amr/hermes-brain
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

per-domain flip: knowledge types and fastest acquisition

Browse Source 
- Sufficiency flip is per-domain, not global. Poetry never flips.
- Three knowledge types: structural (published rules), empirical
  (observations), performance (profiling data)
- Fastest acquisition: active sandboxed probing, contrastive queries
  to human (not waiting for HITL to accumulate), ontology transfer
  from related domains, benchmark harness
- Codified domain: flip within days (hours LLM + hours expert review)
- Uncodified learnable domain: flip within weeks (probe + real use)
- Never-flip domains: system is honest, LLM handles 100%
This commit is contained in:
Hermes
2026-05-21 18:55:11 +00:00
parent b5d59c3360
commit 2d0d6d478a
1 changed files with 71 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
ideas/passepartout-economics.org
View File

@@ -779,6 +779,77 @@ whether the bootstrap succeeds. A patient operator with a
destination as an operator with bottomless API credits — the
second arrives faster, but both arrive.
** The per-domain sufficiency flip
The sufficiency flip is not a single event. It happens
independently for each domain, and some domains never flip.
The flip point is determined by the kind of knowledge the
domain requires.
*** Knowledge types required for a flip
| Domain | Knowledge required | Can it flip? | How fast? |
|--------|-------------------|--------------|-----------|
| Shell safety, path rules | Structural only — the deployment's config, shell semantics | Immediately | Instant — ingested from config |
| Healthcare compliance | Structural (HIPAA text) + empirical (human reviews of edge cases) | Yes | Weeks — one pass of LLM translation + human review cycle |
| Codebase refactoring | Structural (dependency graph, API surface) + empirical (test suite, build results) + performance (latency, throughput) | Yes | Months — depends on how many build cycles the system has observed |
| Microcode optimization | Structural (RISC-V ISA, core topology) + performance (profiling data) | Yes | Weeks — after enough benchmark runs to characterize hardware |
| Poetry, creative writing | Neither structural rules nor empirical ground truth (beauty is subjective) | **Never** | N/A — the gate stack cannot verify aesthetic quality |
| Novel scientific discovery | Structural (known laws) + empirical (experiments) | Eventually | Years — requires experimental data the system must gather through instruments |
*** The fastest acquisition strategy per domain
The goal: reach the flip point with the fewest calendar days
and the fewest human hours.
| Knowledge type | Acquisition strategy | Calendar time | Human time |
|----------------|---------------------|---------------|------------|
| **Structural** (published rules, configs, specs) | LLM translation of source documents + ACL2 consistency verification + one-shot human review | Hours for the LLM pass, days for human review | Days — one domain expert reviewing the output |
| **Structural** (unpublished — your deployment, your codebase) | Automated scanning — the system walks your filesystem, reads your configs, builds the dependency graph | Minutes to hours | Zero — fully automated |
| **Empirical** (what happens when X?) | **Active probing** — the system does not wait for user interactions. It probes its own environment: runs shell commands in sandbox to verify gate rules, executes test suites to verify dependency graph, measures what happens when it pushes boundaries. | Hours to days | Zero — automated sandboxed probing |
| **Empirical** (what does the human prefer?) | **Contrastive queries** — instead of waiting for HITL approvals to accumulate, the system asks targeted questions: "Which of these two interpretations of the regulation is correct? This one or this one?" Each question produces a fact. | Days — batch of targeted questions answered in one session | Hours — one review session with the domain expert |
| **Performance** (latency, throughput) | **Benchmark harness** — the system runs its own workload at varying parameters, records timing data, stores it as facts with `:provenance :benchmark` | Hours — automated sweep | Zero |
| **Transfer** from related domains | **Ontology alignment** — if the system already knows compliance for GDPR, it can ask Screamer: "Which GDPR rules have the same structure as HIPAA rules? Use those as seed hypotheses, flag for human review." The transferred rules flip faster because Screamer starts from a consistent foundation, not from scratch. | Days — Screamer finds alignments automatically, human reviews the suggestions | Hours — review the cross-domain alignment suggestions |
*** The fastest path to flip any domain
1. **Ingest all published text** for the domain (laws, specs, configs)
via LLM translation. One pass. Hours.
2. **Run the benchmark harness** to measure the system's own performance
in the domain. One sweep. Hours.
3. **Run active sandboxed probes** — test each gate rule against
synthetic inputs to verify it behaves as expected. Automated.
4. **Generate contrastive queries** — Screamer identifies the 5%
of rules where the LLM's translation is most uncertain (contradicts
a transferred rule, has no precedent, has multiple valid
interpretations). Present these as yes/no questions to the human
domain expert in a single session.
5. **Start serving real interactions.** Every gate outcome generates
an empirical fact that Screamer feeds back into the rule set.
The empirical loop tightens from the first real interaction.
For a codified domain (healthcare compliance, financial regulation,
industrial safety): flip within days of step 1-4. The only bottleneck
is the domain expert's review session in step 4 — a few hours of
human time.
For an uncodified but learnable domain (codebase refactoring):
flip within weeks of step 3 (benchmark harness) + step 5 (real
interactions). No LLM translation needed — the knowledge comes
from the system probing its own environment.
For a domain that can never flip (poetry, aesthetics): the system
never reaches sufficiency. It never claims to. The TUI shows
"Symbolic index: 0 facts. This domain has no codifiable rules."
The LLM handles 100% of poetry interactions. The gate stack
only checks for safety (no shell commands, no file deletions)
and passes through everything else to the LLM. The system is
honest about its frontier.
Large refactoring projects (extract module, rename API, split monolith)
are the hardest test for any AI agent. Current approaches (Claude Code,
Copilot) handle them probabilistically — every step costs tokens, and