# Hard grade: “What Is Actually Recomputable”

**Quality: 5/10. Below the pay bar of 8. Flag: thin.**

The thesis makes a useful distinction between public records and reproducible model generation. It does not develop that distinction into an IMD/SIMD-specific investigation. Its central criticism targets an unattributed general claim, its challenge example is unnamed, and it conflates independent checking with exact regeneration. This is a competent cautionary outline, not an original technical essay.

Evaluated text: the thesis supplied in the assignment, attributed there to @chidifinance_. [Submitted tweet](https://x.com/chidifinance_/status/2107215226122719449). The tweet could not be retrieved directly; its publication, full contents, and engagement were not independently authenticated. Follower count plays no role in this grade. Research date: 5 October 2026 UTC. Current documentation helps assess the argument; it does not establish what documentation existed when the author wrote it.

## Evidence and claim assessment

| Thesis claim | Attributable evidence | Assessment and limit |
| --- | --- | --- |
| Connected agents, enrolled seats, and completed jobs are publicly queryable. | Direct GETs to [IMD health](https://api.imd.fun/health) and [swarm](https://api.imd.fun/swarm), captured around 21:08 UTC: 611 connected daemons, 621 active enrollments, 8,716 total jobs, 8,534 completed. The [SIMD state API](https://www.si-md.xyz/api/state) separately labels its metrics with the IMD fields they consume. Selected observations are preserved in [evidence.json](evidence.json). | **Observed fact:** these endpoints expose counts without authentication. **Limit:** querying an operator's aggregate is not independently reconstructing its database or proving each completion. |
| Recent snapshots show roughly 590–610 connected agents and over 8,000 completed jobs. | The live count was 611; the browser-retrieved [agents page](https://explorer.imd.fun/agents) displayed 608 online. The swarm API reported 8,534 completed jobs. | **Broadly consistent scale**, not verification of the author's unspecified historical snapshots. Different retrieval times and caches prevent treating these as synchronized readings. Connected, enrolled, completed jobs, and accepted steps are separate measures. |
| Payments and job metadata can be inspected. | [IMD API documentation](https://imd.fun/docs/) exposes job details, submissions, results, and work records. [SIMD's public app source](https://www.si-md.xyz/app.js?v=57), which renders its Docs view, publishes a vault address and describes refunds as IMD transfers. | **Documented inspection surfaces.** Job metadata is not thereby all on-chain. A transfer alone does not establish the associated work's correctness. No transaction receipt or payment-to-job reconciliation was independently verified here; an attempted Etherscan transaction-page retrieval failed. |
| Some challenge conditions can be recomputed directly. | The [SIMD collision API](https://www.si-md.xyz/api/collisions) specifies distinct inputs whose hash prefixes match over 2λ bits; it names keccak256, SHA-256, and RIPEMD-160. At capture, all three λ=24 rungs were open and the successful-results list was empty. | **Concrete documented mechanism:** check two digests and their first 48 bits. **Limit:** this establishes a checkable condition, not a successfully reproduced public solution. The thesis omits the mechanism entirely. |
| Different seats use different models, parameters, and hardware. | The [official worker README](https://github.com/Identity-md/worker#start-and-pair) supports Claude Code and Codex, configurable inference settings, and restricted premium configurations for some work. It permits multiple host platforms. | **Supported:** execution configurations can differ. **Unproven:** every seat differs, or the actual model-serving hardware differs. Worker host platform and inference-provider hardware are not interchangeable evidence. No paired generation experiment is presented. |
| Recomputation is generally overstated and partly aspirational. | Neither the supplied text nor the inspected SIMD Docs view identifies a dated, attributable promise that every model output can be regenerated exactly. Its collision view makes a narrower recomputation claim. | **Unestablished inference.** The author must identify the promise and compare it with the applicable verification contract. Exact regeneration limits do not by themselves refute independent output checking. |

These are observations and documented behaviors, not an audit of deployed backend code. No model-generation, oracle-replay, or successful collision experiment was performed.

## The missing distinction

“Recompute” needs an object and a success criterion. The following distinctions are analytical judgments in this review, not additional substance credited to the author:

1. **Check saved bytes.** IMD result records include hashes and download URLs. Matching a digest establishes consistency with the referenced bytes, not the truth of a report. [IMD documentation](https://imd.fun/docs/).
2. **Recalculate a predicate.** A collision's two inputs and prefix rule can support a cheap check even if the search that discovered them was expensive. That check need not recreate the search. [SIMD collision API](https://www.si-md.xyz/api/collisions).
3. **Replay pinned evidence.** IMD documents chain-evidence oracle recipes such as `log-count` and `call-compare`, rerun at pinned blocks before signing. A signature authenticates an attestation; external replay still needs the inputs and chain access. [IMD documentation](https://imd.fun/docs/).
4. **Validate generated work.** The worker README explicitly separates worker-side website build/type/interaction checks from verifier structure and integrity checks. These checks have different assurance levels. [Official worker README](https://github.com/Identity-md/worker#start-and-pair).
5. **Regenerate the original answer.** Recreating an entire model/tool execution with identical bytes is a stronger demand. The thesis supplies no experiment showing when this succeeds or fails.

The strongest counterargument is therefore straightforward: a stochastic producer can create an artifact with a deterministic acceptance condition. Failure to reproduce the producer's exact wording or search trajectory does not imply failure to verify that condition. Conversely, passing an integrity check cannot establish factual accuracy. The thesis gestures at these limits but does not articulate either verification contract.

## Why the score stops at 5

The piece has a clear argument, restrained language, and a sensible public-inventory recommendation. It avoids treating payment as proof of quality. Those are real strengths.

Its substance remains thin. “Certain challenge formats” substitutes vagueness for a named mechanism. “The strongest epistemic claim” substitutes assertion for attribution. The dashboard figures establish activity scale, not the disputed recomputation boundary. The proposed inventory is neither supplied nor specified: no artifact class, pinned input, verifier, dependency set, success criterion, or known failure is documented by the author.

SIMD specificity is especially weak. Mentioning @SuperIMD_eth does not explain SIMD's relationship to IMD, its collision ladder, or its actual public assurance labels. The report's research cannot be retroactively awarded to the thesis. This earns 5 for a competent outline; it lacks enough concrete work for 6, fails the named-mechanisms requirement for 7, and has neither originality nor depth for 8.

## Unanswered questions and a useful revision

- Which exact IMD or SIMD statement promises universal output regeneration, and when was it made?
- Can an outsider obtain all inputs and dependencies required to rerun a selected acceptance check? Which part still depends on the control plane?
- Which claims are independently checked, which are only operator-reported, and what does an accepted verdict actually establish for each job class?

A stronger essay would identify one promise, inspect one delivered job, and attempt one applicable check. It would publish the input versions, criterion, result, and missing dependencies, then compare that evidence with a second artifact class. This would turn a familiar reproducibility caveat into a falsifiable contribution about IMD/SIMD.

**Impact, assessed separately:** the thesis can improve discourse by asking readers to distinguish observable payments, reported activity, verified conditions, and reproducible generation. Its value is clarification, not a demonstrated discovery. Reach and engagement are unverified and are not scored.

```json
{"quality":5,"impactNote":"Useful clarification of the gap between observable payments, reported activity, condition checking, and reproducible generation; reach is unverified and excluded from quality.","notes":"Clear, restrained argument and sensible inventory recommendation, but no attributable universal recomputation promise, named challenge mechanism, replay experiment, or developed SIMD mechanics. Conflates exact regeneration with independent validation. Current counts broadly support scale without proving historical snapshots or independent correctness.","flags":["thin"]}
```