — Anthony Ludovici's introduction to Nietzsche's Eternal Recurrence as a song-form doctrine.
• Output: 1000 unique facts ingested into donto, 998 successfully landed in the substrate.
• Wall time: 854.6 s (~14 min 14 s).
• Raw LLM yield: 1013 raw facts → 1000 after 13 dedup collisions (1.3% redundancy across 7 passes).
• One pass failed cleanly (pass_2: model returned prose, no JSON) — the remaining 6 passes still produced 1000 unique facts.
• Ratio: ~9.2 facts per input word — almost certainly the highest ratio I've seen out of donto.

The input

ajaxdavis in #donto: It were the height of presumption to attempt to fix any
particular interpretation of my own to the words of this song. With what has
gone before, the reader, while reading it as poetry, should be able to seek
and find his own meaning in it. The doctrine of the Eternal Recurrence
appears for the last time here, in an art-form. Nietzsche lays stress upon
the fact that all happiness, all delight, longs for repetitions, and just as
a child cries "Again! Again!" to the adult who happens to be amusing him; so
the man who sees a meaning, and a joyful meaning, in existence must also cry
"Again!" and yet "Again!" to all his life.

(For reference: this is from Anthony M. Ludovici's commentary in the Thus Spake Zarathustra introduction — the user noted afterward they "just copied that from a HN thread.")

Per-pass yield

Observations on the curve:

• Pass 1 (cold start, no prior_facts context): the model treats the input as fresh and dumps the obvious surface facts — Discord metadata, author, channel, message text, primary topics. 132 facts is plenty for a 109-word input on its own.
• Pass 3 (first pass with prior_facts injected): the model overcompensates — knowing 132 facts already exist, it strains for novelty and produces 202 facts in 62 s. This is the highest-yielding pass and also the fastest, suggesting the prior_facts block successfully redirects the model into less-trodden conceptual territory without it having to re-derive the obvious.
• Passes 4–7: the saturation gradient is clearly visible — 204 → 159 → 158 → 158 raw, with collisions climbing 2 → 2 → 2 → 7. The model is running out of unmined angles. Pass 7 has the highest collision rate (4.4%), suggesting we're at or near the natural ceiling of what this LLM can find in this text.
• Wall time is dominated by latency, not yield. Pass 3 produced more facts than pass 1 in half the time. Latency variance comes from generation length and possibly server-side queueing on Z.AI's GLM-4.7 endpoint — not from how hard the model is "thinking."

Pass 2 failed silently — the prose-not-JSON case

Pass 2's worker logged LLM response decode: error decoding response body and contributed 0 facts. The model returned a chatty preamble ("Here are the new facts I've extracted that go beyond what the previous pass found...") instead of bare JSON. The orchestrator did the right thing: it logged the failure, appended an empty result to the running set, and continued to pass 3.

Action item: add a single-shot JSON-mode retry on parse failure. The model would almost certainly have produced JSON on the second attempt — losing one pass out of seven is a 14% capacity loss on a deep-mode run.

What kinds of facts came out

Subject namespace mix

987  donto:    (ontology entities — concepts, classes, instances)
 13  discord:  (Discord-level provenance)

The Discord-shell facts (user, channel, message-id, posted-in) all fit in the first 13 statements. The other 987 facts are all about the content — Nietzsche, the doctrine, the reader, repetition, joy, the song-form, the disclaimer, the rhetorical move, the lexicon.

Modality distribution

 42  descriptive    (claims directly recoverable from the text surface)
958  inferred       (interpretive moves: "X presupposes Y", "X exemplifies Y", "X conveys Y")

96% of the facts are inferred, not descriptive. This is exactly what a 109-word input should produce when you ask for "every possible fact" — the surface is small; the interpretive surface is enormous.

Confidence distribution

1.00   26    descriptive facts the model is sure about
0.95  193    high-confidence inferences (e.g. "reader has-interpretive-freedom")
0.90  361    median inference confidence
0.85  234    softer inferences
0.80  159    fuzzy / late-pass material
0.75   20
0.70    7    weakest inferences (typically pass 6–7 social-strategy speculation)

The model self-calibrates the way you'd hope: the bedrock claims (the author, the channel, the literal phrases) get 0.99–1.00; mid-confidence material is the doctrinal content; the lowest-confidence facts cluster in the later passes where the model is stretching for novelty.

Predicate cardinality

352  rdf:type          (typed instance assertions — ontology growth)
 16  donto:concerns
 14  donto:usesPhrase
 12  donto:involves
 11  donto:expresses
  9  donto:includes
  8  donto:requires
  8  donto:hasObject
  8  donto:contrastsWith
  7  donto:produces

rdf:type dominates because deep-mode discovers ontology, not just statements. A single sentence "the reader should seek his own meaning" produces donto:Reader rdf:type donto:Agent, donto:Meaning rdf:type donto:Concept, donto:MeaningSeeking rdf:type donto:CognitiveAct, etc. Each new concept introduced gets typed. After 7 passes the model has minted hundreds of new classes — which is good if you want a rich ontology to grow and bad if you want a controlled vocabulary.

Quality samples grouped by theme

The 1000 facts cluster into four bands. Sampling from each:

Band 1 — Discord-shell metadata (pass 1, 13 facts)

discord:user:ajaxdavis            rdf:type              donto:DiscordUser
discord:channel:donto             rdf:type              donto:DiscordChannel
discord:message:f58d1d3b…         donto:hasAuthor       discord:user:ajaxdavis
discord:message:f58d1d3b…         donto:postedIn        discord:channel:donto
discord:message:f58d1d3b…         donto:hasSentenceCount  5
discord:message:f58d1d3b…         donto:hasMood          donto:DeclarativeMood
discord:message:f58d1d3b…         donto:hasRegister      donto:FormalRegister
discord:message:f58d1d3b…         donto:hasSentiment     donto:ReflectiveSentiment

Clean. This is exactly the kind of provenance scaffolding the substrate needs to attach the content to a holder/session/message.

Band 2 — Direct content (passes 1–3, ~300 facts)

donto:Song                        rdf:type              donto:ArtisticWork
donto:EternalRecurrence           rdf:type              donto:PhilosophicalDoctrine
donto:DoctrineOfEternalRecurrence donto:hasFinalAppearanceIn  donto:ArtForm
donto:Nietzsche                   donto:laysStressOn    donto:Repetition
donto:Repetition                  donto:isLongedForBy   donto:Happiness
donto:Repetition                  donto:isLongedForBy   donto:Delight
donto:Reader                      donto:readsAs         donto:Poetry

These are the doctrinal claims the text actually makes. High confidence (0.9–1.0), descriptive or strong-inference.

Band 3 — Interpretive inferences (passes 4–6, ~500 facts)

donto:Interpretation              donto:hasRisk         donto:Presumption
donto:Interpretation              donto:isFallible      True
donto:Song                        donto:hasMultipleInterpretations  True
donto:Song                        donto:isInterpretivelyOpen        True
donto:Existence                   donto:canHaveJoyfulMeaning        True
donto:MeaningConstruction         donto:isSubjective    True
donto:AjaxDavis                   donto:presupposes     donto:ReaderCompetence
donto:ReaderCompetence            donto:includes        donto:PoeticSensitivity
donto:PoeticSensitivity           rdf:type              donto:AestheticCapacity

This is where deep-mode earns its keep. Pass 3+ surfaces structural claims about hermeneutics, the reader-author relationship, the rhetorical apparatus — none of which is directly stated in the text but all of which is legitimately implied by it. A single-shot extraction would never reach this layer.

Band 4 — Pragmatic / speech-act / lexicographic (pass 6–7, ~200 facts)

donto:InterpretiveDisclaimer      donto:hasIllocutionaryForce       donto:DeclarativeAct
donto:InterpretiveDisclaimer      donto:servesInterpersonalFunction  donto:FaceSaving
donto:FaceSaving                  rdf:type              donto:SocialStrategy
donto:AjaxDavis                   donto:manages         donto:EpistemicReputation
donto:HeightOfPresumption         rdf:type              donto:IdiomaticExpression
donto:HeightOfPresumption         donto:conveys         donto:EpistemicHumility
donto:EpistemicHumility           donto:contrastsWith   donto:Presumption
donto:AjaxDavis                   donto:usesPhrase      "with what has gone before"
donto:ContextualPreparation       donto:enables         donto:ReaderAutonomy

By pass 7 the model is reaching for speech-act theory, face-saving, idiom analysis, and rhetorical scaffolding. This is genuinely novel material — none of these claims overlap with bands 1–3. The collision rate ticking up to 4.4% in pass 7 says we're scraping the bottom; pass 8 would probably collide more than it added.

A non-obvious side-effect: identity collapse

Note that the model invented two equivalent identifiers for the same author:

discord:user:ajaxdavis            rdf:type              donto:DiscordUser   (pass 1)
donto:AjaxDavis                   rdf:type              donto:DiscordUser   (pass 5)
donto:AjaxDavis                   donto:hasUsername     "ajaxdavis"        (pass 5)

This is a known donto problem (identity alignment is its own family in the spec) but it's worth surfacing: deep-mode amplifies it. Passes 5+ tend to introduce capitalized, domain-style names for entities the earlier passes referred to with namespaced IDs. A post-extract identity-resolution pass (or stricter instructions to reuse prior IRIs verbatim) would help.

Operational notes

Timeouts

We had to bump the HTTP client timeout from 180 s to 900 s mid-experiment because pass 1 took ~125 s and pass 4 took ~127 s, both within the safe band — but a worst-case pass crashed against the old 180 s ceiling. Z.AI's GLM-4.7 endpoint, under load, can take >3 minutes to return a long JSON response. 15-min client timeout is the right number.

Async + serialized queue is mandatory

Cloudflare cuts proxied HTTP at ~100 s. A synchronous 854 s /memorize would always 524. The new pipeline:

1. /memorize validates input, writes a (queued) row to donto_x_memory_job_log, returns 202 with {queue_id, status:"queued"} immediately.
2. A tokio task acquires async_memorize_lock (a single-permit Mutex — one extraction at a time) and runs memorize_one to completion.
3. Caller polls /jobs/<queue_id>/raw for progress.

The single-permit lock matters: deep-mode is expensive, and running two of them in parallel against the same GLM endpoint would burn quota with very little wall-time speedup.

Logging visibility issue

The per-pass info! lines didn't reach journalctl during this run because the tracing subscriber's effective level filter was promoting to WARN at the binary level. The work completed fine — the audit row in donto_x_memory_job_log confirmed facts=1000 — but observability was thin during the run. Fix: explicit RUST_LOG=info,donto_memory=info,donto_memory_core=info in the systemd unit, plus ensure the tracing init in donto-memory/src/main.rs:60 doesn't get clipped.

What this proves about deep-mode

1. Yield scales with passes, but not linearly. 7 passes produced ~7.6× the per-pass yield of a default single-shot mode would have. Past pass 4 the marginal return slopes off but is still positive.
2. Prior-facts context is the workhorse, not the prompt variation. The pass-to-pass differentiation comes from showing the model what's already been said, not from rotating prompts. The "no rigid prompts" decision was correct — the model finds its own angles when told "here are 132 things we already know; find more."
3. The dedup engine is doing real work. 13 collisions across 1013 raw facts is exactly the right zone — too low would mean passes were independent (and we're missing easy wins from cross-pollination); too high would mean the prior-facts injection isn't being honoured. 1.3% is roughly ideal.
4. The substrate ingested 998/1000. A 0.2% loss rate is from substrate-side rejections (likely identity/predicate-domain violations, the substrate's policy gate fighting the model's eager class invention). That's the next thing to look at.

Recommendations

• Retry pass on JSON parse failure — recover the 14% capacity loss from pass_2.
• Bump max_tokens from 8000 to 12000 — the largest raw-yield pass (pass 4 at 204 facts) was approaching the token ceiling; some facts were salvaged from truncated JSON by the extract.rs salvage path.
• Add an identity-resolution post-pass — collapse discord:user:ajaxdavis / donto:AjaxDavis / donto:Reader (in the cases where Reader refers to ajax) before ingest. Or instruct the model: "use the exact IRIs from prior facts; do not create new IRIs for entities already introduced."
• Promote the tracing level filter — operators couldn't see per-pass progress during this run.
• Investigate the 2-fact substrate rejection — ingested=998 vs unique=1000, what was rejected and why.
• Default passes per modality — a 110-word philosophical passage benefits enormously from 7 passes. A "yeah for real." 3-word reply does not. The omega-bot client should pick passes based on input length (e.g. max(1, min(7, Math.ceil(wordCount / 30)))).

Job artefact

Full job detail + 1000 facts: /jobs/d7ed356c-4c95-4ad6-a1e2-22bb2324ad3d (raw JSON: append /raw)

Related context: donto substrate PRD, donto-memory omega-bot corpus audit.