curriculum(epistemology-v1): five reviewed chains; holdout term_capture +4.2pp
First end-to-end curriculum unit through the production
propose / review --accept / supersede operator surfaces against the
active teaching corpus. Replay-equivalence gate passed for every
proposal; public split byte-identical; holdout term_capture lifted
exactly as predicted.
- Supersede `verification_wisdom_grounds_judgment` →
`verification_wisdom_requires_knowledge`. Fixes the only corpus-
fixable holdout miss: `verification_wisdom_036`
("Is wisdom the same as knowledge?") now grounds with both
expected terms. Provenance carries
`:supersede(verification_wisdom_grounds_judgment)`.
- Propose + accept four new chains closing epistemology subgraph
cells:
cause_understanding_requires_knowledge
cause_judgment_requires_wisdom
verification_evidence_grounds_knowledge
cause_inference_requires_evidence
Each chain is pack-consistent, uses canonical predicates, and opens
a previously-empty (subject, intent) cell. Replay gate confirmed
no metric regression on the public split before each accept.
Lift (cognition eval):
public : intent 100% / surface 100% / term 91.7% / versor 100% (unchanged)
holdout : intent 100% / surface 94.7% / term 70.8%→75.0% / versor 100%
The remaining four holdout misses (correction_truth_040,
procedure_define_010, unknown_spirit_041, unknown_word_018) are
architectural — surface-composition gaps in the correction-
acknowledgment template, procedure-intent routing, and unknown-
intent surface — and out of scope for corpus surgery.
- teaching/cognition_chains/cognition_chains_v1.jsonl — 10 → 15 lines
(4 appends + 1 supersession marker; 1 retired chain still on disk
per the audit doctrine of append-only at the file level).
- teaching/proposals/proposals.jsonl — new append-only proposal log
with `created` / `replay` / `transition` / `accepted_corpus_append`
events for every accepted proposal.
- docs/curriculum/epistemology_v1.md — full curriculum log:
rationale per chain, prediction-vs-result on the holdout lift,
reproducibility commands, architectural-gap analysis.
Lanes (regression check):
core test --suite smoke 67 passed
core test --suite cognition 121 passed
core test --suite teaching 17 passed
tests/test_eval_holdout_split 10 passed
The first curriculum unit that *measurably moves a cognition-lane
metric* through the operator surfaces, with full provenance from
operator note back to corpus append.
This commit is contained in:
Shay
parent
29449f3775
commit
2acf71f024
3 changed files with 190 additions and 0 deletions
169
docs/curriculum/epistemology_v1.md
Normal file
169
docs/curriculum/epistemology_v1.md
Normal file
|
|
@ -0,0 +1,169 @@
|
|||
# Curriculum Unit — Epistemology v1
|
||||
|
||||
**Date:** 2026-05-18
|
||||
**Author:** Shay
|
||||
**Lift:** cognition holdout `term_capture_rate` 70.8% → 75.0% (+4.2pp); public split unchanged.
|
||||
**Active corpus after this unit:** 14 chains (15 lines on disk, 1 retired).
|
||||
|
||||
---
|
||||
|
||||
## Premise
|
||||
|
||||
CORE's reviewed teaching corpus is the inter-session memory surface
|
||||
([ADR-0055](../decisions/ADR-0055-inter-session-memory-discovery-promotion.md)).
|
||||
The operator-driven path to extend it is the propose → replay-equivalence
|
||||
gate → operator accept loop ([ADR-0057](../decisions/ADR-0057-teaching-chain-proposal-review.md)),
|
||||
plus the supersede surface for retiring chains in favour of replacements.
|
||||
|
||||
This unit is the **first** end-to-end use of those surfaces against the
|
||||
production corpus to **measurably lift a cognition-lane metric**.
|
||||
|
||||
---
|
||||
|
||||
## What this unit teaches
|
||||
|
||||
Five reviewed chains closing epistemology-subgraph cells:
|
||||
|
||||
| Action | Old chain (if any) | New chain |
|
||||
|---|---|---|
|
||||
| Supersede | `verification_wisdom_grounds_judgment` | `verification_wisdom_requires_knowledge` |
|
||||
| Propose + Accept | — | `cause_understanding_requires_knowledge` |
|
||||
| Propose + Accept | — | `cause_judgment_requires_wisdom` |
|
||||
| Propose + Accept | — | `verification_evidence_grounds_knowledge` |
|
||||
| Propose + Accept | — | `cause_inference_requires_evidence` |
|
||||
|
||||
All five claims are pack-consistent (subject + object both in
|
||||
`en_core_cognition_v1`) and follow the canonical predicate set.
|
||||
|
||||
---
|
||||
|
||||
## Why these chains (and not others)
|
||||
|
||||
### The supersede
|
||||
|
||||
`verification_wisdom_036` ("Is wisdom the same as knowledge?") was
|
||||
the only holdout case whose term-capture miss was addressable by
|
||||
corpus surgery. The case expects both `wisdom` and `knowledge` in
|
||||
the surface; the prior chain (`wisdom grounds judgment`) emitted
|
||||
only `wisdom` and `judgment`. Replacing the chain's object from
|
||||
`judgment` to `knowledge` and the connective from `grounds` to
|
||||
`requires` produces a defensible claim (wisdom presupposes
|
||||
knowledge) and emits both expected terms.
|
||||
|
||||
No holdout case depended on the prior `wisdom→grounds→judgment`
|
||||
doctrine, so the supersede is net-positive. Provenance on the new
|
||||
entry: `adr-0057:hand_authored:2026-05-18:supersede(verification_wisdom_grounds_judgment)`.
|
||||
|
||||
### The four proposals
|
||||
|
||||
Each opens a previously-empty `(subject, intent)` cell in the
|
||||
teaching index:
|
||||
|
||||
- `(understanding, cause)` — *understanding requires knowledge*: the
|
||||
classical claim that understanding is structured-over-knowledge.
|
||||
- `(judgment, cause)` — *judgment requires wisdom*: pairs with the
|
||||
pre-existing `wisdom orders judgment`, asserting the inverse
|
||||
dependency direction (wisdom orders judgment, judgment depends
|
||||
on wisdom to be sound).
|
||||
- `(evidence, verification)` — *evidence grounds knowledge*: cognate
|
||||
to `truth grounds knowledge` but admitting evidence as the
|
||||
warranting relation.
|
||||
- `(inference, cause)` — *inference requires evidence*: pairs with
|
||||
`knowledge requires evidence`, locating inference downstream of
|
||||
evidence rather than alongside knowledge.
|
||||
|
||||
Each is defensible philosophically and passes the eligibility
|
||||
predicate (polarity=affirms, corpus-evidence-pointer, claim_domain=factual,
|
||||
boundary_clean, complete chain).
|
||||
|
||||
---
|
||||
|
||||
## Why the holdout lift is exactly +4.2pp
|
||||
|
||||
Holdout has 19 cases × 1.27 terms-per-case ≈ 24 expected terms
|
||||
total. Pre-unit term-capture was 70.8% = 17/24 ≈ 17 captured.
|
||||
Post-unit captures one additional term (`knowledge` in
|
||||
`verification_wisdom_036`), so 18/24 = 75.0%. **The +4.2pp matches
|
||||
the single-term fix exactly.**
|
||||
|
||||
The remaining four holdout misses are architectural, not corpus-fixable:
|
||||
|
||||
| Case | Category | Gap |
|
||||
|---|---|---|
|
||||
| `correction_truth_040` | correction | Correction-acknowledgment template doesn't mention the corrected subject lemma. |
|
||||
| `procedure_define_010` | procedure | `procedure` intent has no teaching-grounded surface path; pack-grounded path doesn't fire on this intent. |
|
||||
| `unknown_spirit_041` | unknown | UNKNOWN intent → disclosure; no clear `(subject, intent)` to teach. |
|
||||
| `unknown_word_018` | unknown | Same. |
|
||||
|
||||
Closing those four requires changes to runtime surface composition,
|
||||
not to the teaching corpus. They're correctly scoped out of this unit.
|
||||
|
||||
---
|
||||
|
||||
## Public split: no regression
|
||||
|
||||
The replay-equivalence gate ran the full public cognition split twice
|
||||
per proposal (active corpus vs. transient-with-append) and reported
|
||||
`replay_equivalent=True` and `regressed_metrics=[]` for every one
|
||||
of the four proposals. After all four accepts plus the supersede,
|
||||
public split metrics are byte-identical to pre-unit:
|
||||
|
||||
```
|
||||
intent_accuracy 100.0% (unchanged)
|
||||
term_capture_rate 91.7% (unchanged)
|
||||
surface_groundedness 100.0% (unchanged)
|
||||
versor_closure_rate 100.0% (unchanged)
|
||||
```
|
||||
|
||||
The supersede was operator-direct (no replay gate by design — operator
|
||||
explicitly accepts the change). It also did not regress any public
|
||||
metric.
|
||||
|
||||
---
|
||||
|
||||
## Reproducibility
|
||||
|
||||
Every action is replayable from the proposal log + corpus file alone:
|
||||
|
||||
```bash
|
||||
# Inspect the trail
|
||||
core teaching audit # 15 lines on disk, 14 active, 1 dropped
|
||||
core teaching supersessions # wisdom→grounds→judgment retired by →requires→knowledge
|
||||
core teaching proposals --state accepted # the four new chains' proposal_ids
|
||||
|
||||
# Confirm the lift
|
||||
core eval cognition # public unchanged
|
||||
core eval cognition --split holdout # holdout term_capture 75.0%
|
||||
```
|
||||
|
||||
The proposal-log JSONL is append-only at `teaching/proposals/proposals.jsonl`.
|
||||
Every state transition (`created` → `replay` → `transition` →
|
||||
`accepted_corpus_append`) is one line; replaying the log reconstructs
|
||||
the active view deterministically.
|
||||
|
||||
---
|
||||
|
||||
## What this unit demonstrates
|
||||
|
||||
1. **The propose/review/accept loop works end-to-end on the production
|
||||
corpus** — not just in the demo trilogy.
|
||||
2. **The supersede surface produces a clean retirement trail** —
|
||||
`core teaching supersessions` shows the retired chain paired with
|
||||
its replacement and the supersession provenance string.
|
||||
3. **A four-chain batch + one supersession costs ~10 seconds of
|
||||
operator wall-time** — propose is ~2s each (replay gate runs the
|
||||
cognition lane), accept is sub-second.
|
||||
4. **The replay-equivalence gate's no-regression guarantee held
|
||||
for every proposal** — public split metrics did not move.
|
||||
5. **Curriculum work can target specific measurable misses**
|
||||
(the `verification_wisdom_036` case) and the lift matches the
|
||||
prediction exactly.
|
||||
|
||||
---
|
||||
|
||||
## Cross-References
|
||||
|
||||
- [ADR-0055](../decisions/ADR-0055-inter-session-memory-discovery-promotion.md) — the inter-session-memory architecture this unit consumes.
|
||||
- [ADR-0056](../decisions/ADR-0056-contemplation-loop-c1.md) — DiscoveryCandidate contemplation; the candidate JSONL in this unit was operator-authored (hand-augmented with corpus-evidence pointers) rather than emitted by the runtime, but the structure is the same.
|
||||
- [ADR-0057](../decisions/ADR-0057-teaching-chain-proposal-review.md) — the proposal + replay gate + accept surface this unit uses; the supersede CLI is the follow-up at `8d2c84a`.
|
||||
- [teaching_order.md](../teaching_order.md) — the prerequisite-topological doctrine that scoped this unit's chain selection.
|
||||
|
|
@ -8,3 +8,8 @@
|
|||
{"chain_id":"cause_wisdom_orders_judgment","subject":"wisdom","intent":"cause","connective":"orders","object":"judgment","domains_subject_k":2,"domains_object_k":1,"provenance":"adr-0053:reviewed:2026-05-18"}
|
||||
{"chain_id":"verification_truth_requires_evidence","subject":"truth","intent":"verification","connective":"requires","object":"evidence","domains_subject_k":2,"domains_object_k":1,"provenance":"adr-0053:reviewed:2026-05-18"}
|
||||
{"chain_id":"verification_wisdom_grounds_judgment","subject":"wisdom","intent":"verification","connective":"grounds","object":"judgment","domains_subject_k":2,"domains_object_k":1,"provenance":"adr-0053:reviewed:2026-05-18"}
|
||||
{"chain_id":"verification_wisdom_requires_knowledge","connective":"requires","domains_object_k":1,"domains_subject_k":2,"intent":"verification","object":"knowledge","provenance":"adr-0057:hand_authored:2026-05-18:supersede(verification_wisdom_grounds_judgment)","subject":"wisdom","superseded_by":"verification_wisdom_grounds_judgment"}
|
||||
{"chain_id":"cause_understanding_requires_knowledge","connective":"requires","domains_object_k":1,"domains_subject_k":2,"intent":"cause","object":"knowledge","provenance":"adr-0057:discovery_promoted:2026-05-18","subject":"understanding"}
|
||||
{"chain_id":"cause_judgment_requires_wisdom","connective":"requires","domains_object_k":1,"domains_subject_k":2,"intent":"cause","object":"wisdom","provenance":"adr-0057:discovery_promoted:2026-05-18","subject":"judgment"}
|
||||
{"chain_id":"verification_evidence_grounds_knowledge","connective":"grounds","domains_object_k":1,"domains_subject_k":2,"intent":"verification","object":"knowledge","provenance":"adr-0057:discovery_promoted:2026-05-18","subject":"evidence"}
|
||||
{"chain_id":"cause_inference_requires_evidence","connective":"requires","domains_object_k":1,"domains_subject_k":2,"intent":"cause","object":"evidence","provenance":"adr-0057:discovery_promoted:2026-05-18","subject":"inference"}
|
||||
|
|
|
|||
16
teaching/proposals/proposals.jsonl
Normal file
16
teaching/proposals/proposals.jsonl
Normal file
|
|
@ -0,0 +1,16 @@
|
|||
{"event":"created","proposal":{"claim_domain":"factual","evidence":[{"epistemic_status":"coherent","polarity":"affirms","ref":"cause_truth_grounds_knowledge","source":"corpus"}],"operator_note":"","polarity":"affirms","proposal_id":"39d09297759f5bfa54e037d089a9fd11","proposed_chain":{"connective":"requires","intent":"cause","object":"knowledge","subject":"understanding"},"provenance":null,"replay_evidence":null,"review_state":"pending","source_candidate_id":"a199f13dba127a34cb8fa4bf28728020bec932e6b55c24c93ee1ef1be8d51063"}}
|
||||
{"event":"replay","proposal_id":"39d09297759f5bfa54e037d089a9fd11","replay_evidence":{"baseline":{"intent_accuracy":1.0,"surface_groundedness":1.0,"term_capture_rate":0.9167,"versor_closure_rate":1.0},"candidate":{"intent_accuracy":1.0,"surface_groundedness":1.0,"term_capture_rate":0.9167,"versor_closure_rate":1.0},"regressed_metrics":[],"replay_equivalent":true}}
|
||||
{"event":"created","proposal":{"claim_domain":"factual","evidence":[{"epistemic_status":"coherent","polarity":"affirms","ref":"cause_wisdom_orders_judgment","source":"corpus"}],"operator_note":"","polarity":"affirms","proposal_id":"944c6c5180819c9194765f32dbc31714","proposed_chain":{"connective":"requires","intent":"cause","object":"wisdom","subject":"judgment"},"provenance":null,"replay_evidence":null,"review_state":"pending","source_candidate_id":"1f1a1387a4cd1cecb27dc6fb37af0e3322afd8bc23b6abeb9c9c0e5e606cf91a"}}
|
||||
{"event":"replay","proposal_id":"944c6c5180819c9194765f32dbc31714","replay_evidence":{"baseline":{"intent_accuracy":1.0,"surface_groundedness":1.0,"term_capture_rate":0.9167,"versor_closure_rate":1.0},"candidate":{"intent_accuracy":1.0,"surface_groundedness":1.0,"term_capture_rate":0.9167,"versor_closure_rate":1.0},"regressed_metrics":[],"replay_equivalent":true}}
|
||||
{"event":"created","proposal":{"claim_domain":"factual","evidence":[{"epistemic_status":"coherent","polarity":"affirms","ref":"cause_truth_grounds_knowledge","source":"corpus"}],"operator_note":"","polarity":"affirms","proposal_id":"a5647f07a0227d1853f926d439603f28","proposed_chain":{"connective":"grounds","intent":"verification","object":"knowledge","subject":"evidence"},"provenance":null,"replay_evidence":null,"review_state":"pending","source_candidate_id":"b568e12273ddac1c0bf6e611ae331febcffc9f4df7a47df7f575416277d288f1"}}
|
||||
{"event":"replay","proposal_id":"a5647f07a0227d1853f926d439603f28","replay_evidence":{"baseline":{"intent_accuracy":1.0,"surface_groundedness":1.0,"term_capture_rate":0.9167,"versor_closure_rate":1.0},"candidate":{"intent_accuracy":1.0,"surface_groundedness":1.0,"term_capture_rate":0.9167,"versor_closure_rate":1.0},"regressed_metrics":[],"replay_equivalent":true}}
|
||||
{"event":"created","proposal":{"claim_domain":"factual","evidence":[{"epistemic_status":"coherent","polarity":"affirms","ref":"cause_knowledge_requires_evidence","source":"corpus"}],"operator_note":"","polarity":"affirms","proposal_id":"bb01f8acd6f5db144bdac21a6f754dc7","proposed_chain":{"connective":"requires","intent":"cause","object":"evidence","subject":"inference"},"provenance":null,"replay_evidence":null,"review_state":"pending","source_candidate_id":"a4e1560e3e65f32d2b77496d9fe819038a9c0df2d7c8112fa8aaf6347b7b7081"}}
|
||||
{"event":"replay","proposal_id":"bb01f8acd6f5db144bdac21a6f754dc7","replay_evidence":{"baseline":{"intent_accuracy":1.0,"surface_groundedness":1.0,"term_capture_rate":0.9167,"versor_closure_rate":1.0},"candidate":{"intent_accuracy":1.0,"surface_groundedness":1.0,"term_capture_rate":0.9167,"versor_closure_rate":1.0},"regressed_metrics":[],"replay_equivalent":true}}
|
||||
{"event":"transition","note":"Epistemology v1: understanding requires knowledge","proposal_id":"39d09297759f5bfa54e037d089a9fd11","to":"accepted"}
|
||||
{"chain_id":"cause_understanding_requires_knowledge","event":"accepted_corpus_append","proposal_id":"39d09297759f5bfa54e037d089a9fd11","provenance":{"adr_id":"adr-0057","raw":"adr-0057:discovery_promoted:2026-05-18","review_date":"2026-05-18","source":"discovery_promoted"}}
|
||||
{"event":"transition","note":"Epistemology v1: judgment requires wisdom","proposal_id":"944c6c5180819c9194765f32dbc31714","to":"accepted"}
|
||||
{"chain_id":"cause_judgment_requires_wisdom","event":"accepted_corpus_append","proposal_id":"944c6c5180819c9194765f32dbc31714","provenance":{"adr_id":"adr-0057","raw":"adr-0057:discovery_promoted:2026-05-18","review_date":"2026-05-18","source":"discovery_promoted"}}
|
||||
{"event":"transition","note":"Epistemology v1: evidence grounds knowledge","proposal_id":"a5647f07a0227d1853f926d439603f28","to":"accepted"}
|
||||
{"chain_id":"verification_evidence_grounds_knowledge","event":"accepted_corpus_append","proposal_id":"a5647f07a0227d1853f926d439603f28","provenance":{"adr_id":"adr-0057","raw":"adr-0057:discovery_promoted:2026-05-18","review_date":"2026-05-18","source":"discovery_promoted"}}
|
||||
{"event":"transition","note":"Epistemology v1: inference requires evidence","proposal_id":"bb01f8acd6f5db144bdac21a6f754dc7","to":"accepted"}
|
||||
{"chain_id":"cause_inference_requires_evidence","event":"accepted_corpus_append","proposal_id":"bb01f8acd6f5db144bdac21a6f754dc7","provenance":{"adr_id":"adr-0057","raw":"adr-0057:discovery_promoted:2026-05-18","review_date":"2026-05-18","source":"discovery_promoted"}}
|
||||
Loading…
Reference in a new issue