feat(W-019): learning-arc demo — engine-authored proposal from contemplation (ADR-0152) (#276 )

Two-session arc where engine derives connective+object from corpus
decomposition; operator ratifies rather than authors. Distinguishes
from learning-loop (operator-authored) and directly exercises W-018
checkpoint contemplation and W-017 auto-proposal provenance path.

2026-05-25 13:03:10 -07:00

3.3 KiB

Raw Blame History

ADR-0152 — Learning-Arc Demo (`core demo learning-arc`)

Status: Accepted
Implements: W-019
Depends on: ADR-0150 (W-018 checkpoint contemplation), ADR-0151 (W-017 auto-proposal)

Context

ADR-0055..0057 ships core demo learning-loop, which demonstrates the full cold-turn → discovery → operator-authored proposal → accept → grounded surface arc. In that demo the operator supplies the connective, object, and evidence reference for the proposed chain.

W-018 and W-017 together enable a new capability: the engine enriches discovery candidates through autonomous contemplation at checkpoint and can generate proposal structures without operator-crafted connective or object.

A new demo is needed to make this distinction observable and falsifiable.

Decision

core demo learning-arc (evals/learning_arc/run_demo.py) scripts five scenes:

S1 — Cold session: ChatRuntime(auto_contemplate=True, engine_state_path=tmpdir) turns with an ungrounded prompt. Checkpoint enriches the emitted candidate via contemplate() and persists to engine_state/discovery_candidates.jsonl.
S2 — Checkpoint enrichment: Read the persisted candidate. Assert it carries polarity, claim_domain, and sub_questions populated by contemplate(). Assert the engine's _decompose() enumerated (narrative, cause, reveals, meaning) as a candidate chain from existing corpus shapes.
S3 — Engine-authored proposal: Build the full chain candidate using the engine-derived connective (reveals) and object (meaning) from _decompose() output. Add the corpus evidence reference (cause_creation_reveals_meaning) that the engine found as the shape template. propose_from_candidate with source.kind="contemplation". Replay gate runs.
S4 — Operator ratifies: accept_proposal against a transient corpus. Active corpus is byte-identical before and after. Provenance: adr-0057:discovery_promoted.
S5 — Session 2 grounded: Same prompt against transient corpus → grounding_source == "teaching", surface contains subject / connective / object.

The distinction from learning-loop

	learning-loop	learning-arc
Connective source	operator	engine (_decompose)
Object source	operator	engine (_decompose)
Evidence ref	operator	engine (corpus shape match)
`source.kind`	`"operator"`	`"contemplation"`
Operator action	author + ratify	ratify only

Trust boundary

Writes only to tempfile.mkdtemp() directories (engine state, proposal log, transient corpus)
Active corpus on disk is byte-identical before and after (active_corpus_byte_identical asserted)
No LLM calls, no stochastic sampling, no approximation

Falsifiable claims

test_learning_arc_demo.py (11 tests) pins:

learning_arc_closed — before grounding_source ≠ "teaching", after == "teaching"
active_corpus_byte_identical — no corpus mutation
engine_chain_found in S2 — decomposition found (narrative, cause, reveals, meaning)
source_kind == "contemplation" in S3
replay_equivalent in S3 — replay gate passed, no regression
transient_lines_after == transient_lines_before + 1 in S4
before["surface"] != after["surface"] — measurable change on same prompt

3.3 KiB Raw Blame History

ADR-0152 — Learning-Arc Demo (core demo learning-arc)