gmtigrisva/DMAS

Diagnosing the belief, not the bug

A deductive multi-agent system for automated cognitive misconception diagnosis and Socratic intervention in advanced computer science education

---

Contemporary AI tutors operate in a direct response paradigm: a student brings a problem, the model returns the solution. That is efficient and pedagogically inert. In cumulative disciplines a correct answer handed over does not repair the belief that produced the wrong one, so the same cognitive error reappears in the next, harder problem.

This system inverts the use of the model's reasoning. Instead of pointing it at the problem, it points it at the student: three agents collaborate to find the exact instant a learner's reasoning left the correct path, locate the prerequisite concept that made that misstep possible, and ask the question that exposes the contradiction — without ever revealing the answer.

./dmas session --file attempt.py

Runs immediately from a fresh checkout. No install, no API key, no network, no third-party packages.

The session

The tutor is interactive, because Socratic teaching is. It reveals one rung at a time — a ladder printed all at once is just a worked answer delivered slowly — and it remembers you between attempts, so after you edit the file it can answer the question that matters: was the belief repaired, or the symptom patched?

Session transcript · plain text

dmas ▸ load attempt.py

  loaded  attempt.py  ·  10 lines  ·  task  lower_bound

   DIAGNOSIS   Mixed interval conventions  ·  0.77  ·  L4
  5/7 inputs pass

______________________________________________________________

dmas ▸ hint

   hint 1/3  ·  orienting  ·  Loop invariants

   Take the input ([0], 1) and step through your own function by hand, writing
   down the values at line 4 each time you reach it. Stop at the first moment a
   value surprises you.

   sit with that one first — `hint` again when you are stuck

______________________________________________________________

dmas ▸ hint

  hint 2/3  ·  contradiction  ·  Loop invariants

  At that point your `hi` holds 0, but for the answer to be reachable it would
  have to be 1. Which line last wrote `hi`, and what did you assume about the
  range it was allowed to take?

      … the student edits the file in their editor …

______________________________________________________________

dmas ▸ retry

   NO FINDING   87/87 inputs pass; no structural warning sign either

── Attempt 2 vs 1 ─────────────────────────────────────────────────────────

  resolved  Mixed interval conventions
  checks    5/7 inputs pass → 87/87 inputs pass
  mastery   Loop invariants  0.39 ↑ 0.72

A partial fix that makes things worse is reported just as plainly — persists, new, checks 5/7 → 0/7 — which is the honest thing and the useful thing.

Command reference · copyable text

command
load <file>	read a file; the task is detected from the function it defines, or falls back to structural findings for a file the tutor has no reference for
hint	the next rung, and only the next rung
challenge	a predict-then-run task built from the minimal failing input
why · code · report	the evidence, the marked source, the full report
retry	re-read the file after you edit it, and diff the two attempts
mastery · progress	what the session believes you have mastered; every attempt so far
concept [<id>] · problems · problem <id>	explore the curriculum; pin a task

Tab completion and history work in a terminal; commands can also be piped in, which is how the tests drive it.

---

What it actually does

Given a flawed submission, the system produces this:

Report transcript · plain text

── Cognitive diagnosis ──────────────────────────────────────────────────────

  misconception                    name                        belief
  bs.interval-convention-mismatch  Mixed interval conventions  █████████░░░ 0.77  MAJOR  L4

  The flawed belief: "I think the upper bound is just 'the end of the array'
  — which end does not matter."

  why:
    • [dynamic] execution first diverges at line 4, the same line this rule flags
    • [static]  the guard `lo < hi` treats `hi` as an exclusive bound, but it is
                initialised to `len(a) - 1`, the last valid index

── Knowledge-graph alignment ────────────────────────────────────────────────

  root cause  Loop invariants  ·  loop-invariant, depth 4
  the deepest implicated concept whose own prerequisites are already secure
  learning path  Loop invariants → Half-open intervals → Boundary conditions

── Cognitive intervention ───────────────────────────────────────────────────

  target  Loop invariants  ·  success probability  61%  ·   IN ZPD

  1 · orienting
     Take the input ([0], 1) and step through your own function by hand,
     writing down the values at line 4 each time you reach it.

  2 · contradiction
     At that point your `hi` holds 0, but for the answer to be reachable it
     would have to be 1. Which line last wrote `hi`, and what did you assume
     about the range it was allowed to take?

  3 · conceptual
     Write down the property you believe holds at the top of every iteration.
     Does it still hold immediately after your update lines run?

  answer leakage  0.00  ·  grounding  0.55

Note what is not there: the fix. hi = len(a) never appears, and the gate that guarantees that is a mechanism, not an instruction — see below.

The three agents

Socratic Evaluator. Executes the submission and a reference implementation on the same minimised failing input, matches their variables by behaviour rather than by name, and reports the earliest checkpoint at which the two disagree. Symbolic rules, trace divergence, cost analysis and the reasoning model are fused with Dempster–Shafer belief functions — so agreement between independent evidence types compounds, and disagreement is surfaced as contested rather than averaged away. A leading hypothesis below the settlement threshold is labelled provisional rather than presented as a verdict; on the development bank 16 of 23 diagnoses are settled and 7 are flagged provisional.

Knowledge-Graph Alignment. Propagates blame through a 65-concept prerequisite DAG by personalised PageRank, folds it into a Bayesian knowledge-tracing model as soft negative evidence, and extracts the deficiency frontier: the unmastered concepts whose own prerequisites are already secure. Those are the only places where teaching can succeed — Vygotsky's zone of proximal development, read off the graph.

Cognitive Intervention. Selects a target inside that zone by item response theory, grounds its language in retrieved evidence cards, and emits a three-rung Socratic ladder plus a counter-factual challenge — predict what your function returns, then run it.

The guarantee

No guidance reaches a student without passing the answer-leakage gate.

Enforced by replacement, not by warning. Three independent detection channels: MOSS-style winnowing fingerprints catch a copied passage; AST structural matching catches a statement of the reference the student has not written; edit-proximity catches a modified version of a line the student has written. Prose repairs ("set hi to len(a)") are normalised into assignments first, because a model told not to write code reaches for that phrasing immediately.

Quoting the student verbatim is explicitly allowed. Making someone look at what they actually wrote is the method.

The test suite includes a backend that deliberately emits the full reference solution; the assertion is that the student never sees it, and that the guidance they see instead is still three usable rungs.

Results

Every number below is regenerated by the commands in experiments/ and quoted in the paper through generated macros, so the README, the JSON and the paper cannot disagree. The reasoning backends for the "+ model" rows are openai/gpt-oss-120b (the paper's primary model) and openai/gpt-oss-20b (a comparison), open-weight reasoning models served through an OpenAI-compatible endpoint; the other rows use the deterministic offline generator and run with no network.

RQ1 — diagnostic accuracy

Two labelled banks. The development bank (23 labelled + 6 correct) was built alongside the rules and establishes internal validity only. The held-out bank (25 labelled + 6 correct, six new tasks) was written after the rules were frozen and labelled before the system ran; no rule or threshold was changed afterwards (the protocol is stated in data/heldout.py).

bank	system	top-1	recall	micro-F1	κ	FPR on clean code
development	deterministic	1.00	1.00	0.81	1.00	0.00
development	+ gpt-oss-120b	1.00	1.00	0.81	1.00	0.00
development	+ gpt-oss-20b	0.96	1.00	0.79	0.95	0.00
development	gpt-oss-120b only	0.87	0.91	0.83	0.86	0.00
held-out	deterministic	0.76	0.76	0.71	0.74	0.00
held-out	+ gpt-oss-120b	0.88	0.92	0.78	0.87	0.00
held-out	+ gpt-oss-20b	0.96	1.00	0.82	0.96	0.17
held-out	gpt-oss-120b only	0.80	0.88	0.71	0.79	0.00

The six held-out misses are listed and explained in the paper's appendix: four are symptom patterns the rules do not cover (the failing input is found and localised but no rule names the belief), one is a wrong rather than missing base case, one is recursion inside a nested helper. Fusing gpt-oss-120b closes three of the six without adding a false positive; it proposed no belief on any of the twelve clean programs. The smaller gpt-oss-20b names more (top-1 0.96 held-out) but also accuses one correct program of an asymptotic gap, which the reliability discount (0.62) and the clean run's opposing mass reduce to a provisional 0.37 without silencing — the case for leaving the verdict on clean code to the executional sources.

External validation — 4,225 real student programs

The Refactory corpus (Hu et al., ASE 2019) provides programs written by 361 students for five introductory assignments, each marked correct or wrong by the course's tests. Those labels say nothing about why, so the corpus cannot score diagnostic accuracy; it can score false accusations and detection.

measure	value
course-accepted programs on which a correctness belief was named	4.5%
… with only a cost or hygiene finding	31.2%
course-rejected programs on which a failing input was found	96.5%
… localised to a line	99.6%
… on which a correctness belief was named	50.3%
latency per file	37 ms

Of the rejected programs with no failing input, 59 use the built-in sort in an assignment that forbade it. Of the accepted programs accused of a correctness belief, 51 fail on an input inside the course's contract that the course's own tests did not cover.

The gate

On 274 manufactured leaks (a reference line verbatim or renamed, a prose repair around a code span, the whole solution) the gate rejects 99.3%; on 416 legitimate items (the student's own line quoted back, a question, a fallback rung) it rejects 0.0%. A repair stated in plain words with no code span is beyond the gate's design; the grounding channel happens to catch 50% of those, and the paper says so.

RQ2 — design analysis with simulated learners

These are simulated learners. The response model encodes the scaffolding assumption under test, so any single run reproduces it. What the simulation establishes is narrower: under a null response model in which the arms cannot differ, the family-wise rejection rate over 200 cohorts is 1.0% (the statistics do not manufacture an effect); under the default model the retention effect is Hedges g = 0.56 [0.28, 0.73] and the recurring-error effect g = -0.59 [-0.89, -0.34] over 30 cohorts; and the one quantity that does not depend on the response model — the on-target rate, the share of interventions aimed at a concept the learner actually lacks — is 93% for the full system and 39% without the knowledge graph. Power curves for sizing a human study are in experiments/results/rq2_design_offline.json.

RQ3 — robustness

Semantics-preserving rewrites (rename every identifier, split tuple assignments, expand augmented assignments, invert conditionals, inject dead code, all five at once), each mutant verified behaviourally identical before scoring. Top-1 stability: 0.942 over 138 development mutants, 1.000 over 150 held-out mutants. Renaming costs nothing, because variables are matched by behaviour.

Install and use

git clone <repository> && cd deductive_multi_agent_system_\(MAS\)
./dmas doctor                      # verify the environment
./dmas session --file attempt.py   # the interactive tutor
./dmas demo                        # a non-interactive guided tour

./dmas is a two-line launcher that puts src/ on the import path, so a fresh checkout works with no install step. pip install -e . puts an equivalent dmas command on your PATH; every example below works with either.

dmas session                                       # interactive; `load` a file inside
dmas diagnose --submission lb_inclusive_bound      # one-shot report, from the built-in bank
dmas diagnose --file attempt.py --problem lower_bound
dmas kg show binary-search                         # inspect the curriculum
dmas kg path binary-search sequence-indexing
dmas retrieve "why does my BFS revisit nodes" --full
dmas bench --json > rq1.json                       # the three studies
dmas experiment --students 180 --json > rq2.json
dmas robustness --json > rq3.json

Every command takes --json, and every command is deterministic given --seed.

With a reasoning model

The paper's live results use openai/gpt-oss-120b, an open-weight reasoning model, through Groq's OpenAI-compatible endpoint. That is the default, so with a key in the environment the flags below are optional:

export OPENAI_API_KEY=...
dmas diagnose --submission mc_greedy --backend openai \
  --base-url https://api.groq.com/openai/v1 --model openai/gpt-oss-120b

A locally served model works the same way (--base-url http://localhost:11434/v1 --model <name>), and a Claude backend is available (--backend claude --model claude-opus-5, reads ANTHROPIC_API_KEY). The backend is a swappable component behind a typed task interface with a JSON schema per task, so running the offline generator and a model and diffing the reports is the ablation that measures what the model contributes. A model's chain of thought is captured for the audit trail and never shown to a student — reasoning traces routinely contain the answer, which is exactly what this system must not disclose. Every completion is cached by content hash (keyed by backend and model), so a study can be resumed and re-derived without network access.

As a library

# PYTHONPATH=src, or after `pip install -e .`
from deductive_mas import DeductiveOrchestrator, problem, submission

result = DeductiveOrchestrator().tutor(
    problem("lower_bound"), submission("lb_inclusive_bound")
)

print(result.diagnosis.primary.misconception_id)   # bs.interval-convention-mismatch
print(result.alignment.root_cause)                 # loop-invariant
for hint in result.intervention.hints:
    print(hint.level, hint.text)
assert result.intervention.leakage < 0.18          # the gate held

Why no dependencies

Every numerical, retrieval and statistical routine is implemented here from first principles: one-sided Jacobi SVD, randomised range finding, the Hungarian algorithm, dynamic time warping, Okapi BM25, reciprocal rank fusion, winnowing fingerprints, Dempster's rule, the regularised incomplete beta and gamma functions, BCa bootstrap intervals.

That is a deliberate scientific choice, not asceticism. Every p-value in the results section traces to code in this repository, the statistics are validated against published reference tables in tests/test_stats.py, and the whole system runs identically on any machine with a Python 3.9 interpreter — which is the deployment setting the proposal targets.

Documentation

docs/ARCHITECTURE.md	layering, the blackboard, evidence flow, the compliance gate
docs/ALGORITHMS.md	every algorithm, why it was chosen, and the subtleties that bit
docs/EXPERIMENT_PROTOCOL.md	how each RQ is measured, and what the results do and do not establish
docs/PROPOSAL_TRACEABILITY.md	claim-by-claim mapping to the proposal, including the deviations

Limitations

Stated plainly, because a system that diagnoses overconfidence should not exhibit it.

1. Diagnostic accuracy is measured on a bank developed alongside the rules. Internal validity only.
2. The RQ2 cohort is simulated. No claim is made about human learning.
3. Grounding is support, not entailment. It catches invented vocabulary and off-corpus claims; a fluent, on-topic, subtly false sentence the corpus does not contradict can pass.
4. The guarded interpreter is not a security sandbox. Three independent limits, appropriate for classroom submissions, unsuitable for adversarial code.
5. Cost inference is an upper bound, deliberately over-approximating data-dependent loops; every complexity claim shown to a student is corroborated dynamically first.
6. Cross-algorithm comparison is coarse. When a submission attacks the problem by a different method than the reference — a greedy scan against a dynamic-programming solution — the variable correspondence is weak by construction. Roles below a confidence bar are dropped and the divergence falls back to the output level, but the resulting contradiction is less pointed than it is for two implementations of the same algorithm.
7. Python only. The analysis layer is AST-specific; the knowledge, retrieval, belief and statistics layers are not.
8. Files with no reference get structural findings only. Differential testing, trace alignment and the complexity comparison all need an oracle; for an unknown task the session says so and reports what the rules, the dataflow facts and the cost analysis can still establish.

Layout

Repository layout · plain text

dmas                                  run the CLI from a checkout, no install
src/deductive_mas/
  domain.py  config.py  pipeline.py     shared vocabulary, settings, composition
  
  util/          determinism, tokenisation, linear algebra

  analysis/      CFG, dataflow, cost, tracing, alignment, shrinking, leakage

  knowledge/     ontology, misconception taxonomy, mastery, ZPD

  retrieval/     BM25, LSA, fusion, corpus, MA-RAG

  llm/           Claude, OpenAI-compatible endpoints, offline reasoner, caching

  agents/        the three agents, the gate, the orchestrator

  simulation/    RQ1 benchmark, RQ2 experiment + design analysis, RQ3 robustness,
                 gate benchmark and audit, external validation (Refactory)
  stats/         special functions, tests, effect sizes, power, agreement
  ui/            terminal user interface, the interactive UI
  cli/           terminal presentation, the dmas command, the interactive session
tests/           the test suite
experiments/     result files, figure and number scripts, the live-run driver
docs/

Licence

MIT.