core-labs/core
core-labs/core
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

docs(arc): curriculum-license-loop plan of record + tier division of work #115

Merged
core-labs merged 1 commit from docs/curriculum-license-loop-plan into main 2026-07-25 22:17:50 +00:00
Conversation 0 Commits 1 Files changed 2 +444
core-labs commented 2026-07-25 22:15:23 +00:00
Owner
Copy link

Two docs, no code. Opens the next arc and records the tier split so a cold session can pick up a unit without re-deriving the analysis.

The finding that makes this arc necessary

Two prior arcs closed on "the binding constraint is ratified curriculum volume, not machinery." Verified against code — the machinery half is wrong in a load-bearing way:

  • evals/curriculum_serve/practice/ does not exist. No producer folds run_practice into per-band tallies.
  • seal_ledger exists only for deduction (evals/deduction_serve/practice/runner.py:96), yet chat/curriculum_serve_license.py's own docstring names evals.curriculum_serve.practice.runner.seal_ledger as "the only writer" of the artifact it reads. That writer was never built, and chat/data/curriculum_serve_ledger.json does not exist.
  • The ceremony stops short: RatificationReceipt.pending_stages = ("arena_queue_entry", "ledger_reseal").

So the pipeline is author → ratify → ??? → ??? → license → serve. Authoring curriculum today cannot earn a license at any volume. Content is necessary and currently insufficient.

The build is still modest: evals/curriculum_serve/oracle.py is already an independent decision procedure usable as the gold tether, and ADR-0263 already extracted the sealers for exactly a third instance.

Two measured facts that shape the content work

  • refuted has never been produced. Physics lane gold is entailed 14 / unknown 12 / declined 6. The verdict exists in CurriculumDecision.verdict; no corpus row can express it.
  • Bands key on (domain, operator_family) (chat/curriculum_surface.py:184), so negative content must live inside operator_family: "modal" with a row-level polarity marker. A separate modal_negative family would create a new band at n=0 instead of adding refuted volume to the target band — the easiest available way to waste the whole authoring effort.

The tiering criterion, and why it isn't difficulty

The evidence doesn't support a prestige split. In the same 48-hour window, Sonnet 5 found the DEFAULT_SINK bug that wrote outside the repository and did the six-worktree public_demo archaeology; Opus 5 mis-diagnosed the CGA hot path by multiplying a microbenchmark by a call count. Model tier wasn't the variable — verification discipline was.

So units split on one question: would an error here be caught by a gate, or is it silent?

Two units are silent-failure because they are themselves the gate:

  1. Negative-curriculum epistemology — if wrong, the corpus encodes falsehoods and wrong=0 still passes, because the lane's gold and oracle.py both read the same rows. Structurally unfalsifiable by every gate this repo has.
  2. Volume-honesty invariant — if wrong, conservative_floor cannot distinguish 657 independent facts from 16 asked 41 times, so the ledger clears θ=0.99 having earned nothing, and the license reads identical to a real one.

Those two get Opus. Everything with a loud failure mode goes to the cheapest competent tier: producer implementation (gated by the invariant tests), the reseal verb, the assert_corpus_sound rename, the R7 code move, the Rust parity measurement → Sonnet. The 312-file ADR vocabulary sweep → Haiku, acceptance criterion being that the widened pin goes green.

The brief also carries an escalation section — five conditions under which a cheaper tier must stop and hand back up, framed as the mechanism working rather than a failure.

Ordering correction to my own earlier plan

The epistemology ADR must precede the generator. Building Phase C first means building it twice, since polarity determines what the generator enumerates. Since the ADR is design-only, doing it first costs nothing in parallelism.

[Verification]: docs-only, no code paths touched. Every referenced path confirmed to exist, and all five line citations confirmed accurate at HEAD — evals/deduction_serve/practice/gold.py:1163, evals/curriculum_serve/runner.py:71, teaching/discovery_yield.py:67-70, generate/proof_chain/entail.py:125, chat/curriculum_surface.py:184. Pre-push gate ran on push (smoke 555 / deductive 285 baselines unaffected).

Not merged, per standing policy.

Two docs, no code. Opens the next arc and records the tier split so a cold session can pick up a unit without re-deriving the analysis. ## The finding that makes this arc necessary Two prior arcs closed on *"the binding constraint is ratified curriculum volume, not machinery."* Verified against code — the machinery half is wrong in a load-bearing way: - **`evals/curriculum_serve/practice/` does not exist.** No producer folds `run_practice` into per-band tallies. - **`seal_ledger` exists only for deduction** (`evals/deduction_serve/practice/runner.py:96`), yet `chat/curriculum_serve_license.py`'s own docstring names `evals.curriculum_serve.practice.runner.seal_ledger` as *"the only writer"* of the artifact it reads. **That writer was never built**, and `chat/data/curriculum_serve_ledger.json` does not exist. - **The ceremony stops short:** `RatificationReceipt.pending_stages = ("arena_queue_entry", "ledger_reseal")`. So the pipeline is `author → ratify → ??? → ??? → license → serve`. **Authoring curriculum today cannot earn a license at any volume.** Content is necessary and currently insufficient. The build is still modest: `evals/curriculum_serve/oracle.py` is already an independent decision procedure usable as the gold tether, and ADR-0263 already extracted the sealers for exactly a third instance. ## Two measured facts that shape the content work - **`refuted` has never been produced.** Physics lane gold is `entailed 14 / unknown 12 / declined 6`. The verdict exists in `CurriculumDecision.verdict`; no corpus row can express it. - **Bands key on `(domain, operator_family)`** (`chat/curriculum_surface.py:184`), so negative content must live *inside* `operator_family: "modal"` with a row-level polarity marker. A separate `modal_negative` family would create a **new band at n=0** instead of adding refuted volume to the target band — the easiest available way to waste the whole authoring effort. ## The tiering criterion, and why it isn't difficulty The evidence doesn't support a prestige split. In the same 48-hour window, Sonnet 5 found the `DEFAULT_SINK` bug that wrote outside the repository and did the six-worktree `public_demo` archaeology; Opus 5 mis-diagnosed the CGA hot path by multiplying a microbenchmark by a call count. **Model tier wasn't the variable — verification discipline was.** So units split on one question: **would an error here be caught by a gate, or is it silent?** Two units are silent-failure *because they are themselves the gate*: 1. **Negative-curriculum epistemology** — if wrong, the corpus encodes falsehoods and `wrong=0` still passes, because the lane's gold and `oracle.py` both read the same rows. Structurally unfalsifiable by every gate this repo has. 2. **Volume-honesty invariant** — if wrong, `conservative_floor` cannot distinguish 657 independent facts from 16 asked 41 times, so the ledger clears θ=0.99 having earned nothing, and the license reads identical to a real one. Those two get Opus. Everything with a loud failure mode goes to the cheapest competent tier: producer implementation (gated by the invariant tests), the `reseal` verb, the `assert_corpus_sound` rename, the R7 code move, the Rust parity measurement → Sonnet. The 312-file ADR vocabulary sweep → Haiku, acceptance criterion being that the widened pin goes green. The brief also carries an **escalation section** — five conditions under which a cheaper tier must stop and hand back up, framed as the mechanism working rather than a failure. ## Ordering correction to my own earlier plan The epistemology ADR must precede the generator. Building Phase C first means building it twice, since polarity determines what the generator enumerates. Since the ADR is design-only, doing it first costs nothing in parallelism. [Verification]: docs-only, no code paths touched. Every referenced path confirmed to exist, and all five line citations confirmed accurate at `HEAD` — `evals/deduction_serve/practice/gold.py:1163`, `evals/curriculum_serve/runner.py:71`, `teaching/discovery_yield.py:67-70`, `generate/proof_chain/entail.py:125`, `chat/curriculum_surface.py:184`. Pre-push gate ran on push (smoke 555 / deductive 285 baselines unaffected). Not merged, per standing policy.
core-labs added 1 commit 2026-07-25 22:15:25 +00:00 
Two prior arcs closed on "the binding constraint is ratified curriculum volume,
not machinery." Verified against code 2026-07-25: the machinery half is wrong in
a load-bearing way, and the correction is the reason this arc exists.

evals/curriculum_serve/practice/ does not exist. seal_ledger exists only for
deduction (evals/deduction_serve/practice/runner.py:96), yet
chat/curriculum_serve_license.py's own docstring names
evals.curriculum_serve.practice.runner.seal_ledger as "the only writer" of the
artifact it reads -- that writer was never built, and
chat/data/curriculum_serve_ledger.json does not exist. And the ceremony stops
short: RatificationReceipt.pending_stages = (arena_queue_entry, ledger_reseal).

So the pipeline is author -> ratify -> ??? -> ??? -> license -> serve. Authoring
curriculum today cannot earn a license at ANY volume. Content is necessary and
currently insufficient. The build is still modest, because oracle.py is already
an independent decision procedure usable as the gold tether and ADR-0263 already
extracted the sealers for exactly a third instance.

Two measured facts shape the content work. `refuted` has NEVER been produced --
the physics lane gold is entailed 14 / unknown 12 / declined 6, and no corpus row
can express it. And bands key on (domain, operator_family), so negative content
must live INSIDE operator_family "modal" with a row-level polarity marker; a
separate modal_negative family would create a new band at n=0 instead of adding
refuted volume to the target band -- the easiest available way to waste the whole
authoring effort.

DIVISION-OF-WORK.md records the tiering criterion and why it is not difficulty.
The evidence does not support a prestige split: in the same 48h window Sonnet 5
found the DEFAULT_SINK bug that wrote outside the repository and did the
six-worktree public_demo archaeology, while Opus 5 mis-diagnosed the CGA hot path
by multiplying a microbenchmark by a call count. Model tier was not the variable;
verification discipline was.

The split is therefore: would an error here be caught by a gate, or is it silent?
Two units are silent-failure BECAUSE THEY ARE THE GATE -- the negative-curriculum
epistemology (wrong => corpus encodes falsehoods and wrong=0 still passes, since
lane gold and the oracle both read the same rows) and the volume-honesty invariant
(wrong => conservative_floor cannot tell 657 independent facts from 16 asked 41
times, so the ledger clears theta=0.99 having earned nothing). Those get Opus.
Everything loud goes to the cheapest competent tier.

Also records the ordering correction: the epistemology ADR must precede the
generator, or the generator gets built twice.

[Verification]: docs-only; no code paths touched. Every referenced path confirmed
to exist and all five line citations confirmed accurate at HEAD (gold.py:1163,
curriculum_serve/runner.py:71, discovery_yield.py:67-70, entail.py:125,
curriculum_surface.py:184). Pre-push gate runs on push.
core-labs merged commit 6ada6f7a03 into main 2026-07-25 22:17:50 +00:00
core-labs deleted branch docs/curriculum-license-loop-plan 2026-07-25 22:17:50 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: core-labs/core#115
No description provided.
Delete branch "docs/curriculum-license-loop-plan"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?