docs(session): record substrate affordance foundation pivot
This commit is contained in:
Shay
parent
9fcc029a13
commit
6d827eb433
1 changed files with 334 additions and 0 deletions
|
|
@ -0,0 +1,334 @@
|
|||
# Session 2026-06-19 — Semantic substrate affordances, contract readiness, and the foundation audit pivot
|
||||
|
||||
**Status:** recorded for continuation. **Headline:** After #829–#831 made `ProblemFrame` operational and measurable, the session clarified that the next foundation problem is not another local parser or premature organ migration. The core design target is a **semantic-substrate affordance loop**: words and chunks probe the substrate, surface nearby constructional meanings and relation families, bind roles to exact evidence spans, and only then allow `ContractAssessment` / organs to determine or refuse.
|
||||
|
||||
`Docs-only record. No runtime, serving, eval report, pack, corpus, policy, identity, or sealed artifact changes are made by this session note.`
|
||||
|
||||
---
|
||||
|
||||
## TL;DR
|
||||
|
||||
1. **#831 changed the truth metric.** The project moved from surface/schema recognition to measurable contract readiness: typed mentions, bindings, bound relations, bound question targets, and pure `ContractAssessment` diagnostics. #831's merged evidence recorded `candidate contracts = 42/423`, `runnable contracts = 1/1`, serving still `train 30/20/0`, `holdout 5/495/0`, with `wrong_ids=[]`.
|
||||
2. **GPT-5.5 Thinking XHIGH rejected premature `percent_partition` migration.** The decision memo concluded that readiness was still obligation-loose: broadening extraction or migrating `percent_partition` would preserve or amplify false readiness. The recommended next implementation is proportional-change contract closure: bind `decrease_to_fraction`, delta targets, and organ-specific blockers while correcting the false runnable holdout.
|
||||
3. **The constitutional program was refined, not adopted monolithically.** The correct split is three layers: a small constitutional kernel (`ADR-0223` later), a mutable critical-program roadmap, and versioned evidence profiles. The canonical spine must use semantic roles, not force `PropositionGraph` into a universal reasoning carrier; ADR-0144 distinguishes articulation `PropositionGraph`, epistemic `EpistemicGraph`, and math problem state.
|
||||
4. **The operator's design correction became the key doctrine.** Dead giveaways are not isolated trigger words. They are **constructional affordances**. A chunk such as `buys 3 more apples` should probe the substrate and surface acquisition / transaction / positive possession-delta affordances, then bind actor, quantity, item, prior state, optional cost, and target.
|
||||
5. **CGA semantic nearness remains essential but not sufficient.** Substrate proximity should propose candidate nodes and relation families (`buy ≈ purchase ≈ acquire ≈ gain`), while exact span bindings and contracts determine whether anything can be asserted. The governing rule: **closeness proposes; bindings ground; contracts determine.**
|
||||
6. **A foundation audit is required before broad substrate work.** The current risk is accumulating local recognizers and false contract candidates instead of building the intended semantic substrate. A new ADR is required to mandate a deep, repository-wide substrate-alignment audit before agents make broad foundational changes.
|
||||
|
||||
---
|
||||
|
||||
## Context entering the session
|
||||
|
||||
The active stack after the latest merged work:
|
||||
|
||||
| PR | Meaning |
|
||||
|---|---|
|
||||
| #829 | Kernel Substrate Tranche 1: scalar equivalence, unit dimensions, ambiguity hazards, process-frame schemas, `ProblemFrame` skeleton, morphology foundations. |
|
||||
| #830 | `ProblemFrame` operationalization and legacy-deprecation guardrails. |
|
||||
| #831 | Span-grounded mentions/bindings/relations, bound question targets, `ContractAssessment`, adequacy reporting, and contract-gap morphology recommendations. |
|
||||
|
||||
#831's evidence established that readiness must be measured by **runnable contracts**, not by recognized schema presence. The merged PR body recorded:
|
||||
|
||||
```text
|
||||
Train 50:
|
||||
entity mention present: 50
|
||||
quantity binding present: 46
|
||||
bound process relation present: 16
|
||||
bound question target present: 42
|
||||
candidate contracts: 42
|
||||
runnable contracts: 1
|
||||
|
||||
Holdout 500:
|
||||
entity mention present: 494
|
||||
quantity binding present: 452
|
||||
bound process relation present: 124
|
||||
bound question target present: 402
|
||||
candidate contracts: 423
|
||||
runnable contracts: 1
|
||||
```
|
||||
|
||||
This means CORE now has a measurable bridge from text into typed readiness, but the bridge is thin. It proves the method, not broad serving readiness.
|
||||
|
||||
---
|
||||
|
||||
## The XHIGH decision memo
|
||||
|
||||
The next-decision prompt asked GPT-5.5 Thinking XHIGH to choose the highest-leverage PR after #829–#831. The decision memo recommended:
|
||||
|
||||
```text
|
||||
feat(kernel): close proportional-decrease contracts and make readiness obligation-sound
|
||||
branch: codex/problemframe-proportional-change-closure
|
||||
```
|
||||
|
||||
The reasons:
|
||||
|
||||
- Current readiness blockers are not yet sufficiently precise.
|
||||
- `percent_partition` readiness is too loose: candidates can arise from broad `partition`/`consumption` recognition rather than full topology and target-direction proof.
|
||||
- The existing false runnable holdout is an inverse-reconstruction case outside the forward `percent_partition` contract.
|
||||
- `fraction_decrease` has mature confusers and a clear train case (`0005`) that can prove a second genuine ProblemFrame-runnable contract without serving migration.
|
||||
|
||||
Expected diagnostic movement:
|
||||
|
||||
```text
|
||||
Train runnable contracts: 1 → 2
|
||||
- 0005: fraction_decrease
|
||||
- 0046: percent_partition
|
||||
|
||||
Holdout runnable contracts: 1 → 0
|
||||
- remove the false positive; this is a truthfulness improvement, not a regression.
|
||||
```
|
||||
|
||||
Hard boundary: no serving integration, no derivation-organ migration, no `math_candidate_graph.py` change, no `report.json`, no sealed mutation, and no raw-text fallback behind a `ProblemFrame` call.
|
||||
|
||||
---
|
||||
|
||||
## The constitutional-program correction
|
||||
|
||||
A broader AGI-capable constitutional plan was discussed. The useful direction was kept, but two important corrections were made:
|
||||
|
||||
### 1. Split constitution, roadmap, and evidence profiles
|
||||
|
||||
The correct structure is:
|
||||
|
||||
```text
|
||||
1. Small constitutional ADR kernel.
|
||||
2. Mutable critical-program roadmap.
|
||||
3. Versioned evidence profiles with frozen benchmarks/thresholds.
|
||||
```
|
||||
|
||||
This keeps benchmark churn and implementation sequencing from rewriting the constitution. A future claim such as `agi-capable-v1` should be reproducible against a frozen evidence profile, not embedded forever in the constitutional core.
|
||||
|
||||
### 2. Do not make `PropositionGraph` universal
|
||||
|
||||
The first constitutional spine over-named `PropositionGraph`. ADR-0144 already separates:
|
||||
|
||||
- `generate.graph_planner.PropositionGraph` — articulation planner.
|
||||
- `recognition.carrier.EpistemicGraph` — epistemic/provenance carrier.
|
||||
- `MathProblemGraph` / math state — mathematical reasoning state.
|
||||
|
||||
The corrected constitutional spine uses semantic roles:
|
||||
|
||||
```text
|
||||
observation
|
||||
→ grounded evidence
|
||||
→ domain frame
|
||||
→ admissible domain model
|
||||
→ verified derivation/proof
|
||||
→ epistemic standing
|
||||
→ articulation plan or authorized action
|
||||
```
|
||||
|
||||
For the current math path, the realization is narrower:
|
||||
|
||||
```text
|
||||
text
|
||||
→ KernelFacts
|
||||
→ ProblemFrame
|
||||
→ ContractAssessment
|
||||
→ contract-backed derivation organ
|
||||
→ independently verified resolution
|
||||
→ existing admission and articulation path
|
||||
```
|
||||
|
||||
This preserves ADR-0207's anti-universal-IR posture: use typed adapters between domain structures; do not introduce another universal reader/graph by accident.
|
||||
|
||||
---
|
||||
|
||||
## The operator's substrate doctrine correction
|
||||
|
||||
The operator challenged a simplified explanation that treated `buys` as a generic cue for addition. The corrected doctrine is:
|
||||
|
||||
```text
|
||||
Dead giveaways are constructional affordances, not isolated trigger words.
|
||||
```
|
||||
|
||||
A word such as `buys` does not merely mean `add`. It probes a semantic neighborhood:
|
||||
|
||||
```text
|
||||
buy / bought / purchase / acquire / get / gain / receive
|
||||
→ acquisition
|
||||
→ transaction
|
||||
→ possession change
|
||||
→ positive object delta
|
||||
→ optional money/cost delta
|
||||
```
|
||||
|
||||
But the meaningful unit is often the chunk:
|
||||
|
||||
```text
|
||||
buys 3 more apples
|
||||
```
|
||||
|
||||
That chunk should surface:
|
||||
|
||||
```text
|
||||
surface chunk: "buys 3 more apples"
|
||||
semantic neighborhood: buy, purchase, acquire, gain possession, transaction
|
||||
construction: actor buys quantity more entity
|
||||
roles:
|
||||
buyer
|
||||
acquired_quantity = 3
|
||||
acquired_entity = apples
|
||||
same_entity_continuation = true
|
||||
object_delta = positive
|
||||
optional roles:
|
||||
price
|
||||
seller
|
||||
money_delta
|
||||
operator candidates:
|
||||
add_to_inventory
|
||||
compute_cost
|
||||
compute_remaining_money, if money target is present
|
||||
hazards:
|
||||
entity mismatch
|
||||
target mismatch
|
||||
negation
|
||||
comparison rather than acquisition
|
||||
```
|
||||
|
||||
The same principle applies to proportional change:
|
||||
|
||||
```text
|
||||
decrease to 3/4 of 84 degrees
|
||||
```
|
||||
|
||||
should surface:
|
||||
|
||||
```text
|
||||
semantic affordance: proportional state change
|
||||
construction: state decreases to fraction of base
|
||||
roles:
|
||||
base quantity = 84 degrees
|
||||
target scale = 3/4
|
||||
relation = decrease_to_fraction
|
||||
question target:
|
||||
delta decrease, if asked "how much did it decrease by"
|
||||
blocked cases:
|
||||
final-value question
|
||||
affine-more-than fraction
|
||||
multiple fractions
|
||||
ambiguous base
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Working model: substrate-mediated cognition loop
|
||||
|
||||
The intended loop is:
|
||||
|
||||
```text
|
||||
chunk enters the field
|
||||
→ nearby semantic / constructional affordances resonate
|
||||
→ candidate nodes and relation families surface
|
||||
→ roles bind to exact evidence spans
|
||||
→ organ-specific contracts assess obligations
|
||||
→ determination, refusal, articulation, or action follows
|
||||
```
|
||||
|
||||
Or shorter:
|
||||
|
||||
```text
|
||||
chunk
|
||||
→ semantic neighborhood
|
||||
→ constructional affordance
|
||||
→ role obligations
|
||||
→ bound relation candidate
|
||||
→ contract assessment
|
||||
```
|
||||
|
||||
This is the bridge between the CGA substrate and the typed problem-solving stack.
|
||||
|
||||
The CGA inner product / substrate should provide **semantic gravity**: related meanings and constructions become easy to retrieve. But closeness does not license truth. Truth promotion requires grounded bindings and contracts.
|
||||
|
||||
Canonical rule:
|
||||
|
||||
```text
|
||||
closeness proposes;
|
||||
bindings ground;
|
||||
contracts determine.
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Foundation risk identified
|
||||
|
||||
The operator's warning was accepted: if the project builds a pile of local recognizers, it will become harder each sprint. Every later capability will fight earlier shortcuts.
|
||||
|
||||
The dangerous path:
|
||||
|
||||
```text
|
||||
surface keyword
|
||||
→ local regex
|
||||
→ one-off organ behavior
|
||||
→ patch more confusers forever
|
||||
```
|
||||
|
||||
The intended path:
|
||||
|
||||
```text
|
||||
surface chunk
|
||||
→ substrate affordance retrieval
|
||||
→ constructional role binding
|
||||
→ ProblemFrame relation
|
||||
→ ContractAssessment obligations
|
||||
→ verified derivation or refusal
|
||||
```
|
||||
|
||||
Therefore, before broad foundation changes, CORE needs a deep audit of how the current codebase aligns or misaligns with the intended substrate-mediated cognition loop.
|
||||
|
||||
---
|
||||
|
||||
## New ADR required
|
||||
|
||||
This session produced the requirement for a new ADR:
|
||||
|
||||
```text
|
||||
ADR-0223: Semantic Substrate Affordance Audit and Foundation Alignment
|
||||
```
|
||||
|
||||
The ADR must require a deep-dive research pass over small and large details that may be blocking the intended design, including but not limited to:
|
||||
|
||||
- lexical vs constructional affordances;
|
||||
- local regex recognizers vs substrate-backed constructional retrieval;
|
||||
- ProblemFrame binding quality;
|
||||
- false runnable contracts;
|
||||
- relation topology and question-target direction;
|
||||
- graph/carrier boundaries (`ProblemFrame`, `EpistemicGraph`, `PropositionGraph`, math state);
|
||||
- legacy parser bypasses;
|
||||
- substrate pack coverage;
|
||||
- morphology/planner recommendation evidence;
|
||||
- proof/admission boundaries;
|
||||
- exact recall / CGA-neighborhood usage;
|
||||
- places where the code's foundation makes later correctness harder.
|
||||
|
||||
The ADR should authorize research and recommendations, not immediate destructive edits. Any proposed removal, replacement, or consolidation must return as evidence-backed proposals and separate implementation PRs.
|
||||
|
||||
---
|
||||
|
||||
## Immediate plan of record
|
||||
|
||||
1. Continue with the already-selected implementation PR:
|
||||
|
||||
```text
|
||||
feat(kernel): close proportional-decrease contracts and make readiness obligation-sound
|
||||
branch: codex/problemframe-proportional-change-closure
|
||||
```
|
||||
|
||||
2. Record this session and ADR as docs-only work.
|
||||
3. Keep `percent_partition` deferred until its topology, target direction, and heterogeneous positive evidence are sufficient for ProblemFrame-only migration.
|
||||
4. Use the new ADR to guide a high-skill research/audit effort, not to authorize unreviewed rewrites.
|
||||
|
||||
---
|
||||
|
||||
## Open / next
|
||||
|
||||
- Run or commission the substrate-affordance audit required by ADR-0223.
|
||||
- Ensure the proportional-decrease PR interprets `decrease to <fraction> of <state>` as a constructional affordance, not as keyword matching.
|
||||
- Update future constitutional documents to use semantic roles rather than concrete universal graph classes.
|
||||
- Consider adding the doctrine below to architecture docs after ADR-0223 is ratified:
|
||||
|
||||
```text
|
||||
Words and chunks are probes into the semantic substrate.
|
||||
They surface candidate affordances and relation families.
|
||||
Only grounded bindings and contracts can promote those affordances into determinations.
|
||||
```
|
||||
Loading…
Reference in a new issue