a0edbb4bdb
Second curriculum unit through the production operator surfaces.
Pure saturation — no cognition-lane lift expected (the eval splits
test fixed 32 cases that don't overlap with this unit's subjects),
but the live-prompt grounding surface expands materially: seven
prompts that previously fell through to disclosure now route to
deterministic teaching-grounded surfaces.
Three coherent clusters:
A. Cognition-source
cause_thought_reveals_meaning
cause_question_reveals_understanding
cause_recall_reveals_memory
B. Conceptual structure (bidirectional)
cause_definition_grounds_concept
verification_concept_requires_definition
C. Semantic content
cause_meaning_grounds_understanding
cause_analogy_reveals_relation
All pack-consistent (subject + object in en_core_cognition_v1),
canonical predicates (reveals / grounds / requires), each opens a
previously-empty (subject, intent) cell.
Replay-equivalence gate reported replay_equivalent=True for all
seven proposals (public cognition lane byte-identical pre/post
every accept).
Cognition lane:
public : intent 100% / surface 100% / term 91.7% / versor 100% (unchanged)
holdout : intent 100% / surface 100% / term 83.3% / versor 100% (unchanged)
Saturation lift is visible at the live-prompt level, not at the
eval level:
Why does thought exist? → [teaching] thought reveals meaning (...)
Why does a question exist? → [teaching] question reveals understanding (...)
Why does definition exist? → [teaching] definition grounds concept (...)
Why does meaning exist? → [teaching] meaning grounds understanding (...)
Why does an analogy exist? → [teaching] analogy reveals relation (...)
Does a concept require definition? → [teaching] concept requires definition (...)
Why does recall exist? → [teaching] recall reveals memory (...)
Why saturation matters: the cognition pack has 78 lemmas; we've
now covered ~21 (subject, intent) cells of the hundreds available.
Without saturation, prompts outside the 32 fixed eval cases are
coin-flips between vault recall and disclosure. Saturation moves
marginal prompts to deterministic teaching-grounded surfaces — the
foundation the composed-surface ADR (next) will compose over.
- teaching/cognition_chains/cognition_chains_v1.jsonl — 15 → 22 lines
(7 appends). Active set: 14 → 21 chains.
- teaching/proposals/proposals.jsonl — 7 new (created → replay →
transition → accepted_corpus_append) event sequences appended.
- docs/curriculum/cognition_saturation_v2.md — full curriculum log:
cluster rationale, live-prompt lift, operator-wall-time profile,
saturation-state-of-the-pack.
Lanes (regression check):
core test --suite smoke 67 passed
core test --suite cognition 121 passed
core test --suite teaching 17 passed
The non-negotiable field invariant (versor_condition < 1e-6) is
unaffected: this is corpus growth only; no code path changed.
7.3 KiB
Curriculum Unit — Cognition Pack Saturation v2
Date: 2026-05-18 Author: Shay Active corpus: 21 chains (was 14; +7 this unit). Lift: cognition lane metrics unchanged (already at architectural ceiling); saturation lift is in the live-prompt grounding surface — 7 new prompts that previously fell through to vault/disclosure now route to teaching-grounded surfaces deterministically.
Premise
The first curriculum unit
(epistemology_v1.md) targeted the
corpus-fixable holdout misses. ADR-0060 and ADR-0061 closed the
architectural holdout misses (correction template, procedure
routing). After those three pieces, the cognition-lane public
split is at 100/100/91.7/100 and holdout at 100/100/83.3/100 —
both at the architectural ceiling of the current pack +
surface composition.
The remaining 4/24 holdout term misses are concentrated on
UNKNOWN-intent cases that need a distinct selector (out of
scope for this unit).
This unit is not about lifting eval metrics. It's about
saturating the cognition pack's chain coverage so that more
of the pack's 78 lemmas have reviewed (subject, intent) cells
to route through teaching-grounded surfaces. Lift is observable
at the live-prompt level, not in the public/holdout eval splits
(which test a fixed 13/19 cases).
What this unit teaches
Seven new chains across three coherent clusters:
Cluster A — Cognition-source chains
The cognitive act produces meaning; questions produce understanding-gap exposure; recall produces awareness of memory.
| Chain | Defensible claim |
|---|---|
cause_thought_reveals_meaning |
Thought is the cognitive act through which meaning emerges. |
cause_question_reveals_understanding |
Socratic: questions reveal where understanding is incomplete. |
cause_recall_reveals_memory |
Recall is the mechanism that brings memory into awareness. |
Cluster B — Conceptual structure chains
Definitions ground concepts; concepts require definitions (bidirectional grounding).
| Chain | Defensible claim |
|---|---|
cause_definition_grounds_concept |
To define X is to ground the concept X. |
verification_concept_requires_definition |
A concept must be definable to count as a concept. |
Cluster C — Semantic-content chains
Meaning underlies understanding; analogies reveal relations.
| Chain | Defensible claim |
|---|---|
cause_meaning_grounds_understanding |
Semantic content underlies understanding. |
cause_analogy_reveals_relation |
Analogies reveal hidden structural relations. |
All seven are pack-consistent (subject and object both in
en_core_cognition_v1), use canonical predicates (reveals,
grounds, requires), and open previously-empty
(subject, intent) cells.
Lift profile
Cognition lane (fixed 13/19 cases)
| Split | Pre-v2 | Post-v2 | Δ |
|---|---|---|---|
| public | 100 / 100 / 91.7 / 100 | 100 / 100 / 91.7 / 100 | byte-identical |
| holdout | 100 / 100 / 83.3 / 100 | 100 / 100 / 83.3 / 100 | byte-identical |
No movement — expected. The 19 holdout cases don't include any of the subjects this unit teaches. Saturation isn't visible at the canonical-eval level.
Live-prompt lift
Seven prompts that previously fell through to the universal disclosure (or to a vault recall that may or may not return a coherent surface) now route to deterministic teaching-grounded surfaces:
Why does thought exist? → [teaching] thought reveals meaning (...)
Why does a question exist? → [teaching] question reveals understanding (...)
Why does definition exist? → [teaching] definition grounds concept (...)
Why does meaning exist? → [teaching] meaning grounds understanding (...)
Why does an analogy exist? → [teaching] analogy reveals relation (...)
Does a concept require definition? → [teaching] concept requires definition (...)
Why does recall exist? → [teaching] recall reveals memory (...)
Each emits a surface containing the subject, the humanised
connective, the object, and the relevant pack semantic_domains
strings — pack-grounded discipline preserved.
Why saturation matters even when eval numbers don't move
The eval splits test 32 fixed cases (13 public + 19 holdout).
The cognition pack has 78 lemmas; pack-consistent
(subject, intent) cells across {cause, verification} number
in the hundreds. Today's 21 chains cover ~21 of those cells —
roughly 12% saturation.
Without saturation, every prompt outside the 32 eval cases is a coin flip between vault recall (session-stateful, may return incoherent surface) and the universal disclosure (no grounding). Saturation moves the marginal prompt from "coin flip" to "deterministic teaching-grounded surface."
This is the foundation the composed-surface ADR will sit on: chain-of-chains composition only produces fluent surfaces when the constituent chains exist. v2 builds the constituent surface.
Process notes — operator wall-time
Total operator wall-time for this unit, end-to-end:
- Candidate authoring (7 JSONL files): ~3 minutes (mostly typing).
- Proposals (7 ×
core teaching propose): ~13 seconds each — the replay gate runs the full cognition lane on each. ~90s total. - Accepts (7 ×
core teaching review --accept): sub-second each. - Eval + audit + lane regression check: ~45 seconds.
Total: ~5 minutes of operator wall-time for 7 reviewed chains. The replay-equivalence gate is the dominant cost. Worth the cost: every accepted chain has a CI-grade no-regression guarantee at the moment of admission.
Cumulative corpus state
| Phase | Chains active | Cumulative additions |
|---|---|---|
| Pre-curriculum (2026-05-18 morning) | 10 | — |
| After epistemology v1 | 14 | +4 (1 supersede) |
| After cognition saturation v2 | 21 | +7 |
Now-covered subjects: light, knowledge, memory, correction, creation, truth, wisdom, understanding, judgment, evidence, inference, thought, question, definition, meaning, analogy, concept, recall.
Pack-resident subjects still without any teaching chain (~60+ lemmas): account, answer, ask, beginning, belong, cause, compare, comparison, contrast, context, define, discourse, distinction, distinguish, explain, figure, follow, ground, identity, image, infer, learn, life, mean, metaphor, narrative, order, person, precede, principle, procedure, reason, register, relate, relation, remember, reveal, rhetoric, simile, spirit, story, style, support, symbol, teach, tone, verify, voice, word.
The next saturation pass could target a verb-noun balance
(compare, contrast, distinguish, relate, explain,
teach, learn) — useful for procedure-intent prompts after the
ADR-0061 path landed.
Cross-References
- Curriculum: epistemology v1 — the first unit; chain selection rationale and operator workflow.
- ADR-0057 — the propose/review/accept surface.
- ADR-0060
- ADR-0061 — architectural surface composition fixes that close the remaining-fixable holdout misses.
- Next: composed-surface ADR (chain-of-chains realization) — the fluency surface that saturation v2 unlocks.