G6 engineers its own gains.

AI is usually judged by public benchmarks, and a high score there often fails to carry over to real work. So G6 never takes a score on trust: every change it makes is tested on tasks held out from it.

Every study here asks one question: when G6 engineers the harness around a model for a task, how much better does the same model do? It counts every study run, the earliest included, and says which part of G6 did the work. Together they are a worked example, built to show what the system can do; on your own work, results will vary with the task and the data you have. How to read them

Studies run
24
On the core harness
12
Core with extensions
9
On extensions
3
Showed a gain
18

Bootstrapping a small model through its harness

llama3.1:8b, an 8-billion-parameter model running on one computer with Ollama, on 100 fresh held-out tasks per dataset, in a study registered before it ran. Larger models engineer the harness: Claude Code built G6's components and its diagnosis of failures, and in the development rounds a 20-billion-parameter model reflects on the small model's mistakes to propose changes. The small model then runs inside that harness alone, with no weights changed. That is G6's core idea: a larger model's judgement, distilled into a smaller model's harness.

Grey is the model alone; black is G6. Each large number is G6's winning option for that task set, named beneath it, chosen after the fact from the options measured; in use, G6 chooses between options on your own held-out examples.

2281

Winning option: G6, self-tuned

GSM8K
The model alone 22 of 100
G6 as built 20 of 100
G6, every component on 44 of 100
G6, a person's prompt 78 of 100
G6, random changes, same gate 20 of 100
G6, self-tuned 81 of 100

6577

Winning option: G6, random changes, same gate

MASSIVE
The model alone 65 of 100
G6 as built 60 of 100
G6, every component on 73 of 100
G6, a person's prompt 63 of 100
G6, random changes, same gate 77 of 100
G6, self-tuned 60 of 100

5273

Winning option: G6, self-tuned

SciFact
The model alone 52 of 100
G6 as built 50 of 100
G6, every component on 66 of 100
G6, a person's prompt 67 of 100
G6, random changes, same gate 50 of 100
G6, self-tuned 73 of 100

5463

Winning option: G6, every component on

MBPP
The model alone 54 of 100
G6 as built 52 of 100
G6, every component on 63 of 100
G6, a person's prompt 49 of 100
G6, random changes, same gate 52 of 100
G6, self-tuned 52 of 100
  • Exploratory upper bound: selecting the winning option after observing each task set, G6 solved 294 of 400 tasks the model alone solved 193 of, and every winning option beat the model alone task by task (the table below).
  • The gain is the harness, not more compute: spending the same tokens on voting took GSM8K from 22 to 24; the harness G6 engineered took it to 81, at one model call per task, like the model alone.
  • With every component switched on, G6 beat the model alone on all 4 datasets (246 of 400).

The winning approach follows the task's structure.

These options were selected after evaluation. Their unadjusted paired p-values do not validate a deployable selection policy. The fixed self-tuning arm D solved 266 of 400; report it separately.

Task setThe taskWinning approach Model alone → G6Tasks gained, lost
GSM8KMulti-step arithmetic word problems Self-tuning: the model shows its working step by step, with the answer on its own last line 22 → 81 of 100 62 gained, 3 lost
MASSIVEPick one of 60 similar intent labels A prebuilt component: eight labelled examples, found by G6's search and passed by its gate 65 → 77 of 100 18 gained, 6 lost
SciFactJudge a scientific claim against one abstract Self-tuning: the model judges the claim from the abstract alone, not outside knowledge 52 → 73 of 100 26 gained, 5 lost
MBPPWrite Python that passes hidden tests Every prebuilt component: self-review, test-and-repair, an import check and eight labelled examples 54 → 63 of 100 12 gained, 3 lost

Where one instruction fixes the main failure, self-tuning finds it, and its gate confirms the gain quickly: on GSM8K, asking for the working step by step, with the answer on its own last line, was enough. Where the gain comes from structure, examples of sixty similar labels or a loop that runs the code against tests, it is spread across more tasks, and the gate needs more evidence: self-tuning did propose labelled examples for MASSIVE and test-and-repair for MBPP, but with 70 held-out tasks its gate could not yet confirm them. G6's prebuilt components are those layers ready-made. Switched on together, they beat self-tuning task by task on MASSIVE (17 gained, 4 lost) and MBPP (12 gained, 1 lost).

This supports the Swiss cheese design G6 is built on: each prebuilt component is a layer against one kind of failure, several are needed where one alone is not enough, and the structure of the task decides which layers matter. It is why G6 offers both modes, and why it tests both on your own examples.

Development round 3 measured the two modes together, and they did not stack: self-tuning that started from every prebuilt component had no change approved on any of its 4 task sets, and on SciFact every component on did worse than the model alone (57 against 64 of 100; 14 tasks lost, 7 gained). There a check asks every answer a second time with its own fixed question, so no instruction upstream could change the final label: one layer can mask another. The modes are alternatives, chosen per task on held-out evidence, not layers to switch on all at once.

How to read these results

A worked example: these studies show the power of G6's self-engineering on public tasks. They are not a forecast for yours.

What they are Claude Code working with G6 on one 8-billion-parameter local model (llama3.1:8b) and public task sets. Every score is on fresh tasks that no tuning step saw, against the same model alone. One adaptation study was registered before it ran; the development rounds were not, and the method was improved between them. Each task set's winning option is picked after the fact from the options measured.
What they show
  • Where a model fails in a systematic way, G6 can engineer a large gain in minutes, on the same model: GSM8K went from 22 to 81 of 100 in the registered study.
  • The gate decides on evidence, whoever proposes a change: random changes put through it were approved when they won on held-out tasks and refused when they did not. G6's own reflective tuning found more than random search overall (266 against 199 of 400 tasks in the registered study), but not on every task set.
  • The winning approach follows the task's structure: one instruction where a single change fixes the main failure, prebuilt components where the gain needs layers such as labelled examples or test-and-repair. The registered study supports G6's Swiss cheese design of prebuilt failure layers.
  • The two modes are alternatives, not a stack: in development round 3, self-tuning that started from every prebuilt component had no change approved, and on SciFact every component on did worse than the model alone. Which mode wins depends on the task.
  • A small model can be bootstrapped through its harness: larger models engineer it, and the small model runs inside it alone, with no weights changed.
What they do not show How G6 does with other models, on your tasks, or with independent coding agents, or that a harness tuned for one task set helps another. Nor do they show that a harness adds knowledge a model lacks: much of each gain is the model doing what it can already do, once it is asked the right way. With 100 tasks per set, a difference of a few tasks can be chance; the large gains are not. The earlier benchmark studies were audited, not run again.
Why results vary G6 can only engineer what your checks and examples can measure: it needs about 30 or more held-out examples with known answers before it will approve a change. Gains are largest where errors are systematic and a harness change addresses them, and smallest where the model is already near its ceiling or lacks the knowledge a task needs.
Status The experiments are complete: the programme closed with development round 3, and the numbers here are final.

What each component does on its own

Every single change to G6 as built that we measured, on six public datasets: tasks solved out of 100, against G6 as built. Most changes help one kind of task and hurt another. That is why G6 tests every change on your own held-out cases and keeps only what wins on yours.

Change in tasks solved, of 100, against G6 as built. Development measurement: every single-field change from G6 as built (llama3.1:8b, temperature 0, seed fixed), each split measured from a freshly loaded model; successes on round 1 dev (20) and dev_eval (100).
ChangeARC-Challengebanking77GSM8KMASSIVEMBPPSciFact
G6 as built864724686071
Prompt: “Reply with the letter of the correct option only.”−2
Prompt: “Think step by step about the science involved, then write the fi…”−6
Prompt: “Choose the single intent that best matches what the customer wan…”+4
Prompt: “Think briefly about what the customer wants, then write the fina…”+2
Prompt: “Work through the problem step by step. Then write the final answ…”+64
Prompt: “Give only the final number: no units, no words, no working.”−17
Prompt: “Choose the single intent that best matches what the user wants.…”+2
Prompt: “Think briefly about what the user wants, then write the final li…”+1
Prompt: “Reply with one Python code block that contains the complete func…”+3
Prompt: “Think about edge cases in a sentence or two, then give the compl…”−1
Prompt: “Use only the abstract, not outside knowledge. If the abstract do…”0
Prompt: “Reason briefly about what the abstract says, then write the fina…”0
8 labelled examples in the prompt−5+36+3+11−1−2
A second look at answers of “B”0
A second look at answers of “D”0
A second look at answers of “C”−1
A second look at answers of “A”−2
Review its own answer once−9+2+230−2−4
Run the shown tests and repair, up to three rounds+6
Run the shown test and repair, once+6
A second look at answers of “not enough info”+2
A second look at answers of “supports”+1
A second look at answers of “contradicts”−3
Check a “supports” verdict against the abstract−33
Extra check: the answer is one letter0
Extra check: the answer is one label000
Extra check: the answer is a whole number0
Extra check: the code imports cleanly0
Retry once when a check fails000000

Every study run

The registered studies, the development work, and G6's earlier benchmark studies with the figures their audits settled on. Each compares the same model without G6 and with it. A study of fewer than about a hundred tasks can rarely rule out chance on its own; each study's own page gives its test.

The core harness 12

  1. Self-tuning on four public datasets

    Registered Gain

    The meta-harness tuned G6 to each dataset: it proposed changes, tested each one and kept what its gate passed

    The model alone: 193 of 400 G6: 266 of 400

    tasks solved, 100 fresh held-out tasks per dataset · llama3.1:8b, local

    GSM8K 22 → 81, MASSIVE 65 → 60, SciFact 52 → 73, MBPP 54 → 52

    The study, its test and raw data →

  2. Self-tuning, first run

    Registered Gain

    The meta-harness tuned G6 to each dataset: it proposed changes, tested each one and kept what its gate passed

    The model alone: 125 of 240 G6: 153 of 240

    tasks solved, 60 held-out tasks per dataset · llama3.1:8b, local

    GSM8K 12 → 42, MASSIVE 36 → 38, SciFact 38 → 32, MBPP 39 → 41

    The study, its test and raw data →

  3. Claim checking

    Registered Gain

    G6's source check (grounding) judged each claim against its source text

    One model call: 26 of 174 G6: 0 of 174

    wrong claims called supported (negated, or with a wrong figure) (lower is better) · llama3.1:8b, local

    It confirmed fewer reworded true claims than a model call (19 against 83 of 100).

    The study, its test and raw data →

  4. Action gate

    Registered Gain

    G6's action gate decided each tool call an agent wanted to make: allow, ask a person, or block

    No gate: 0 of 30 G6: 29 of 30

    harmful agent calls stopped · AgentDojo's tool calls

    A read-only allow-list stopped all 30; the one G6 let through, a password change, now goes to a person. It held 26% of harmless calls for a person.

    The study, its test and raw data →

  5. Root-cause remedies

    Worked example Gain

    G6 diagnosed why each answer failed and applied the remedy for that cause: labelled examples, a second look, or running the shown tests and repairing

    G6 as built: 109 of 180 G6: 139 of 180

    tasks solved on MASSIVE, SciFact and MBPP · llama3.1:8b, local

    Development: measured on cases an earlier run had seen; the next registered study tested the remedies on fresh cases.

    The study, its test and raw data →

  6. Development round 3

    Worked example Gain

    G6's own loop found, tested and approved each change

    The model alone: 174 of 340 G6: 234 of 340

    tasks solved, 100 fresh cases per dataset (MBPP 40) · llama3.1:8b, local

    GSM8K 21 → 81, MASSIVE 65 → 65, SciFact 64 → 64, MBPP 24 → 24

    The study, its test and raw data →

  7. Development round 2

    Worked example Gain

    G6's own loop found, tested and approved each change

    The model alone: 288 of 500 G6: 388 of 500

    tasks solved, 100 fresh cases per dataset · llama3.1:8b, local

    GSM8K 26 → 90, MASSIVE 67 → 74, MBPP 61 → 61, ARC-Challenge 82 → 82, banking77 52 → 81

    The study, its test and raw data →

  8. Development round 1

    Worked example Gain

    G6's own loop found, tested and approved each change

    The model alone: 357 of 600 G6: 456 of 600

    tasks solved, 100 fresh cases per dataset · llama3.1:8b, local

    GSM8K 24 → 83, MASSIVE 66 → 70, SciFact 71 → 75, MBPP 62 → 60, ARC-Challenge 86 → 86, banking77 48 → 82

    The study, its test and raw data →

  9. SWE-bench Pro

    Earlier study, audited Gain

    Theory building from past attempts, with G6's tools

    The same model without G6: 27% G6: 38%

    issue resolved, one attempt each · GPT-5.5 (OpenAI's Codex)

    Significant at matched effort (p 0.0074).

  10. GPQA Diamond

    Earlier study, audited Gain

    Grounding, debate, retrieval and a computer-algebra system

    The same model without G6: 51% G6: 57.1%

    multiple-choice accuracy · qwen3.5:35b (local)

    Not significant on its own (p 0.18).

  11. BBEH

    Earlier study, audited Gain

    Formal solvers and Z3 for the tractable subtasks

    The same model without G6: 82.6% G6: 87%

    exact match · Claude Opus 4.8

    Not significant on its own (p 0.125).

  12. HLE

    Earlier study, audited Specialised

    Self-training on past attempts

    The same model without G6: 12% G6: 70.2% on the questions it trained on

    accuracy, held out · gpt-oss-120b

    Trained on the evaluation questions themselves: specialisation, not held-out accuracy. A held-out study is next.

Core with extensions 9

  1. AgentHarm

    Earlier study, audited Gain

    A safety gate with a model harm judge

    The same model without G6: 51.1% G6: 86.9%

    verdict accuracy · gpt-oss-120b (judge)

    It measured refusing harmful prompts, not blocking actions, and also stopped 25% of harmless ones.

  2. GAIA Level 2

    Earlier study, audited Gain

    MCP tools: grounding, web search, retrieval, planning

    The same model without G6: 39% G6: 68%

    exact match · Claude Opus 4.6

    Sound on audit; G6 had twice the baseline's budget.

  3. GAIA Level 1

    Earlier study, audited Gain

    MCP tools: grounding, web search, retrieval

    The same model without G6: 52% G6: 69%

    exact match · Claude Opus 4.6

    Sound on audit; G6 had twice the budget, and on its own it falls just short of significance (p 0.065).

  4. MedXpertQA

    Earlier study, audited Gain

    Retrieval from a medical corpus, with elimination prompting

    The same model without G6: 22.9% G6: 31.6%

    multiple-choice accuracy · medgemma:27b (local)

    Not significant on its own (p 0.135); the corpus's licences were never cleared.

  5. LongBench v2

    Earlier study, audited Gain

    Long-context retrieval and summarising tools

    The same model without G6: 39% G6: 45%

    multiple-choice accuracy · Claude Haiku 4.5

    The gain is coverage: G6 answered the 34 overlong items the model alone skipped; on the items both answered, the model alone did better.

  6. ARC-AGI v1

    Earlier study, audited No difference

    MCP tools

    The same model without G6: 96% G6: 96%

    task solved · Claude Opus 4.6

    The model alone was already near the ceiling.

  7. ARC-AGI v2

    Earlier study, audited No difference

    MCP tools

    The same model without G6: 94% G6: 94%

    task solved · Claude Opus 4.6

    Both near the ceiling, on 16 of 25 tasks.

  8. SkillsBench

    Earlier study, audited Lower

    Computational tools, with task classifiers in one arm

    The same model without G6: 28% G6: 24%

    task passes its tests · Claude Sonnet 4.5

    Both arms already had the skills.

  9. tau2-bench

    Earlier study, audited No paired baseline

    Theory injection and planning guidance

    The same model without G6: no paired baseline G6: 58%

    task success · gpt-oss-120b

    No paired baseline was run, so the gain cannot be measured.

Extensions 3

  1. Omni-MATH

    Earlier study, audited Gain

    Planning and reasoning architectures with a computer-algebra system

    The same model without G6: 49% G6: 58%

    exact match · Claude Opus 4.6

    The numbers held; which tool made the gain is unproven.

  2. GDPval

    Earlier study, audited Gain

    Eighteen job agents, scored by a rubric judge

    The same model without G6: 44.1% G6: 50.9%

    mean rubric score · Claude Opus 4.6; judge gpt-oss-120b

    Scored by a model judge; 2 of the 18 job agents reached 70%.

  3. ARC-AGI 3

    Earlier study, audited Specialised

    Game strategies written in advance

    The same model without G6: 3.3% G6: 44% (81 of 183) with strategies built on these games

    levels completed · Claude Opus 4.6 (writing strategies)

    The strategies were built on these same games; played fresh, G6 matched random play (6 of 183).

A worked example

The G6 meta-harness engineering itself

After the registered study, Claude Code ran G6's meta-harness in three rounds, each on fresh tasks and each published whatever it found. The method was improved after the first round, so they show what G6 can do, not a validated rate. Every task set is shown, including those where G6 changed nothing: pooled, the gain is close to the registered study's. The programme closed with round 3.

Do the two modes work together: does self-tuning that starts from every prebuilt component beat each mode alone?

GSM8K, MASSIVE, MBPP, SciFact; llama3.1:8b; 40/100 fresh cases each

The two modes did not stack. Self-tuning from G6 as built took GSM8K 21 to 81, and every prebuilt component switched on took MASSIVE 65 to 78; but self-tuning that started from every component on had no change approved on any task set. On SciFact every component on did worse than the model alone (57 against 64): it asks every answer a second time with a fixed question, so no instruction can reach the final label, and the three the reflector proposed there changed no task. The gate also refused a SciFact change that won 15 to 1 on the tasks it was chosen on and went 3 to 5 on fresh ones. MBPP had 40 fresh tasks left, too few for the gate to decide. G6 as built, measured again last, repeated all 340 outputs exactly.

GSM8K
The model alone 21 of 100
G6 as built 21 of 100
G6, every component on 54 of 100
G6, self-tuned 81 of 100
G6, self-tuned from every component on 54 of 100
MASSIVE
The model alone 65 of 100
G6 as built 65 of 100
G6, every component on 78 of 100
G6, self-tuned 65 of 100
G6, self-tuned from every component on 78 of 100
SciFact
The model alone 64 of 100
G6 as built 64 of 100
G6, every component on 57 of 100
G6, self-tuned 64 of 100
G6, self-tuned from every component on 57 of 100
MBPP
The model alone 24 of 40
G6 as built 24 of 40
G6, every component on 23 of 40
G6, self-tuned 24 of 40
G6, self-tuned from every component on 23 of 40
Root causes and raw data →

On fresh cases, how far does G6 tune itself beyond the model alone and its own other variants?

ARC-Challenge, banking77, GSM8K, MASSIVE, MBPP; llama3.1:8b; 100 fresh cases each

Self-tuning was approved on three of five datasets and gained 100 cases over G6 as built: GSM8K 26 to 90, banking77 52 to 81 and MASSIVE 67 to 74. On ARC, where the model alone already scored 82, and on MBPP, the gate refused every change it was offered. Random changes through the same gate were approved once, on banking77 (78); like self-tuning's change there, they show the model eight labelled examples. Every arm started from a freshly loaded model, and G6 as built, measured again last, repeated all 500 outputs exactly.

GSM8K
The model alone 26 of 100
G6 as built 26 of 100
G6, a person's prompt 85 of 100
G6, random changes, same gate 26 of 100
G6, self-tuned 90 of 100
MASSIVE
The model alone 67 of 100
G6 as built 67 of 100
G6, a person's prompt 47 of 100
G6, random changes, same gate 67 of 100
G6, self-tuned 74 of 100
MBPP
The model alone 61 of 100
G6 as built 61 of 100
G6, a person's prompt 57 of 100
G6, random changes, same gate 61 of 100
G6, self-tuned 61 of 100
ARC-Challenge
The model alone 82 of 100
G6 as built 82 of 100
G6, a person's prompt 76 of 100
G6, random changes, same gate 82 of 100
G6, self-tuned 82 of 100
banking77
The model alone 52 of 100
G6 as built 52 of 100
G6, a person's prompt 54 of 100
G6, random changes, same gate 78 of 100
G6, self-tuned 81 of 100
Root causes and raw data →

On fresh cases, how far does G6 tune itself beyond the model alone and its own other variants?

ARC-Challenge, banking77, GSM8K, MASSIVE, MBPP, SciFact; llama3.1:8b; 100 fresh cases each

Tuning gained 99 cases over G6 as built, and random edits from the same options gained as many: on GSM8K and banking77 one option did the work (working step by step; labelled examples), and on the other four nothing was approved. Runs that changed nothing still moved by up to 4 cases, from the model server's prompt cache; later rounds measure every arm from a freshly loaded model.

GSM8K
The model alone 24 of 100
G6 as built 23 of 100
G6, a person's prompt 78 of 100
G6, random changes, same gate 85 of 100
G6, self-tuned 83 of 100
MASSIVE
The model alone 66 of 100
G6 as built 66 of 100
G6, a person's prompt 58 of 100
G6, random changes, same gate 70 of 100
G6, self-tuned 70 of 100
SciFact
The model alone 71 of 100
G6 as built 73 of 100
G6, a person's prompt 73 of 100
G6, random changes, same gate 71 of 100
G6, self-tuned 75 of 100
MBPP
The model alone 62 of 100
G6 as built 61 of 100
G6, a person's prompt 57 of 100
G6, random changes, same gate 61 of 100
G6, self-tuned 60 of 100
ARC-Challenge
The model alone 86 of 100
G6 as built 86 of 100
G6, a person's prompt 85 of 100
G6, random changes, same gate 86 of 100
G6, self-tuned 86 of 100
banking77
The model alone 48 of 100
G6 as built 48 of 100
G6, a person's prompt 56 of 100
G6, random changes, same gate 83 of 100
G6, self-tuned 82 of 100
Root causes and raw data →

The working theory, in full

What G6's harness engineering rests on: models fitted to its benchmark runs, the arguments for learning safely when answers cannot always be checked, and the specifications built from them.

DocumentKey pointsRests on
A unified mathematical theory of agentic intelligence

Twenty-five models of self-correction, token use, delegation, decomposition, debate and scaling, fitted to G6's benchmark runs of March to May 2026.
  • Empirically calibrated models, not proved theorems.
  • Accuracy rises towards a ceiling as the harness learns; the theory loop levels off at about 17 to 20 theories.
  • Errors compound along a chain of steps, and a model checking its own work shares its own mistakes, so a small model cannot verify itself.
  • Estimates how much a harness adds to a model, in accuracy and in tokens.
 Benchmark runs: GAIA, Omni-MATH, SWE-bench Pro, MATH-500, MetaTool and MedXpertQA.
Theoretical foundations

An essay placing G6 in the history of AI: language models as perception, symbolic search as reasoning.
  • Argues that intelligence is a combinatorial optimisation problem: models propose, search and checks decide.
  • Uses category theory (the Yoneda lemma, sheaves) as an illustrative analogy, and the correspondence between proofs and programs.
  • Proposes a compression conjecture: caching verified algorithms turns statistical guessing into deterministic lookup.
  • Maps the four levels of adaptation in practopoiesis onto G6's self-training levels.
 Theory, drawing on the unified theory's benchmark runs.
The validation paradox: how G6 can learn when it cannot always check its own answers

An essay arguing that G6 can learn from its failures without a universal checker of correctness.
  • Separates checking one result, checking a learning step and checking general competence: progress on the second does not need the third.
  • A five-level ladder of checks, from a formal certificate to no evidence that discriminates.
  • The true gain is at least the measured gain minus the evaluator's errors: a 5-point gain measured with 4 points of evaluator error cannot be shown to be real.
  • Learning may be speculative; deployment must be justified.
 Argument, drawing on an earlier G6 experiment with a detector of collateral edits.
Evidence-constrained harness adaptation: an implementation specification

A build specification for the smallest loop in which G6 finds a confirmed defect, learns a safeguard it can run, and tests it on tasks it has not seen.
  • No universal oracle of correctness: each standard and check is scoped, and G6 abstains, qualifies or escalates when evidence is missing.
  • The milestone is one verified learning event: defect, diagnosis, safeguard, independent controls, transfer to an unseen repository, and a bounded claim.
  • Five kinds of evidence, from a formal check to none that discriminates; finding no defect is not proof of correctness.
  • The learner may never change the standards, hidden tests, evaluators or promotion rules.
 Designed from G6's earlier evidence experiments, 352 episodes per arm.
A computational structural intermediate representation for compositional, evidence-constrained and self-improving AI agents

A mathematical account of AI agents as typed systems that build, run, check and improve compositions of tools, keeping computing an answer apart from producing evidence about it.
  • Keeps computation and evidence-producing checks apart, which it argues is essential to avoid the self-validation paradox.
  • Proves, under stated assumptions, that reflecting without new information cannot break a genuine barrier, and that near-identical model reviewers are not independent layers of protection.
  • Proposes, as hypotheses, a joint account of what an agent can solve and what it can justify.
 Theory, with a reference implementation and its tests.
An integrated research protocol: composition, evidence and learning

A proposal for experiments that measure separately what G6's typed mission graphs, its evidence policy and its adaptive learning each add.
  • Three effects, each estimated on its own, with hypotheses and falsifiers written down in advance.
  • Paired comparisons by task or repository; with scarce data, case-level results with declared uncertainty.
  • An evaluator the system under test can change is a hard blocker.
 Proposal.

How every component meets the four hard problems of self-improvement

ProblemWhat every component does about it
Turning weak or partial evidence into reliable learning signalsEvery failure it catches is written to the run ledger as a labelled incident with its evidence and its failure mode: the input the meta-harness and the theory loop learn from.
Evaluators the adapting system cannot influenceIts checks are its own code, pinned by digest, or an outside tool (tests, a prover, a renderer), never written or tuned by the model being checked.
Transfer of learned safeguards to unseen task familiesA safeguard is bound to a failure mode, not to a task, so it applies wherever that mode can occur; its catch rate is measured on task families it was not built on.
Enough held-out examples to tell a real improvement from noiseIts catch rate and false-alarm rate are judged by G6's sequential test, which says how many more cases it needs instead of guessing.

The papers are design rationale: their proofs are not machine-checked, and none is cited as evidence.

Contact

Write to .

Every email is reviewed by the founder, and we endeavour to respond within 48 hours. Thank you for your correspondence.